Novel approach to improve the assessment of biodiversity of phytoplankton communities based on hyperspectral data analysis
Abstract
La gestió sostenible dels ecosistemes marins requereix d'un millor coneixement de la distribució de certs paràmetres ecològics com les comunitats de fitoplàncton, inclosos aquells grups algals causants d'afloraments. La caracterització acurada dels patrons espacio-temporals de la biodiversitat del fitoplàncton és essencial per tal d'avaluar el paper de cada grup algal en l'ecosistema global marí i els cicles biogeoquímics. En l'intent d'abordar aquesta qüestió, observacions espectromètriques adquirides in situ i per teledetecció han demostrat proporcionar informació valuosa sobre la distribució de diversos components òpticament actius presents a l'aigua de mar, tant a escala local com global, i en particular, sobre l'estructura de les comunitats fitoplanctòniques. En aquest sentit, l'aparició de sensors òptics d'alta resolució espectral (hiperespectrals) ha augmentat les expectatives per discriminar la composició de les comunitats de fitoplàncton, permetent anar més enllà de l'estimació del pigment primari del fitoplàncton, la clorofil·la-a, utilitzat convencionalment com un indicador global de la biomassa i la producció primària de fitoplàncton ja que és un pigment comú a tots els grups taxonòmics. Aquesta tesi doctoral s'ha dut a terme amb l'objectiu de millorar la nostra capacitat d'extraure informació sobre l'estructura de les comunitats de fitoplàncton en l'oceà, mitjançant el desenvolupament i avaluació d'una nova aproximació basada en l'anàlisi de dades hiperespectrals. En particular, s'ha proposat una tècnica de classificació, on s'examinen les dissimilituds entre les signatures hiperespectrals de cada mostra d'aigua, i que ha estat aplicada en combinació amb l'espectroscopia derivativa, tècnica amb la qual s'exploren les característiques de la forma de l'espectre analitzat. Com a novetat, es proposa també una eina de validació per demostrar l'eficàcia d'aquesta nova tècnica de classificació. En el procés de validació es demostra que les classificacions obtingudes amb les dades òptiques i la metodologia proposada són molt semblants a les obtingudes amb dades basades en l'anàlisi de la composició pigmentària (utilitzant un cromatògraf líquid d'alta resolució, HPLC), tècnica habitualment utilitzada per la comunitat científica com a indicador de la composició de fitoplàncton. La viabilitat d'aquesta metodologia ha estat demostrada inicialment utilitzant una aproximació basada en simulacions, on la distribució del fitoplàncton està predeterminada i on s'han generat diferents escenaris lumínics d'aigües en mar obert i costaneres mitjançant un model de transferència radiativa. Per altra banda, escenaris reals de mar obert corresponents a diferents estacions en l'oceà Atlàntic han estat classificats satisfactòriament mitjançant les tècniques proposades, aplicades a dades hiperespectrals incloent espectres d'absorció i reflectància, així com els seus espectres derivats. Aquesta classificació ha servit per identificar una aplicació potencial de la metodologia proposada: l'establiment de diferents províncies bio-òptiques a partir de l'anàlisi de mesures hiperespectrals oceanogràfiques, donant lloc a l'examen de la seva rellevància biogeogràfica en comparació amb províncies ecològiques proposades prèviament en la literatura. Aquesta tesi conclou amb la confirmació de la hipòtesi principal: una millor discriminació de l'estructura i dinàmica de les comunitats de fitoplàncton és possible mitjançant l'ús d'observacions hiperespectrals oceanogràfiques. Cal destacar que l'aproximació proposada és en general aplicable a diferents conjunts de dades, més enllà de la composició pigmentària o dades òptiques obtingudes in situ també a dades obtingudes per teledetecció, dades biogeoquímiques i altres paràmetres hidrogràfics.
Full text
Programa de Doctorat en Enginyeria Òptica Novel approach to improve the assessment of biodiversity of phytoplankton communities based on hyperspectral data analysis Elena Torrecilla Ribalta Supervisors Dr. Jaume Piera Fernández, UTM/CSIC Dr. Meritxell Vilaseca Ricart, CD6/UPC A thesis submitted for the degree of Philosophy Doctor (PhD) July 2012
Abstract Sustainable management of marine ecosystems requires a better knowledge about the space-time distribution and dynamics of ecological parameters such as phytoplankton communities, including critical bloom-forming algal groups. Better understanding of phytoplankton biodiversity and dynamics is essential in evaluating the role of each algal group in the global marine ecosystem and biogeochemical cycles. In attempting to address this question, in situ and remotely-sensed spectrometric optical observations have demonstrated to provide previously unavailable information regarding several optically active constituents in seawater at local and global scales, in particular, regarding phytoplankton community structure. In this sense, the advent of high spectral resolution (hyperspectral) optical sensors have raised new expectations about the possibilities of discriminating phytoplankton community composition in the ocean, beyond the estimation of only the primary pigment in phytoplankton, chlorophyll-a, a proxy for the phytoplankton biomass and primary production since it is common to all taxonomic groups. This PhD thesis has been carried out with the aim of improving our ability to extract information regarding phytoplankton community structure in the ocean by developing and evaluating a novel approach based on hyperspectral data analysis. In particular, a dissimilarity-based cluster technique, which accounts for complete spectral behaviour of hyperspectral data of each seawater sample, has been applied in combination with derivative spectroscopy, which exploits the spectral shape features of each analyzed spectrum. As a novelty, a validating tool has been proposed and proven useful to illustrate the eectiveness of the optical-based classication for discriminating dierent phytoplankton assemblages. This novel validation approach is based on the pigment composition analyzed in conjunction with concurrently obtained optical data, information which has been commonly used by the scientic community as a proxy for the phytoplankton composition. The feasibility of this methodology has initially been demonstrated using a simulation-based approach, i. e. using a radiative transfer modeling framework for open ocean and shallow coastal environments. In addition, dierent real open ocean environments corresponding to several stations in the eastern Atlantic Ocean have successfully been classied by applying the cluster analysis to dierent hyperspectral data sets including absorption and remote-sensing reectance spectra and their second derivative spectra. This classication has served to identify a potential application of the proposed methodology: the establishment of dierent bio-optical provinces from the analysis of hyperspectral oceanographic observations, leading to examination of its biogeographical relevance by comparison to ecological provinces previously proposed in the literature. This thesis concludes by conrming the main hypothesis that discrimination of phytoplankton community structure and dynamics iii
in the ocean can be enhanced while using hyperspectral oceanographic observations. It is noteworthy that the proposed approach is generally applicable to dierent data sets, besides in-situ pigment or optical data data also to remotely-sensed, biogeochemical or hydrographic data sets. iv
Resum La gestió sostenible dels ecosistemes marins requereix d'un millor coneixement de la distribució de certs paràmetres ecològics com les comunitats de toplàncton, inclosos aquells grups algals causants d'aoraments. La caracterització acurada dels patrons espacio-temporals de la biodiversitat del toplàncton és essencial per tal d'avaluar el paper de cada grup algal en l'ecosistema global marí i els cicles biogeoquímics. En l'intent d'abordar aquesta qüestió, observacions espectromètriques adquirides in situ i per teledetecció han demostrat proporcionar informació valuosa sobre la distribució de diversos components òpticament actius presents a l'aigua de mar, tant a escala local com global, i en particular, sobre l'estructura de les comunitats toplanctòniques. En aquest sentit, l'aparició de sensors òptics d'alta resolució espectral (hiperespectrals) ha augmentat les expectatives per discriminar la composició de les comunitats de toplàncton, permetent anar més enllà de l'estimació del pigment primari del - toplàncton, la clorol · la-a, utilitzat convencionalment com un indicador global de la biomassa i la producció primària de toplàncton ja que és un pigment comú a tots els grups taxonòmics. Aquesta tesi doctoral s'ha dut a terme amb l'objectiu de millorar la nostra capacitat d'extraure informació sobre l'estructura de les comunitats de toplàncton en l'oceà, mitjançant el desenvolupament i avaluació d'una nova aproximació basada en l'anàlisi de dades hiperespectrals. En particular, s'ha proposat una tècnica de classicació, on s'examinen les dissimilituds entre les signatures hiperespectrals de cada mostra d'aigua, i que ha estat aplicada en combinació amb l'espectroscopia derivativa, tècnica amb la qual s'exploren les característiques de la forma de l'espectre analitzat. Com a novetat, es proposa també una eina de validació per demostrar l'ecàcia d'aquesta nova tècnica de classicació. En el procés de validació es demostra que les classicacions obtingudes amb les dades òptiques i la metodologia proposada són molt semblants a les obtingudes amb dades basades en l'anàlisi de la composició pigmentària (utilitzant un cromatògraf líquid d'alta resolució, HPLC), tècnica habitualment utilitzada per la comunitat cientíca com a indicador de la composició de toplàncton. La viabilitat d'aquesta metodologia ha estat demostrada inicialment utilitzant una aproximació basada en simulacions, on la distribució del toplàncton està predeterminada i on s'han generat diferents escenaris lumínics d'aigües en mar obert i costaneres mitjançant un model de transferència radiativa. Per altra banda, escenaris reals de mar obert corresponents a diferents estacions en l'oceà Atlàntic han estat classicats satisfactòriament mitjançant les tècniques proposades, aplicades a dades hiperespectrals incloent espectres d'absorció i reectància, així com els seus espectres derivats. Aquesta classicació ha servit per identicar una aplicació potencial de la metodologia proposada: l'establiment de diferents províncies bio-òptiques a partir de l'anàlisi de mesures hiperespectrals v
oceanogràques, donant lloc a l'examen de la seva rellevància biogeogràca en comparació amb províncies ecològiques proposades prèviament en la literatura. Aquesta tesi conclou amb la conrmació de la hipòtesi principal: una millor discriminació de l'estructura i dinàmica de les comunitats de toplàncton és possible mitjançant l'ús d'observacions hiperespectrals oceanogràques. Cal destacar que l'aproximació proposada és en general aplicable a diferents conjunts de dades, més enllà de la composició pigmentària o dades òptiques obtingudes in situ també a dades obtingudes per teledetecció, dades biogeoquímiques i altres paràmetres hidrogràcs. vi
Als meus pares, l'Elena i el Sebastià, perseverança i bondat
Agraïments/Acknowledgements Durant tots els anys de desenvolupament d'aquesta tesi, he estat envoltada de persones que han estat fonamentals per poder fer-la realitat. Ara ha arribat el moment de que els agraeixi la seva contribució. En primer lloc, el major agraïment és pels meus directors de tesi en Jaume Piera i la Meritxell Vilaseca. A en Jaume, creador inesgotable d'idees, li agraeixo l'haver-me proposat aquelles idees que van servir-me de llavor per anar desenvolupant el que ha esdevingut aquesta tesi. També l'haver-me introduït en aquest àmbit de recerca i la seva dedicació, així com la seva actitud motivadora i ment oberta en tot moment, valors que m'han permès anar conduint el meu treball amb llibertat i conança. Ja als inicis de col · laborar al seu grup, em va dir que fer una tesi era com una travessia del desert. Jo em sento afortunada, en aquesta travessia hi he trobat molta aigua. A la Meritxell, treballadora incansable, li estic molt agraïda pels coneixements i experiència que m'ha transmès al llarg d'aquests anys de supervisió. La seva actitud sempre activa m'ha servit en molts moments de referent per tirar endavant cada etapa de la tesi. Aquesta tesi no hauria vist la llum sense el nançament per part del Ministeri d'Educació i Ciència, mitjançant els programes Junta d'Ampliació d'Estudis (JAE Predoc, I3P-BPD2005) i Estades Breus I3P de l'Agència Estatal Consell Superior d'Investigacions Cientíques (CSIC), així com el programa de Menció Europea en el Títol de Doctor de la Secretaria General d'Universitats. La tesi ha estat realitzada en el marc dels projectes de recerca HIDRA (CSIC, PIE06-301102) i ANERIS (CSIC, PIF08-015). També m'agradaria donar les gràcies als meus companys de grup i despatx, l'Isma i el Sergi. Tot i que cadascú ha hagut d'anar resolent la seva tesi des de punts de partida diferents, ha estat un plaer poder fer-ho en paral · lel amb ells. Crec que tots hem après molt amb l'experiència, no sempre ha estat senzill, però espero que aviat també puguin culminar la seva travessia doctoral. Tampoc vull deixar de mencionar el meu agraïment a la resta de companys de la UTM pel bon ambient i estones d'esbarjo o dinars que hem compartit. Així com a tot el personal directiu i administratiu del CMIMA per posar-m'ho fàcil en els múltiples tràmits que he hagut de resoldre durant aquests anys. Alguns investigadors de l'ICM també han contribuït a que pogués dur a terme aquesta tesi. Entre ells, a l'Elisa Berdalet li agraeixo que sempre estigués a disposició per resoldre'm tots aquells dubtes en relació a alguns conceptes de biologia marina que estaven fora dels meus coneixements. A l'Antonio Turiel, pel seu interès en la nostra recerca i per tot sovint aconseguir-me alguns articles als quals jo no tenia accés. ix
List of Figures 3.6 Hyperspectral (1 nm) determinations of the remote-sensing reectance Rrs(λ) obtained from radiative transfer simulations (solid line) compared with in situ multispectral measurements at 13 discrete bands (solid circles). Each panel illustrates a dierent station location. . . . 64 3.7 Example of similarity between two hierarchical cluster trees, computed using the cophenetic and Rand indices. In the latter, an optimal number of clusters ( k ) needs to be pre-dened in the cluster tree. . . . . . 67 3.8 (a) Dendrogram obtained for the nine stations using 24 pigments to total chlorophylla (TChl a ) ratios determined from the CHORS HPLC analysis. (b) Linkage distances obtained from the cluster analysis shown in (a) as a function of distance along the dendrogram. . . . . . 70 3.9 Chlorophyll-specic normalized absorption coecients for the nine stations corresponding to: (a) phytoplankton, a∗ n,ph(λ) , (b) pure seawater plus phytoplankton, a∗ n,w+ph(λ) , and (c) total absorption, a∗ n(λ) . . . . 71 3.10 Results of cluster analysis applied to absorption data from the nine stations. The left panels represent dendrograms obtained using: (a) a∗ n(λ) , (c) a∗ n,w+ph(λ) , and (e) a∗ n,ph(λ) , and the right panels (b, d, and f) illustrate results obtained using each component's second derivative spectrum. .................................. 72 3.11 Similarity indices between absorption-based and pigment-based cluster trees obtained for the 9 stations using dierent combinations of spectral range for (a, b) a∗ n(λ) , (c, d) a∗ n,w+ph(λ) and (e, f) a∗ n,ph(λ) . The yaxis indicates the lower limit of the spectral range ( λmin ) and the x-axis the upper limit of the spectral range ( λmax ) used in the cluster analysis. Left and right panels depict the cophenetic and Rand indices, respectively. ................................ 74 3.12 Similar to Fig. 3.11, but based on cluster trees obtained using different spectral range combinations for the second derivative spectra of (a, b) a∗ n(λ) , (c, d) a∗ n,w+ph(λ) and (e, f) a∗ n,ph(λ) . Optimal values for band separation and window size determined from prior analyses ( BS = WS =9 nm, see Section 3.3.4) were used in the calculation of derivativespectra. ............................. 75 3.13 Histograms of cophenetic indices obtained for all combinations of spectral ranges shown in Fig. 3.11e and 3.12e based on (a) the absorption of phytoplankton, a∗ n,ph(λ) , and (b) its second derivative spectra. . . . 76 3.14 Hydrolight-simulated Rrs(λ) spectra, normalized at 555 nm, computed for the nine stations using measured IOPs as input. . . . . . . . . . . 77 xvi
List of Figures 3.15 Dendrograms resulting from cluster analysis of the nine stations calculated using four dierent sets of input data vectors: (a) three re- ectance band ratios of Rrs(λ) based on 4 SeaWiFS wavebands obtained from measurements with SPMR instrument, (b) 13 band ratios corresponding to multispectral measurements of Rrs(λ) with SPMR instrument, (c) hyperspectral (1 nm) ordinary (non-dierentiated) normalized Rrs(λ) spectra computed from the Hydrolight simulations, and (d) second derivative of hyperspectral normalized Rrs(λ) spectra obtained using optimal values for band separation and smoothing lter window (i.e., BS = WS = 27 nm, see Section 3.3.4). . . . . . . . . . 78 3.16 Linkage distances as a function of distance along the dendrogram for each cluster tree depicted in Fig.3.15. . . . . . . . . . . . . . . . . . . 79 3.17 Similarity indices between reectance-based and pigment-based cluster trees obtained for the nine stations using dierent combinations of spectral ranges for the second derivative of the hyperspectral normalized Rrs(λ) . Optimal values of BS = WS = 27 nm, determined from prior analyses, were used for the calculation of derivative spectra. The y-axis indicates the lower limit of the spectral range ( λmin ) and the x-axis the upper limit of the spectral range ( λmax ) utilized in the cluster analysis. Panels (a) and (b) depict the cophenetic and Rand indices,respectively. ............................ 80 3.18 The second derivative spectra at each station of the (a) chlorophyllspecic normalized phytoplankton absorption coecient, a∗ n,ph(λ) , and (b) normalized hyperspectral remote-sensing reectance, Rrs(λ)/Rrs(555) . Optimal values were used for the derivative calculations, i.e., BS = WS = 9 nm for the absorption data and BS =WS = 27 nm for the reectancedata. .............................. 82 3.19 Cophenetic indices obtained from similarity analysis between the pigmentbased cluster tree with trees obtained using the second derivative of hyperspectral normalized spectra for (a) phytoplankton absorption, a∗ n,ph(λ) , and (b) remote-sensing reectance, Rrs(λ) , in which dierent parameter sets for the derivative analysis are considered. The y-axis indicates the size of the lter window used in smoothing of the spectra ( WS ), and the x-axis represents the band separation used in the calculation of the derivative ( BS ). The analysis was conducted using the optimal spectral regions of 420 to 515 nm for a∗ n,ph(λ) , and from 435 to 495 nm for Rrs(λ) . ......................... 82 4.1 Map of the investigated area depicting the 48 stations sampled during the ANT XXV-1 cruise track during November 2008. Stations are identied by symbols corresponding to the groups dened based on the pigment composition (see Section 4.2 for details) and shown superimposed the oceanic provinces dened by Longhurst (2006) (see Section 4.2.1). .................................... 86 xvii
List of Figures 4.2 Ratios of 23 dominant pigments to TChl a corresponding to the 48 stations. Pigment abbreviations are: Chl c 3 = chlorophyllc 3, Chl c1+c 2 = chlorophyllc 1+ c 2, Peri = peridinin, But = 19' - butanoloxyfucoxanthin, Fuco = fucoxanthin, Neo = neoxanthin, Hex = 19' - hexanoyloxyfucoxanthin, Viola = violaxanthin, Asta = astaxanthin, Pra = prasinoxanthin, Dia = diadinoxanthin, Allo = alloxanthin, Diato = diatoxanthin, Zea = zeaxanthin, Lut = lutein, DVChl b = divinyl chlorophyllb , Chl b = monovinyl chlorophyllb , DVChl a = divinyl chlorophylla , Chl a = monovinyl chlorophylla , Phytb = phaoephytinb , Phyta = phaoephytina , α - caro = α - carotene and β - caro = β - carotene. . 90 4.3 Cluster tree obtained for the 48 stations using pigment data determined by HPLC. Main clusters are identied with symbols and roman numbers. .................................. 91 4.4 Spectral absorption of phytoplankton ( aph(λ) ) for all stations. . . . . 93 4.5 Results of cluster analysis applied to aph(λ) from the 48 stations. Symbols correspond to pigment-based clusters (see Fig. 4.3). . . . . . . . 94 4.6 Remote-sensing reectance spectra ( Rrs(λ) ) for the 21 stations measured. .................................... 95 4.7 Cluster tree resulting from analysis of the 21 stations calculated using second derivative of hyperspectral Rrs(λ) spectra (optimal values: WS =BS = 9 nm). Symbols correspond to pigment-based clusters (Fig.3.8). ................................. 96 4.8 Cophenetic indices between pigment-based and optical-based cluster trees obtained using dierent combinations of spectral range for (a) aph(λ) and (b) second derivative spectra of Rrs(λ) . Optimal values of WS =BS = 9 nm, were used for the calculation of derivative spectra. The y-axis indicates the lower limit of the spectral range ( λmin ) and the x-axis the upper limit of the spectral range ( λmax ) utilized. . . . . 96 4.9 Temperature-salinity diagram for the north-south transect in the eastern Atlantic Ocean (see Fig. 4.1). Points corresponding to stations are labeled with their station number and their symbol according to the pigment-based cluster partition (see Fig. 4.3). . . . . . . . . . . . . . . 99 5.1 (a) Map depicting the location of the Alfacs Bay. (b) Dierent toxic algal blooms recorded during a weekly sampling in 2007 performed by the IRTA. (c) Chlorophylla specic absorption coecient of two groups of phytoplankton: diatoms and dinoagellates (Kim and Philpot, 2006), reported to be present in the Alfacs Bay. These groups show distinct optical features, in particular in the spectral region from 435 to 580 nm where absorption characteristics of main accessory pigments appear.105 5.2 Bottom irradiance reectances provided by the Hydrolight code and considered for the present study. . . . . . . . . . . . . . . . . . . . . . 107 5.3 Diagram of the nonuniform vertical chlorophyll prole expressed in equation (5.2) and utilized in part of the RT simulations. . . . . . . . . 108 xviii
List of Figures 5.4 Examples of absorption (a) and backscattering (b) spectra utilized in one of the simulations for selected values of: Chl0 (1 mg/m3 , uniform vertical prole) and SS (2 mg/L ). In this case, the underwater scenario was characterized by the presence of the algal group of diatoms.108 5.5 Remote-sensing reectance, Rrs(λ) , of ve dierent bottom types seen through clear water with a maximum depth equal to 6.5 m. . . . . . . 110 5.6 (a) Examples results of radiative transfer simulations of optically shallow waters, showing the Rrs(λ) spectra corresponding to uniform vertical proles of two phytoplankton groups at dierent concentrations and given dierent bottom types. (b) MDS plot visualizing similarities between the Rrs(λ) spectra shown in (a). Symbols and colors represent each algal group and bottom type, respectively. Labels indicate the chlorophyll concentration considered in each modeled underwater scenario....................................111 5.7 (a) Examples results of radiative transfer simulations of optically shallow waters, showing the Rrs(λ) spectra corresponding to uniform vertical proles of two phytoplankton groups at a concentration of 0.5 mg/m3 and given dierent concentrations of suspended matter and bottom types. (b) MDS plot visualizing similarities between the Rrs(λ) spectra shown in (a). Symbols and colors represent each algal group and bottom type, respectively. Labels indicate the concentration suspended sediment considered in each modeled underwater scenario. . . . 112 5.8 (a) Examples results of radiative transfer simulations of optically shallow waters, showing the Rrs(λ) spectra corresponding to uniform (indicated in black symbols and lines) and nonuniform vertical proles (in red) of two phytoplankton groups. (b) MDS plot visualizing similarities between the Rrs(λ) spectra shown in (a). . . . . . . . . . . . . 114 5.9 Results of cluster analysis based on the irradiance reectance along the water column corresponding to a very shallow area characterized by the presence of a thin layer of phytoplankton at 4 m depth, obtained by means of radiative transfer simulation. . . . . . . . . . . . . . . . . 115 xix
List of Tables 3.1 Stations sorted into dierent classes characterized by diering pigment assemblages based upon the ratios of the concentrations of two dominant accessory pigments to TChl a (see the 2nd and 3rd columns from the left). The ratios of six dominant pigments to TChl a are also displayed, with the two most dominant accessory pigments indicated within the shaded areas. Pigment abbreviations are: MVChl a = monovinyl chlorophylla , DVChl a = divinyl chlorophylla , Fuco = fucoxanthin, Hex = 19' - hexanoyloxyfucoxanthin, But = 19' - butanoloxyfucoxanthin, MVChl b = monovinyl chlorophyllb , Chl c 2 = chlorophyllc 2, Zea = zeaxanthin, Pra = prasinoxanthin, Dia = diadinoxanthin, and α - caro = α -carotene. ............... 68 3.2 A comparison of similarity indices between pigment-based clusters and reectance-based clusters for the dierent sources of reectance data that are depicted in Fig. 3.15. For the case of the second derivative of hyperspectral reectance, the result of computations for two dierent spectral regions is given. . . . . . . . . . . . . . . . . . . . . . . . . . 79 4.1 Summary of cluster properties obtained for the 48 stations using the pigment information. Properties and pigments abbreviations are: TChl a = Chl a + DVChl a (Mean ± SD); Temp. = Surface temperature (Mean ± SD); Salinity = Surface salinity (Mean ± SD); MVChl a = monovinyl chlorophylla , DVChl a = divinyl chlorophylla , Fuco = fucoxanthin, Hex = 19' - hexanoyloxyfucoxanthin, But = 19' - butanoloxyfucoxanthin, Chl b = monovinyl chlorophyllb , Chl c1 = chlorophyllc1 , Chl c2 = chlorophyllc2 , Zea = zeaxanthin, and Dia = diadinoxanthin; = signicantly dierent from sub-cluster IV-S (p<0.05). (Table adapted from Taylor et al. ,2011). ................................ 92 4.2 Summary of results including cophenetic indices between pigmentbased and optical-based cluster trees obtained using dierent spectral ranges of aph(λ) , Rrs(λ) and second derivative spectra of Rrs(λ) . In parenthesis, the cophenetic indices obtained once the outliers stations (i.e., 11, 12, 13, 19, 20 and 38) were ruled out from the analysis. . . . . 97 xxi
Introduction Motivation The oceans cover two-thirds of the Earth's surface and contain 97% of the planet's water, the most important substance for life on Earth. They have a profound inuence on its environment since the oceans contain 99.5% of our planet's livable habitat. Monitoring and protecting water resources, including the oceans, is therefore a must and a challenge for all countries of the world. In this sense, the United Nations (UN) recommended the proclamation of an International Decade for Action from 2005 to 2015, Water for Life, which provided an excellent opportunity for the international community to advance towards a truly integrated approach to the management of the world's water. In accordance, authorities and initiatives like the European Union Water Framework Directive (WFD, 2000/60/EC), the Global Ocean Observing System (GOOS) of UNESCO, or the European Network Global Monitoring for Environment and Security (GMES) were established as frameworks to assess the marine role in Earth system dynamics by promoting a dense and frequent monitoring of the ocean's composition. Of particular importance are climate research related monitoring applications as to record indicators of the ocean's response to climate change, and to calibrate and validate biogeochemical ocean-ecosystem models (LeQuéré et al. , 2005). Sustainable management of oceanic ecosystems requires comprehensive information on phytoplankton community composition and distribution. Phytoplankton, as primary producers that convert light and CO2 into biomass through the photosynthesis, stand at the base of the oceanic food web. They are important to monitor for being an indicator of the biological state of the oceans. Their importance as a food source for the pelagos and as a potential sink of atmospheric carbon have been widely recognised (Holligan, 1992). Some phytoplankton groups have also been identied to have an important role in the climate regulation and the atmospheric sulfur cycle (Fig. I.1a). These groups named DMS-producers are able to produce the cellular component dimethylsulfoniopropionate (DMSP), which once converted into sulfate molecules are emitted to the atmosphere and act as cloud condensation nuclei, increasing general cloud cover (Vallina and Simó, 2007). In addition, some phytoplankton species are able to cause harm through the production of toxins or by their accumulated biomass. Such outbreaks which typically occur in coastal areas are known as harmful algal blooms (HABs, Babin et al. , 2008). Social impact of environmental changes associated to algal blooms is signicant (Fig. I.1b). When blooms are formed the toxins travel up the food chain and pose serious human-health threats and severely aect numerous industries and commercial sheries, causing substantial economic losses in the form of reduced sales or diminished tourist activity. 1
Introduction Figure I.1: (a) Diagram illustrating the role of phytoplankton in the sulfur cycle. Source: modied from Takahashi et al. (2011). (b) Episode of a "red tide" algal bloom (non-toxic) of Noctiluca scintillans in New Zealand. Source: M. Godfrey. Phytoplankton have a pivotal role in driving and maintaining aquatic biodiversity and the marine food web as a whole. Causative links between phytoplankton growth, abundance, species success and succession and how they are aected by the driving environmental forces need to be established. Moreover, the understanding of the regulation of phytoplankton productivity and biomass in the ocean is crucial for the understanding on eects of global climate change by radiative forcing to the ocean's food web structures and carbon uxes. In this sense, there is an important need for better information on the spatial distribution of biological activity in the upper ocean and its temporal variability, especially in the case of oceanic phytoplankton biomass. Further research is therefore needed to develop new techniques that enable to recognize phytoplankton-related ecological changes and perturbations. For all the above-mentioned compelling reasons, increasing emphasis is being placed upon scientic research and monitoring of the phytoplankton biodiversity and dynamics. In attempting to monitor spatial and temporal variations of the phytoplankton community structure at a local and global scale, in situ and remote sensing optical observation represents the most suitable, fastest and less invasive method currently available for synoptically measuring wide-area properties of ocean ecosystems. Traditional description of global phytoplankton abundance and distribution involved the monitoring of the primary pigment in phytoplankton, chlorophylla , a proxy for the phytoplankton biomass. In fact, numerous studies over the past three decades have focused on the development of bio-optical algorithms linking dierent processes to this pigment (O'Reilly et al. , 2000; Reynolds et al. , 2001). However, biogeochemical cycling in marine systems is intimately linked to the activity of specic phytoplankton groups (see Table I.1), meaning that phenomena such as the production of sulfur components like DMSP or proliferation of algal blooms are taxon-dependent (VilaCosta et al. , 2006; Stumpf et al. , 2003). Further eorts are thus needed to go beyond the estimation of only chlorophylla , which is common to all taxonomic groups. 2
Motivation Table I.1: Summary of properties of dierent phytoplankton groups, which play a specic role in the marine ecosystem and are characterized by specic bio-marker pigments. Source: generated from tables in Nair et al. (2008), Hirata et al. (2008) and Hirata et al. (2011). Mapping marine phytoplankton community structure from a hyperspectral perspective Today, in situ and remotely-sensed spectrometric optical observations of ocean waters have the potential to provide information regarding the concentrations of optically signicant constituents in seawater, in particular, regarding phytoplankton community structure. In fact, eorts to expand the use of spectrometric measurements for estimating some biogeochemically important ocean variables and phytoplankton-related phenomena have increased considerably. For example, in recent years, a variety of biooptical methods have been established (Nair et al. , 2008; Brewin et al. , 2011) that use satellite spectrometric data to identify dierent phytoplankton communities through the detection of either Phytoplankton Functional Types (PFTs, i.e., phytoplankton that have a specic biogeochemical function such as calcication, silicication, DMS production or nitrogen xation) or Phytoplankton Size Classes (PSCs, i.e., the autotrophic pool is partitioned according to cell size, which may be useful to better understand their contribution to global ocean primary production). There are two main types of approaches to derive PFTs and PSCs from remote sensing ocean color data: (1) the abundance-based approach, in which the eutrophic status of the waters, indicated by the chlorophyll concentration or the magnitude of the absorption optical coecient of phytoplankton, is related to community structure (Sathyendranath et al. , 2001; Uitz et al. , 2006; Devred et al. , 2006 and 2011; Aiken et al. , 2007; Hirata et al. , 2008 and 2011; Brewin et al. , 2010a; Mouw and Yoder, 2010) and (2) the spectral response approach, in which dierent PFTs and PSCs are detected providing they have contrasting optical signatures or traits (Sathyendranath et al. , 2004; Alvain et al. , 2005 and 2008; Ciotti and Bricaud, 2006; Bracher et al. , 3
1 Bio-optical oceanography: general background higher sunlight penetration and the runo of water from inland bringing nutrients from soil and plants, contribute to the fact that although coastal ocean comprise less than 10% of the ocean's total area, 90% of marine species inhabit this region. Coral reeefs, known for their great variety of organisms, can be found in this area. Figure 1.1: Diagram illustrating the major life zones and vertical zones in an ocean (not drawn to scale), based on the available light ( Source : Miller, 2007). Variability of light in the ocean is strongly inuenced by the distribution of the components in the water column, which varies across horizontal and vertical space and time scales. In this sense, ocean waters vary in color from the deep blue of the open ocean to yellowish-brown in a turbid region. The most optically signicant water constituents are briey described as follows 1 : Pure water Water itself limits the wavelength range of interest in optical oceanography from the near-ultraviolet to the near infrared because it is highly absorbing at wavelengths below 250 nm and above 700 nm. Inorganic particulate matter (minerals) These particles are created by weathering of terrestrial rocks and are entered the water as wind-blown dust settles on the sea surface, as rivers carry eroded soil to the sea or as currents resuspended bottom sediments. They can dominate water optical properties when present in sucient concentrations. 1 Air bubbles generated by breaking waves at sea surface can signicantly contribute to scattering properites of a water body (Stramski et al. , 2004). Nevertheless, their eect is beyond the scope of the research presented in this thesis and therefore it is not described in detail. 10
1.1 Oceanic radiative transfer and phytoplankton biology Colored Dissolved Organic Matter (CDOM or gelbsto) These compounds are produced during the decay of plant matter. In high concentrations these compounds can color the water yellowish-brown. CDOM absorbs very little in the red, but absorption increases rapidly and exponentially with decreasing wavelength. It can be the dominant absorber at the blue end of the spectrum, especially in coastal waters inuenced by river runo. Scattering by CDOM is generally considered negligible. Particulate organic matter Phytoplankton are microscopic plants (see Fig. 1.2), which occur with incredible diversity of species, cell shapes, concentrations (from less than 0.03 mgm−3 in oligotrophic waters up to about 30 mgm−3 in eutrophic waters) and sizes (3 phytoplankton size classes (PSCs): microphytoplankton from 20 to 200 µ m, nanophytoplankton from 2 to 20 µ m and picophytoplankton < 2 µ m). Figure 1.2: An astonishing diversity of phytoplankton live in Earth's oceans ( Source : Smithsonian Environmental Research Center and U.K. Fisheries Research Service). These algal organisms have an essential role in the ocean since, through the process of photosynthesis , they convert inorganic materials into organic compounds (Kirk, 1994). Phytoplankton, as primary producers of biomass, form the rst link of the oceanic food chain that reaches through all trophic levels. In order to photosynthesise, phytoplankton require carbon dioxide ( CO2 ), water ( H2O ) and light energy. The process itself occurs within the chloroplasts which contain the photosynthetic pigments that convert the radiant to chemical energy. The dominant pigment is chlorophylla which is common to all species, but there are also cholorphyllsb and - c plus several accessory pigments (i.e., carotenoids, xantophylls and the photoprotective phycobiliproteins: phycoerythrin and phycocyanin). Each of the above pigments is able to absorb light of wavelengths within the visible range with strong spectral selectivity (see Fig. 1.3). This range (approx. from 400 to 700 nm but occasionally slightly extended towards the near infrared and near ultraviolet regions) is referred to as photosynthetically available radiation (PAR) because it provides energy to marine 11
1 Bio-optical oceanography: general background phytoplankton to perform photosynthesis. In fact, the visible and near infrared (approx. from 400 to 800 nm) constitute the only portion of the electromagnetic spectrum that penetrates water and directly probes the water colum (see Fig. 1.4). The solar spectrum peaks correspond to the maximum transparency of water and the peak in phytoplankton absorption. Thus, phytoplankton photosynthesis is tuned to the spectral range of maxium light. Figure 1.3: Absorption spectra of several individual phytoplanktonic pigments in solvents. Pigment abbreviations are: Chl a = monovinyl chlorophylla , DV a = divinyl chlorophylla , Chl b = monovinyl chlorophyllb , DV b = divinyl chlorophyllb , Chl c = chlorophyllc 1+ c 2, Fuco = fucoxanthin, 19HF = 19'-hexanoyloxyfucoxanthin, 19-But = 19'-butanoloxyfucoxanthin, Zea = zeaxanthin, Peri = peridinin ( Source : Taylor et al. , 2011). Figure 1.4: Solar irradiance, phytoplankton specic absorption and water absorption spectra ( Source : McClain, 2001). 12
1.1 Oceanic radiative transfer and phytoplankton biology Each phytoplankton group dier signicantly with respect to motility, cell-wall and pigment composition, likewise to nutrition and reproduction strategies. At the same time, all species are at the mercy of oceanic currents for transport to areas that are suitable for their survival and growth. Thus, physical processes and light conditions for phytoplankton photosynthesis play both a signicant role in determining the composition and distribution of phytoplankton species. Phytoplankton, with a great diversity of species, are responsible for determining the optical properties of most oceanic waters. Bacteria are the smallest living organisms on Earth (i.e., size range 0.2 - 1.0 µ m). They can be signicant scatterers and absorbers of light at blue wavelengths in oceanic waters, where the larger phytoplankton are relatively scarce. Detritus are non living organic particles of various sizes that are produced when phytoplankton die and their cells break apart or when zooplankton graze on phytoplankton and leave cell fragments and fecal pellets. They have signi- cat absorption only at the blue wavelengths. They can, however, contribute signicantly to scattering in the open ocean. Having identied the most optically signicant water consituents, which are responsible for diversity in color of oceanic waters (see some examples in Fig. 1.5), a bipartite classication scheme for oceanic waters has traditionally been used in optical oceanography. It was introduced by Morel and Prieur (1977) and rened later by Gordon and Morel (1983) and Sathyendranath and Morel (1983). A pictorical representation of the two cases (i.e., Case 1 and Case 2) is presented in Fig. 1.5. Moreover, the component absorption and scattering coecients for two hypothetical water bodies representing each of them are depicted in Fig. 1.6. Figura 1.5: Examples of dierent colored sea areas and diagram representing the Case 1 and Case 2 waters. Case 1 waters represent the phytoplanktondominant cases, whereas Case 2 waters represent all other possible cases ( Source : Sathyendranath, 2000). 13
1 Bio-optical oceanography: general background `Case 1 waters' are those in which the contribution by phytoplankton to the total absorption and scattering is high compared to that by other substances. Therefore, absorption by chlorophyll and related pigments plays the dominant role in determining the total absorption in such waters, although covarying detritus and dissolved organic matter derived from the phytoplankton can also contribute to absorption and scattering in these waters. Case 1 waters can range from very clear (i.e., oligotrophic, phytoplankton-poor) to very productive (i.e., eutrophic, phytoplankton-rich) water, depending on the phytoplankton concentration. On the other hand, `Case 2 waters' are `everything else,' namely, waters where inorganic particles in suspension or colored dissolved organic matter (CDOM) from land drainage contribute signicantly to the absorption and scattering. In this sense, the eect of pigments is relatively less important in determining the total absorption. Roughly 98% of the world's open ocean and coastal waters fall into the Case 1 category, but near-shore and estuarine Case 2 waters are disproportionately important to human interests such as recreation, sheries and military operations. Even this classication scheme cannot be implemented in a very strict quantitative fashion, its value lies most in its use as a simple device to dierentiate waters where phytoplankton-related features dominate the signal (see Chapters 2, 3 and 4) from more optically complex water bodies, where such simplifying assumptions would not hold (see Chapter 5). Figura 1.6: Absorption and scattering coecients for the Case 1 and 2 water bodies and the contribution by each water consituent ( Source : Mobley, 2001). 14
1.1 Oceanic radiative transfer and phytoplankton biology For the Case 1 scenario, the total absorption is dominated by phytoplankton (i.e., chlorophyll-bearing particles) at blue wavelengths and by the pure water at wavelengths greater than 500 nm. The total scattering is also dominated by phytoplankton. In the Case 2 scenario, mineral particles present an absorption comparable to or greater than that by the chlorophyll-bearing particles and are the primary scatterers. In both Case 1 and Case 2 waters, oceanographic optical measurements can be important indicators of the health of the ocean in the form of changing turbidity together with changing diversity and distribution of phytoplankton species. In this sense, the assessment of optical properties of the ocean can help elucidating the link between the phytoplankton communities and several biogeochemical processes in which phytoplankton play a key role. For instance, as it has been described in the thesis introduction (see Fig. I.1), practical applications of knwoledge of plankton diversity and distribution in the oceans include ecosystem responses to climate change (including the carbon and sulfur cycle), detection of harmful algal blooms in coastal waters, food web modelling and management of marine resources. Knowledge of the variability of bio-optical properties of the ocean is essential for understanding all these processes. The subdiscipline of ocean optics , which concerns the assessment of the propagation of light through the oceanic water column and surface, has become fundamental for understanding water dynamics and composition, including phytoplankton communities distribution. This subdiscipline is based on the radiative transfer (RT) theory (Kirk, 1994; Mobley, 1994), utilized as the theoretical framework and dened to better comprehend light propagation in the optically complex ocean. Radiative transfer is therefore a complex eld that requires physical understanding of the absorption and scattering processes in the atmosphere and oceans, as well as sea surface and bottom reection characteristics (see Fig. 1.7). In this sense, optical properties of sea water are divided into two classes: Inherent Optical Properties (IOPs) are those properties that are only affected by the constituents of the aquatic medium and are independent of the ambient illumination conditions. Examples of IOPs are the spectral absorption and scattering coecients. IOPs are additive, which means that they result from additive contributions associated with all individual water components that absorb and scatter light (see Fig. 1.6). Therefore, their variation is generally attributed to variability in four constituents of the aquatic medium in natural oceanic waters: pure sea water, phytoplankton, non-algal particles and colored dissolved organic matter (CDOM). Apparent Optical Properties (AOPs) are not additive and depend both on the medium (i.e., on the IOPs) and on the directional structure of the ambient light eld. Examples of AOPs are the spectral radiance and irradiance distributions. 15
1 Bio-optical oceanography: general background Historically, the IOPs were easy to dene and interpret in terms of water constituents, but hard to measure. This is, however, less true today because of advances in instrumentation. In contrast, the AOPs are much easier to measure and readily available, but are often more dicult to interpret because of the confounding environmental eects. For instance, a change in the state of the sea surface or in the sun's position modies the spectral radiance distribution, and hence the AOPs, even though the IOPs are kept unchanged. Figura 1.7: Depiction of the various optical pathways and interactions within the atmosphere-ocean system. Radiative transfer theory provides the connection between the IOPs and the AOPs of a water body and is expressed as a mathematical structure by the Radiative Transfer Equation (RTE) . The physical environment of a water body waves on its surface, the character of its bottom, the incident radiance from the sky enters the theory via the boundary conditions necessary for solution of the equations arising from the theory. Figure 1.8 summarizes the relationships between the various radiometric quantities, including IOPs and AOPs. Note that all of them have wavelength dependence. A detailed description of the oceanic RTE equation and its link with some of the optical properties is given in the Appendix A. Nevertheless, the denition of the most important optical properties necessary to understand the results presented in the following chapters (and also shown in Fig. 1.8) is provided below. 16
1.1 Oceanic radiative transfer and phytoplankton biology Figura 1.8: Relationships among the various optical quantities commonly used in hydrological optics, which are dened below ( Source : Mobley, 1994). Inherent Optical Properties (IOPs) IOPs can be dened by using the diagram shown in Fig. 1.9, which illustrates a small volum ( 4V ) of water of thickness 4r , that is illuminated by a collimated beam of monochromatic light wavelength λ and spectral radiant power Φi(λ) [Wnm−1] . Part of the incoming beam of light is absorbed, Φa(λ) [Wnm−1] , and the other part is scattered, Φs(λ) [Wnm−1] , from the incident to other directions. 17
1 Bio-optical oceanography: general background Figura 1.9: Geometry used to dene the Inherent Optical Properties (IOPs) of a water body ( Source : Mobley, 1994). Absorption coecient - a(λ) is the absorbance (i.e., the ratio of the absorbed to incident spectral radiant ux) per unit distance: a(λ) = lim∆r→01 Φi(λ) Φa(λ) ∆rm−1 (1.1) Volume scattering function (VSF) - β(ψ, λ) is the radiant ux per unit solid angle ( ∆Ω ) scattered in the ψ direction with respect to the direction of the incident light, per unit distance, expressed as a proportion of the incident ux: β(ψ, λ) = lim∆r→0lim∆Ω→01 Φi(λ) Φs(ψ,λ) ∆r∆Ω m−1sr−1 (1.2) Scattering coecient - b(λ) is obtained by integrating β(ψ, λ) over all directions: b(λ) = ´4πβ(ψ, λ)dΩ=2π´π 0β(ψ, λ) sin ψdψ m−1 (1.3) Beam attenuation coecient - c(λ) is the total loss of radiant energy from the original beam, which is the sum of absorption and scattering coecients: c(λ) = a(λ) + b(λ)m−1 (1.4) Backscatter coecient - bb(λ) is obtained when restricting the integration of β(ψ, λ) to the interval π/2≤ψ≤π : bb(λ)=2π´π π/2β(ψ, λ) sin ψdψ m−1 (1.5) Single-scattering albedo - wo(λ) is the ratio of scattering eciency to total attenuation eciency: w0(λ) = b(λ) c(λ) (1.6) 18
1.1 Oceanic radiative transfer and phytoplankton biology Volume scattering phase function - e β(ψ, λ) provides the angular distribution of the scattered light and is obtained dividing β(ψ, λ) by b(λ) : ˜ β(ψ, λ) = β(ψ,λ) b(λ)sr−1 (1.7) It is important to note that in the preceding denitions, no inelastic scattering processes (e.g. uorescence by phytoplankton or CDOM and Raman scattering by the water molecules) were considered. This factor appears as a source term ( S(z, θ, ϕ, λ) ) in the radiative transfer equation (see Appendix A for further details). Moreover, due to the additive property, the total IOPs of a water body (i.e., a(λ), b(λ), β(ψ, λ), bb(λ) and c(λ) ) are the sums of the IOPs associated with each of the various components present in this water body. In this sense, for instance, atotal(λ) = aw(λ) + aph(λ) + anap(λ) + acdom(λ) where: w = pure water, ph = phytoplankton, nap = non-algal particles such as detritus or minerals and cdom = colored dissolved organic matter (see the examples for absorption and scattering coecients in Fig. 1.6). Fundamental radiometric quantities in ocean optics Spectral radiance - L(z, θ, ϕ, λ) , as shown in Fig. 1.10, is the radiant energy incident in a time interval ( ∆t ), onto a surface of area ∆A located at a position ( x, y, z ) and arriving through a set of directions contained in a solid angle ( ∆Ω ) about the direction (zenith angle θ and azimuthal angle ϕ ) normal to the area ∆A and produced by photons in a wavelength interval ∆λ : L(x, y, z, θ, ϕ, λ) = ∆Q ∆t∆A∆Ω∆λWm−2sr−1nm−1 (1.8) Figura 1.10: Geometries used to dene the spectral radiance ( Source : Mobley, 1994). Scalar spectral irradiance - Eo(z, λ) is obtained by integrating the radiance over all directions that reaches a certain surface. It is commonly measured using a spherical collector (see Fig. 1.11) and as the sum of the downwelling and upwelling components: 19
1 Bio-optical oceanography: general background Other molecular techniques that are beyond the scope of this thesis exploit genetic variations to distinguish between phytoplankton groups and provide a solution to the limitations encountered with HPLC. DNA sequencing techniques have opened avenues to distinguish organisms at all taxonomic levels, from the level of classes to ecotypes (Rocap et al. , 2003). Despite of the advances, probes are not available for all possible phytoplankton functional types (PFTs) and specicity of probes remains an area of ongoing research. Bulk particle analysis Secchi disc and Forel-Ule scale are two simple and fast methods that were described at the end of the 19 th century to establish the ecological state of surface water (Wernand, 2011). The optical properties of seawater such as color and transparency in water have been measured for a long period: color through the Forel-Ule scale and transparency through the Secchi disc. The Forel-Ule Scale is a method to approximately determine the color of a water body by comparison with a standard color palette produced with mixtures of colored chemical solutions in a series of numerically designated vials (1-21). The Secchi disc is a circular white disk 30 cm in diameter, which is lowered over the side of a ship. The depth at which it is no longer visible is known as the Secchi depth and has been used as a rough measure of optical clarity or transparency. It is noted that these methods depend on the eyesight of the observer, thus, the major problem is the lack of absolute calibration. In addition, these methods are just helpful to classify gross phytoplanktonic activity of a water body. Spectrometric techniques Traditionally, as described in almost all previous methods, phytoplankton species discrimination has involved exhaustive and time consuming eldwork, implying a ship-based approach to collect discrete water samples for later laboratory analysis to retrieve taxonomical information. These traditional methods often require ltering of discrete samples or even solvent extractions, which are time consuming and challenging to use when at sea. After the introduction of spectral radiometers to in situ and remote sensing oceanic observing platforms, the extrapolation of such information became easier, faster, much less invasive and usually non-destructive. In fact, by directly measuring the amount of light that is absorved, scattered or reected from a bulk, which is a function of its pigment and chemical composition, one can possibly distinguish the phytoplankton composition of it. Scientists are now engaged in developing inversion methods to obtain biogeochemical products from these type of measurements. Next, a description of several in-water spectrometric-based technologies is provided: Measurement of Inherent Optical Properties Recent advances in optical instrumentation and methodologies now enable the in situ measurement of some IOPs (see Fig. 1.14). These instruments generally 26
1.2 From single to bulk particle measurements employ an active light source given that IOP measurements are independent to the ambient radiance eld. Therefore, they can be used at any time of day and are not subject to variable cloud or surface conditions and shadows from the observing platform or the instrument itself. The derivation of absolute coecients of optical transfer properites is performed over relatively short path lengths and with a resolution of in-water variability at scales ranging down to a few centimeters. Beam attenuation meters (also known as transmissometer s) operate upon the principle by which the attenuation (i.e., loss of light due to both absorption and scattering) of a narrow, well collimated beam of light through a path length of water is measured (Fig. 1.14a). Modern transmissometers for in situ ocean research operate at single or multiple wavelengths (Moore et al. , 2004a) and are used to estimate species composition, particulate organic carbon, suspended particles and visibility (Claustre et al. , 2008; Zaneveld and Pegau, 2003). Absorption meters measure the light loss through a given path of water due only to the molecular absorption of the water and the components contained therein. The conceptual framework behind underwater absorption meters is based on minimizing measurement losses due to highly scattering media found in natural waters. In order to retrieve the absorption component, dierent designs have been developed: (1) Refractive tube absorption meters use a collimated beam propagating through a xed path surrounded by a reective tube and impinging upon a large area detector (Fig. 1.14b). Commercial units incorporating this design are available in multispectral and hyperspectral con- gurations (Zaneveld et al. , 2004). (2) The Point Source Integrating Cavity Absorption Meter (PSICAM) was designed to detect absorption spectra in an integrating cavity sphere lled with the water sample. The principle is based on multiple reection and scattering, i.e. the light eld is diused inside the cavity (Röttgers et al. , 2005). In comparison to the refractive tube meters, it reduces the scattering eects on absorption measurements (i.e., the overestimation of the sample absorption resulting from backscattering). Submersible versions of this design have been developed by Dana and Maone (2006) (Fig. 1.14c), likewise instruments built for operation on Autonomous Underwater Vehicles (Kirkpatrick et al. , 2006). It is important to note that there exist similar approaches that measure the absorption coecient by planktonic algae contained in a bulk but performed in the laboratory. These approaches deserve a mention since they also represent an essential source of information in terms of phytoplankton composition and have been used within the framework of this thesis (see Section 3.1.1.2 and 4.1.3). They are performed by combining light-transmission and light-reection measurements of the algae previously concentrated onto a lter (i.e., the lter pad and T-R techniques, see Tassan and Ferrari, 1995) and using an integratingsphere attached with a dual-beam spectrophotometer. In order to minimize the photon loss by backscattering, another solution has been developed in which 27
1 Bio-optical oceanography: general background the sample is placed in the center of the integrating sphere (Simis et al. , 2005). In all these cases, as the lter pad technique is used to prepare the sample, it has to be considered that multiple scattering in the lter increases the optical path length. Therefore, a correction for the path length amplication factor has to be applied to convert the measured absorption of the algal particles on the lter to the absorption coecient of the algal particles in suspension. Figure 1.14: Schematic diagrams for IOP sensors. (a) Beam attenuation meter or transmissometer. (b) Reective tube absorption meter. (c) Submersible spherical cavity absorption meter. (d) Backscattering sensor ( Source : Moore et al. , 2009a; Dana and Maone, 2006). Backscattering meters are used to measure light scattered in the backward direction (i.e., between 90 and 180 degrees). The backscattering coecient is important in understanding remote sensing signals seen for instance by satellites. These sensors are most widely used in applications focused upon validating ocean color measurements from satellites or also to retrieve information regarding the particle size distribution of a bulk, including phytoplankton. Optical backscattering sensors have been congured for measurement typically between 115 and 145 degrees (Fig. 1.14d) and multispectral congurations (Lee and Lewis, 2003; Moore et al. , 2000). Fluorescence sensors are used to estimate concentrations and provide invaluable physiological information from phytoplankton materials. Due to its cellular pigment composition, phytoplankton species are characterized by specic excitation wavelengths triggering specic emission wavelengths. Fluorometers have been congured for in situ single channel detection, which provide reasonable proxies for phytoplankton biomass. Moreover, multichannel uorometers have 28
1.2 From single to bulk particle measurements been developed to detect spectral excitation and/or emission uorescence signatures, which can be used to infer taxonomic composition and identify specic phytoplankton types (Moore et al. , 2004b; Chekalyuk et al. , 2006). Measurement of Apparent Optical Properties Sensors measuring AOPs are in general passive since these properties vary as the angular radiance distribution varies. It implies that they use as their light source the incident solar beam directly transmitted and scattered by the atmosphere, the sea surface and the ocean interior. In this sense, passive detection of sunlight results in relatively low energy consumption and implies that AOP sensors measure light levels experienced by the surrounding biota. Radiance and irradiance sensors are used to determine and derive all the AOPs. They can be dierentiated by the angular integration performed by the frontend optics, which is the component that captures the ambient light eld. Narrow angular eld-of-view (FOV) collectors are used to measure radiance, while plane collectors provide cosine weighting to yield plane irradiance, and spherical collectors, which weight all directions equally, measure scalar irradiance (see diagrams in Fig. 1.10 and 1.11). These sensors have been manufactured in multispectral and hyperspectral congurations to provide full spectral distribution over the ultraviolet, visible and near infrared bands. The spectral dispersion to separate the broadband eld into narrow spectral intervals is accomplished by prism or grating elements. For some applications, spectral weighting lters have also been employed to return the integral energy over the entire visible spectrum and are known as irradiance PAR sensors. Radiance and AOP sensors are now operationally for applications including the determination of phytoplankton biomass. For instance, it is estimated via the measurement of the in-water irradiance, the water leaving radiance or the reectance at the sea surface. These signals are inuenced by the absorption and scattering properties of the water column and hence, by the red uorescence (approx. at 680 nm) resulting from phytoplankton activity. It is also under study the identication of dierent phytoplankton functional groups (PFTs) from the in-water and remote-sensing radiance characteristics and related AOPs (Nair et al. , 2008). These instruments are also used routinely for ocean color applications such remote sensing validation and calibration and for capturing optical changes associated with some phytoplankton-related episodic events (e.g., harmful algal blooms). As it is emphasized below, the increasing use of high spectral resolution (hyperspectral) sensors is being crucial in all these applications, in particular for phytoplankton species monitoring (see Section 1.3). It is important to note that protocols for eective use of instruments measuring IOPs and AOPs have been developed during the last years (see Mueller et al. , 2003b; Mueller et al. , 2003a and references found therein). For instance, there exist international standards of radiance and irradiance measurements for ocean 29
1 Bio-optical oceanography: general background color applications, which provide a reference to which instruments anywhere in the world (i.e., including onboard space-based platforms) and at any time can be calibrated and intercompared with a high degree of condence. Other approaches that are beyond the scope of this thesis include the use of holographic instruments and confocal technologies for the 3D in-situ visualization of plankton (Malkiel et al. , 1999; Matsumoto, 1993). 1.3 From multito hyperspectral applications It has been described how phytoplankton community composition in the ocean may be assessed using a wide range of optical-based techniques in the laboratory and in the eld. Due to the faster and less invassive characteristics, however, a noticeable increase has been in the number of coastal and open ocean studies based on spectrometric measurements. In situ and remotely-sensed spectrometric observations of ocean waters (i.e., IOPs and AOPs) provide information regarding the concentrations of optically signicant constituents in seawater, and oer the ability to observe important biological and biogeochemical variables (e.g., Chang et al. , 2006). Numerous studies over the past three decades have focused on the development of bio-optical algorithms linking measurable spectral optical properties to the primary pigment in phytoplankton, chlorophylla , a proxy for the phytoplankton biomass and primary production (e.g., Bricaud et al. , 1998; Morel, 1988; O'Reilly et al. , 2000 and the illustrating example shown in Figure 1.15 from Reynolds et al. , 2001). Current global algorithms use simple wavelengthratios of reectances to retrieve chlorophyll concentrations, such as the NASA OC2 version 2 algorithm, also depicted in Fig. 1.15 and the NASA OC4 version 4 algorithm, both developed by O'Reilly et al. (2000). In particular, the OC4 algorithm is a four-band maximum band ratio formulation that uses a fourth order polynomial to estimate chlorophyll concentration and is given by the relation: [Chl] = 10(0.366−3.067X+1.93X2+0.649X3−1.532X4) (1.16) where X=log(Rrs(443)/Rrs(555) > Rrs(490)/Rrs(555) > Rrs(510)/Rrs(555)) . More recently, eorts to expand the use of spectrometric measurements for estimating other biogeochemically important ocean variables and phytoplankton-related phenomena have increased considerably. For example, spectrometric measurements including satellite remote sensing have been used to detect harmful algal blooms (Cullen et al. , 1997; Stumpf et al. , 2003), surface concentrations of particulate inorganic and organic carbon (Balch et al. , 2005; Stramski et al. , 2008), particle size distribution (Kostadinov et al. , 2009), phytoplankton community composition and size structure (Aiken et al. , 2007, 2005; Ciotti and Bricaud, 2006; Nair et al. , 2008; Uitz et al. , 2006), and class-specic primary production (Uitz et al. , 2010). It is noteworthy, however, and in accordance with the central topic of this thesis, that recent advances in measuring ocean spectrometric properties and light elds 30
1.3 From multito hyperspectral applications within and leaving the ocean have included a progressive shift from using multispectral to high spectral resolution (hyperspectral) acquisition systems (Chang et al. , 2004). The capability to obtain measurements at hundreds of narrow and closely spaced wavelength bands from the ultraviolet to near-infrared, with a resolution better than 10 nm, has become one of the most powerful and fastest growing areas of technology in the eld of ocean optics. The rapid maturing of optical instrumentation has led to hyperspectral technology, which has opened the possibility for optical oceanographers to more accurately characterize complex oceanic environments. Figure 1.15: (a) Particulate absorption coecients at particular wavelengths (i.e., 490 and 555 nm) as a function of chlorophyll concentration for dierent stations within the Southern Ocean. Solid and open symbols represent surface data from the Ross Sea and Antarctic Polar Front Zone, respectively. The lines illustrate the predicted relationship from the model developed by Bricaud et al. (1998). (b) Modeled relationships between the spectral band ratio of remote-sensing reectance, Rrs(490)/Rrs(555) , and chlorophyll obtained in the same locations shown in (a). The lines inllustrate the predicted relationship from dierent models, among them the NASA OC2 version 2 algorithm developed by O'Reilly et al. (2000) ( Source : Reynolds et al. , 2001). The advantage of a hyperspectral over a multispectral inversion for phytoplankton species identication stems from the fact that more accurate spectral information, such as spectral features related to characteristic phytoplankton pigment absorption peaks, is resolved. This is a key factor in the identication of phytoplankton types because it relies on very small signal and changes in the shape of spectral optical characteristics. In the past, analyses of single band or single band-ratios obtained in discrete multispectral bands were employed to retrieve concentrations of the pigment chlorophylla (O'Reilly et al. , 2000, see Fig. 1.15), resolve the presence of specic phytoplankton functional groups (Sathyendranath et al. , 2004; Alvain et al. , 2005) or determine phytoplankton size classes (Hirata et al. , 2008). However, in most studies phytoplankton species mapping failed or was limited to the identication of some 31
1 Bio-optical oceanography: general background phytoplankton groups given the limited number of spectral bands. The multispectralbased approach usually implied a lack of spectral resolution necessary to dierentiate more species in the ocean. In this sense, the advent of hyperspectral sensors provides ample spectral information and facilitates to go beyond the estimation of the chlorophylla and identify more phytoplankton species, which are spectrally unique and complex. For instance, Craig et al. (2006) assessed the feasibility of remote detection and monitoring of the toxic dinoagellate, Karenia brevis , accounting for complete spectral behavior of in situ hyperspectral measurements of remote-sensing reectance, Rrs(λ) and Lohrenz et al. (2008) directly explored the hyperspectral patterns of Rrs(λ) to better characterize water mass properties in coastal areas (i.e., Case 2 waters). More recently, the composition and concentrations of two major phytoplankton functional groups (i.e., diatoms and cyanobacteria) have been derived from hyperspectral measurements of the satellite sensor SCIAMACHY (Scanning Imaging Absorption Spectrometer for Atmospheric Chartography) on ENVISAT and analysed with the PhytoDOAS method (Bracher et al. , 2009; Sadeghi et al. , 2012). As for what this thesis concerns, all hyperspectral-based applications developed up to now and the increasing availability of hyperspectral observations of ocean waters clearly lead to a need for ongoing exploration of its potential to better map phytoplankton community composition and dynamics. Expanding role of hyperspectral oceanography New technologies and the miniaturization of electro-optical components have permitted the development of accurate, low-cost and energy-ecient hyperspectral sensors designed to measure oceanic high spectral resolution optical properties (i.e., including IOPs and AOPs). As a result, their use is no longer limited to bench-top applications but they are being deployed on fully edged sea-going observational platforms. New interdisciplinary ocean observing platforms incorporating hyperspectral sensors are being deployed such as in-water vertical proling systems, moorings, drifters, autonomous vehicles, air-borne and space-borne platforms (Dickey et al. , 2006; Twardowski et al. , 2005; Perry and Rudnick, 2003). Various satellite-based hyperspectral imagers are currently available: NASA's Hyperion sensor on the EO-1 satellite (Fig. 1.16a), ESA's CHRIS sensor on PROBA satellite or US Air Force Research Lab's FTHSI sensor on the MightySat II satellite. In addition, examples of existing hyperspectral airborne sensors providing high spatial resolution images of specic areas are the Ocean Portable Hyperspectral Imager for Low-Light Spectroscopy (PHILLS, Davis et al. , 2002), the Compact Airborne Spectrographic Imager (CASI) or the Airborne Visible Infrared Imaging Spectrometer (AVIRIS). Other in situ ocean observing platforms for which hyperspectral sensors may be suitable are ships and moored buoys (Fig. 1.16b, Kuwahara et al. , 2007). These platforms oer continuous measurements at a high spatial and temporal resolution delivering data even under cloudy conditions, thereby complementing discrete observations by satellites and aircrafts and providing the data essential for calibration and validation purposes. Furthermore, Lagrangian platforms that follow a particular water mass (e.g. oats and drifters) 32
1.3 From multito hyperspectral applications can provide hyperspectral measurements in oceanic regions not normally accessible by satellites or oceanographic research vessels. The goal of these monitoring missions is to determine the material composition of the water mass under investigation through the accurate analysis of simultaneous optical and hydrographic measurements. The number of newly emerging ocean observing platforms capable of incorporating hyperspectral instrumentation is continuously growing. For instance, new autonomous underwater vehicles (AUVs) have been developed recently with integrated equipment for adaptive sampling and the ability to perform a wide-range of preprogrammed monitoring surveys over extended periods of time (Perry and Rudnick, 2003). Another member of this family of instruments are the special AUVs called gliders that propel themselves through the water by changing their buoyancy with a minimal expenditure of energy (Fig. 1.16c). The glider provides a wide spatial coverage measuring among others - the optical properties during periods of three or four weeks (Wood, 2009). Another approach to monitor the variability of optical properties in the ocean is the use of vertical proling systems (Fig. 1.16e, f) which can, for instance, be part of some of the emerging cabled observatories. Instrument platforms such as the Vertical Proler System (VPS), part of Neptune Canada (Barnes et al. , 2008), the world's rst regional-scale underwater ocean observatory plugged directly into the Internet, oer great power and bandwidth for gathering large quantities of hyperspectral data. Figure 1.16: Examples of ocean observing platforms suitable to incorporate hyperspectral sensors. The amount of hyperspectral data sets available collected using all these types of oceanographic observing platforms, covering a wide range of temporal and spatial scales, will increase in the near future. New interdisciplinary research initiatives are being successfully carried out such as the Hyperspectral Coastal Ocean Dynamics 33
1 Bio-optical oceanography: general background Experiment (HYCODE, see Pegau, 2002) conceived to exploit the new capabilities of hyperspectral ocean color sensors. Within this framework, several collaborative short-term and long-term eld experiments, involving the simultaneous use of hyperspectral devices deployed in situ and remotely, are being conducted for calibrating, sea truthing and relating subsurface optical properties to remote sensing ocean color measurements. These investigations are essential to develop and validate optical radiative models (see next Section) and to further our understanding of the processes that contribute to the temporal and spatial variability of IOPs and AOPs in the ocean. One of the most important challenges of HyCODE experiment is the standardization of the quality control protocols, including the denition of procedures to perform detailed analysis of the reliability of measurements and an evaluation of the impact of instrumentations inaccuracies. 1.4 Radiative Transfer (RT) models In situ and remote sensing hyperspectral measurements in the ocean might be the two methods providing sucient information on phytoplankton distribution in their spatial and temporal variability. However, collecting signicant amount of hyperspectral data from such a high dynamic environment and under fully-controlled conditions is still not straightforward. In fact, there are currently no hyperspectral sensors that provide validated observations with a high temporal and spatial coverage of extended marine areas. In this sense, there are still several limitations in the applications where multispectral and hyperspectral data is used. For instance, despite current global remote sensing products (that rely on algorithms based on spectral ratios of the remote-sensing reectance for the operational determination of chlorophylla , O'Reilly et al. , 2000) benet from reduction in uncertainties partially cancelled out through ratios, residual uncertainties may still become the source of large uncertainties in derived products. These uncertainties are related to error estimates for atmospheric correction or to errors in the performance of the sensors themselves and are a limiting factor for the successfull development of the satellite-derived products. One approach to nding a solution to these problems and exploring the feasibility of hyperspectral oceanographic observations to detect phytoplankton functional and toxic groups is using radiative transfer (RT) modelling. On the one hand, model-based approaches may be useful to show the feasibility of dierent optical measurements and their results for drawing the attention to the possibilities of how the problems that might arise could be solved. On the other hand, and as it is shown in this thesis, optical modelling can be an optimal and essential tool for developing and testing new methods for estimating variability and dynamics of phytoplankton communities. In any case, it is noteworthy that models are only useful for these applications when we have good knowledge on the phytoplankton properties and other marine components. The use of in-situ measurements of in-water properties, together with theoretical RT models and laboratory experiments is the most suitable combination to validate in situ and remote sensing observations and investigate biological activity, marine optical 34
1.4 Radiative Transfer (RT) models properties, and changes in the concentration and composition of material in surface waters, over larger temporal and spatial scales (Tzortziou et al. , 2006). In this thesis, the commercially available RT numerical model Hydrolight/Ecolight 5.0 (Sequoia Scientic, Inc., Mobley and Sundman, 2008) has been utilized. Timeindependent radiance distributions within and leaving any plane-parallel water body (and derived quantities) have been computed as a function of depth, direction and wavelength given water column absorbing and scattering properties (i.e., the IOPs) and other oceanographic environmental conditions (see Fig. 1.8 for relationships between the various radiometric quantities). The Hydrolight/Ecolight code employs mathematically sophisticated invariant imbedding techniques to solve the radiative transfer equation (see Appendix A for details on this equation) and oers the possibility of performing numerical simulations in controlled environments (e.g., Mobley and Stramski, 1997 and Albert and Mobley, 2003). When computing the full radiance distribution, invariant imbedding is computationally extremely fast compared to other solution methods such as Monte Carlo simulations. Hydrolight/Ecolight 5.0 has been used in a variety of studies ranging from biooptical oceanography to remote sensing. In this thesis, the IOPs of dierent water types have been used to simulate dierent in-water light scenarios (each characterized by an hyperspectral optical signature) with the purpose of evaluating the feasibility of hyperspectral oceanographic observations to better map phytoplankton communities. As it is described in each chapter, modeled, laboratory and in situ data were collected and considered in the simulations performed for each of the presented studies. 35
2 Model-based assessment to use hyperspectal data in monitoring phyto. communities 2.1.2 Hierarchical cluster analysis As it has been mentioned, several of the studies dedicated to derivative analysis of hyperspectral remote-sensing reectances have traditionally focused on the use of ratios of discrete values of derivative spectra to characterize or classify oceanic environments (Lubac et al. , 2008; Louchard et al. , 2002). However, more advanced and contemporary methods are beginning to take advantage of the information contained in the whole derivative spectrum. For example, Filippi (2007) proposes a combination of derivative spectroscopy and articial neural network (ANN) algorithms. The derivative-neural approach has been proven to be eective for providing bathymetry, bottom type and constituent concentration estimations from Rrs(λ) measurements. In this chapter, the assessment of similarity between hyperspectral optical spectra has been made using a similarity-based cluster algorithm. This type of technique accounts for complete spectral behavior of optical data with no need to identify specic spectral features (Pekalska and Duin, 2000). This approach is particularly useful in optical oceanography when the global contribution over the entire spectral range of all pigments present in the water sample wants to be considered (Duin et al. , 1997). A method based on a hierarchical cluster analysis (HCA) has been applied using hyperspectral optical data as input vectors (or objects). The HCA method utilizes an unsupervised classication algorithm which creates a hierarchical cluster tree (or dendrogram) by partitioning a given collection of unlabeled input data into clusters or groups of objects (Berkhin, 2006; Jain et al. , 1999). Each cluster includes objects that have a similar pattern between them and a dissimilar pattern from objects distributed in the other clusters. Figure 2.4 illustrates the process of creation of each cluster tree in which a linkage algorithm is utilized based on initial calculations of a pairwise distance between all objects included in the input data set. Figure 2.4: Diagram of stages in unsupervised hierarchical cluster analysis (HCA). Hyperspectral optical data (or derivative) are utilized as input. Data in each cluster of a dendrogram share a proximity according to some dened distance measure. The selected distance measure determines how the similarity of two input objects is calculated. In this study, the partitioning of data into clusters 42
2.2 Results and discussion has been based on the determination of an angular distance between each pair of examined input data objects (i.e., each pair of reectance or absorption spectra). In particular, the similarity between each pair of objects has been evaluated using the cosine distance, d , calculated as one minus the cosine of the angle θ between each pair of objects: d(x1, x2) = 1 −cos θ = 1 −x1·x2 ||x1||×||x2|| (2.3) where x1 and x2 include the two considered input data objects and the cosine of the angle between the objects ( θ ) is obtained as the ratio of the dot product of the objects to the product of norms of the objects. Note that as the angle between the objects decreases the cosine approaches 1, resulting in a smaller distance between the input data objects and therefore higher similarity. Other measures of similarity between the input objects have been tested, e.g., Euclidean distance using a similar approach to that proposed in Robila (2005). Finally, the cosine distance has been selected as the most appropriate measure for this study because it reects mainly the dierences in the spectral shape of optical data rather than magnitude. The cosine distance is also advantageous because it is scale invariant, i.e., insensitive to normalization of optical spectra at a specic wavelength. As a linkage algorithm, the shortest distance D , also referred to as the nearest neighbor, has been computed to measure the distance between two clusters of objects in the tree: D(a, b) = min [dist(xai, xbj)] i (1, . . . , na)and j (1, . . . , nb) (2.4) where xai is the i th object in cluster a and xbj is the j th object in cluster b . In the traditional graphical representation of a dendrogram, the individual objects appear at one end and a single cluster containing all objects at the other end (e.g., Jain et al. , 1999). In this study, pairs of objects showing a small cosine distance between them (because of a similar pigment composition and spectral optical pattern) will provide a small linkage distance and therefore appear closer in the cluster tree. 2.2 Results and discussion 2.2.1 Experimental design The potential oered by hyperspectral observations to identify phytoplankton communities in open ocean waters is investigated based on the analysis tools described in the previous section (i.e., derivative spectroscopy and HCA). A simulation-based framework is used to achieve this goal as a primary approach, which includes the use of the Hydrolight/Ecolight version 5.0 radiative transfer (RT) model (Mobley and Sundman, 2008). As one of the main interests of this study lies in a methodology 43
2 Model-based assessment to use hyperspectal data in monitoring phyto. communities for estimating variability in phytoplankton communities from remote-sensing optical observations, this study is based on a set of remote-sensing reectance, Rrs(λ) , spectra. A brief description of the methodology proposed and carried out in this research work is provided here (and schematically presented in Figure 2.1). As a rst stage, the generation of a set of simulated hyperspectral Rrs(λ) spectra is performed using a RT oceanic model, which takes into account the constituents present in the water column, their inherent optical properties (i.e., absorption and scattering properties of each constituent) and their concentration prole. The next stage encompasses the computation of the second derivative spectrum of each modeled Rrs(λ) . Previous to derivative analysis, however, a mean lter smoothing type is also applied to Rrs(λ) data. It consists of a simple average of points within the chosen lter window. As stated in Section 2.1.1, the ecient way to accomplish the smoothing and derivative processing is carefully adapting the lter size ( WS ) and the sampling interval ( BS ) in order to better match the scale of the spectral features of interest in each case. As a nal stage, the goal of validating the potential of hyperspectral Rrs(λ) measurements for identifying dierent phytoplankton communities is addressed through a similarity assessment between Rrs(λ) spectra and also between second derivative spectra of Rrs(λ) . To make such a similarity assessment, an approach based on a hierarchical cluster analysis is used (see Section 2.1.2). Spectra corresponding to open ocean scenarios with a similar phytoplankton composition (i.e. with a similar spectral pattern) are expected to appear closer in the dendrogram than those having a dierent phytoplankton composition. The feasibility of considering hyperspectral measurements of Rrs(λ) to identify phytoplankton communities is assessed by analyzing how close ordinary Rrs(λ) spectra (or derivative spectra) dominated by the same phytoplankton group appear in the computed cluster tree. 2.2.2 Modeling of underwater optical scenarios A synthetic data set of hyperspectral Rrs(λ) has been created covering a range of environmental conditions in terms of phytoplankton composition of an hypothetical open ocean scenario. Despite natural phytoplankton assemblages usually consist of several species, the study presented in this chapter has focused on dierent simpli- ed open ocean scenarios. As a rst approach for testing the proposed methodology, dierent scenarios have been created under fully-controlled conditions and each dominated only by a single phytoplankton group. The RT optical simulations have been carried out within the spectral region 350 to 750 nm with a spectral resolution of 1 nm. In order to emulate the real conditions for an open ocean scenario, the ocean has been assumed to be innitely deep and the calculations included the inelastic Raman scattering and uorescence by chlorophyll and CDOM within the ocean. The sea surface boundary conditions have been congured using the default settings in the Hydrolight/Ecolight RT model, being the same for all modeled scenarios (i.e., wind speed equal to 5 m/s, solar zenith angle equal to 30 º and null cloudiness). The inherent optical properties (IOPs) of the water column 44
2.2 Results and discussion required as input to the simulations have been derived from Pope and Fry (1997) for the pure water component and from Kim and Philpot (2006) for the phytoplankton component, which included specic absorption and scattering spectral coecients corresponding to six dierent phytoplankton groups. Figure 2.5 shows the significant variability in spectral features of the specic absoprtion coecients, a∗ ph(λ) , of these phytoplankton groups: cryptophyceae, cyanophyceae, diatoms, dinophyceae, prasinophyceae and prymnesiophyceae. The contribution of the inorganic component of particulate matter (i.e., minerals) to total absorption and scattering coecients has been considered negligible for open ocean scenarios (see Fig. 1.6 for Case 1 waters). For the colored dissolved organic matter, CDOM, absorption has been congured to covary with particle absorption coecient according to the commonly used model from Morel and Maritorena (2001). Figure 2.5: Specic absorption coecients corresponding to six phytoplankton groups, each indicated in a dierent color. A set of hyperspectral Rrs(λ) spectra have been simulated, each of them characterized by the presence of only one phytoplankton group along the water column, which has been assumed to be optically homogeneous (i.e., constant concentration depth prole for all components). Therefore, a total of thirty Rrs(λ) spectra have been generated from combining the six dierent dominating phytoplankton groups at ve dierent concentration values (i.e., 0.01, 0.03, 0.05, 0.07 and 0.09 mg/m3 ). 45
2 Model-based assessment to use hyperspectal data in monitoring phyto. communities It is noted that these levels correspond to typical concentrations encountered under non-bloom sea conditions. Figure 2.6 illustrates how the modeled Rrs(λ) spectra display a great variability in both magnitude and spectral shape, in correspondence with the variable phytoplankton composition at dierent concentration levels. Figure 2.6: Results of RT simulations showing the dierences in hyperspectral Rrs(λ) spectra (1 nm resolution) corresponding to dierent dominating phytoplankton groups at dierent concentration levels (i.e., 0.01, 0.03, 0.05, 0.07 and 0.09 mg/m3 ). 2.2.3 Automatic classication of phytoplankton communities HCA clustering techniques have been used independently to identify and group similar phytoplankton communities from hyperspectral Rrs(λ) spectra. The cluster tree obtained from the analysis applied to the ordinary hyperspectral Rrs(λ) spectra is displayed in Fig. 2.7. Each scenario (i.e., each simulated Rrs(λ) ) is identied with a specic label, consisting of the name of the dominating phytoplankton group and the concentration value. For instance, if diatoms are the dominating phytoplankton group, with a concentration of 0.05 mg/m3 along the homogeneous water column, the label identifying that case will be Diat_0.05. In addition, a specic color is used to identify all cases corresponding to the same phytoplanktonic group. It is worth noting that, when ordinary Rrs(λ) spectra are used as input data for the cluster analysis, the dendrogram provides dierent clusters but most of them composed by cases corresponding to oceanic scenarios with a dierent phytoplanktonic dominance. The only phytoplankton group which is clustered satisfactorily and separately from the rest is the one corresponding to the group of prasinophyceae (identied 46
2.2 Results and discussion in color cyan), except for the lowest concentration (i.e., 0.01 mg/m3 ). Whereas, the remaining cases dominated by the other algal groups (i.e., diatoms, cryptophyceae, cyanophyceae, dinophyceae and prymnesiophyceae) at dierent concentration levels are grouped all mixed. In addition, no improvements are obtained from testing the cluster analysis of ordinary Rrs(λ) using an Euclidean distance measure instead of an angular distance (not shown). Figure 2.7: Cluster analysis based on modeled ordinary hyperspectral Rrs(λ) spectra shown in Fig. 2.6. These results point out the real challenge that implies the identication of phytoplankton community composition from the analysis of the apparent optical property: Rrs(λ) . The reason is that the spectral Rrs(λ) of the ocean is highly inuenced by the optical response of the phytoplankton pigment composition but also by many other optically signicant non-phytoplankton constituents (e.g., CDOM). The unsatisfactory results obtained from the cluster analysis based on the ordinary Rrs(λ) spectra suggests the utilization of derivative spectra of Rrs(λ) in order to improve our ability to classify dierent open ocean scenarios dominated by dierent phytoplankton groups. Figure 2.8 depicts the second derivative spectra of Rrs(λ) , obtained with the parameters WS and BS arbitrarily selected to be equal to 5 nm and 11 nm, respectively. These values provided a good trade-o between denoising and enhancement of spectral details of interest. Although derivative calculations have been made using data from the whole spectral range (i.e., 350 - 750 nm), the results from the derivative analysis of Rrs(λ) spectra are reported between 361 nm ( ≡ 47
2 Model-based assessment to use hyperspectal data in monitoring phyto. communities 350+11 ) and 739 nm ( ≡750−11 ) according to the BS considered. As it is observed, derivative analysis of Rrs(λ) data enhances shape singularities in hyperspectral data, which are signicant since they are related to absorption features of phytoplankton pigments present in the samples, specially within the spectral region below 540 nm (see Fig. 1.3). Figure 2.8: Second derivative spectra of the modeled Rrs(λ) shown in Fig. 2.6. Derivative spectra for the range of 470-500 nm is enlarged with regard to the sensitivity analysis discussed at the end of this section (see Fig. 2.11). This is an optically-signicant region in terms of detecting the phytoplankton community composition. The cluster analysis based on the second derivative of Rrs(λ) spectra and an angular distance measure provides better results (see Fig. 2.9) in comparison to the results obtained based on the ordinary Rrs(λ) spectra. When second derivative spectra over the whole spectral range are considered for the cluster analysis, the majority of scenarios dominated by the same phytoplankton groups at dierent concentration levels are identied and connected by closer dendrites in the dendrogram. In this sense, derivatives of Rrs(λ) spectra corresponding the same phytoplankton group (i.e., the same color) are located closer in the dendrogram and create clusters that are clearly separated one from each other, specially for the dinophyceae and prasinophyceae. The degree of separation between clusters which group scenarios of the same phytoplanktonic dominance is higher. However, there exist some cases which are not well grouped, corresponding to those with the lowest concentration levels (i.e. 0.01 mg/m3 and in some case 0.03 mg/m3 ). In fact, two multi-color clusters can be identied grouping all these scenarios due to Rrs(λ) spectra corresponding to low concentra48
2.2 Results and discussion tions have a similar signature regardless the dominant phytoplankton group. This is something expected given that spectral features of absorption of phytoplankton are less evident and have a smaller eect in the Rrs(λ) spectrum as the concentration decreases. Figure 2.9: Cluster analysis based on the second derivative of hyperspectral Rrs(λ) spectra for the total spectral range from 361 to 739 nm (see Fig. 2.8). Despite the previous analysis does not allow to resolve dierent phytoplankton communities at low concentrations, the results provide the rst detailed demonstration of the advantages and limitations of integrating hyperspectral Rrs(λ) data, derivative spectroscopy and HCA cluster techniques for an automatic identication of phytoplankton communities in open ocean environments. The results conrm the potential of accounting for complete spectral behavior of derivative of hyperspectral Rrs(λ) spectra in comparison with the use of multispectral observations or band ratios of discrete spectral values of derivatives (Lubac et al. , 2008; Louchard et al. , 2002). The HCA cluster analysis based on the hyperspectral data as input, with variable optical conditions, has been able to automatically bring together scenarios corresponding to the same phytoplanktonic dominance, going beyond the studies focused on the detection of the presence of a single phytoplankton group (Craig et al. , 2006). It is important to note that the performance of cluster analysis applied to derivative spectra of Rrs(λ) is clearly dependent on the examination of the spectral data and the optimal selection of involved parameters (i.e. smoothing lter size and band separation) for each particular data set and for each specic purpose of the analysis. For instance, if smoothing and derivative parameters are selected to be too coarse, 49
2 Model-based assessment to use hyperspectal data in monitoring phyto. communities spectral features of interest might be lost, probably causing a worse performance. Figure 2.10 illustrates an example in which the cluster analysis is applied to second derivative spectra of the modeled hyperspectral Rrs(λ) obtained using as parameters a BS = 40 nm and WS = 40 nm. In this case, scenarios dominated by the groups prymnesiophyceae and diatoms are clustered together and scenarios corresponding to the group of cyanophyceae (in color blue) appear also mixed with other phytoplankton groups. Therefore, the resulting analysis based on the derivative settings for spectra of Rrs(λ) provides a worse partitioning, which prevents a satisfactory identication of dierent phytoplankton communities. It is important to note that a wide range of values for the BS and WS parameters were also tested to conrm that variability of our performance but results are not shown. Figure 2.10: Cluster analysis based on the second derivative of Rrs(λ) spectra using analyzing parameters that are too coarse (i.e. BS = WS = 40 nm). In order to keep on investigating the importance of selecting the optimal parameters involved in the proposed analysis, it is also essential to take into account the spectral region considered for the analysis. Because the absorption characteristics of main accessory pigments occur at specic wavelength ranges, it is therefore useful to examine whether the cluster analysis of derivative spectra of Rrs(λ) yields better results if a specic spectral range is considered. Following the examination of derivative spectra of Rrs(λ) (see Fig. 2.8), a region in which the dierent scenarios present distinct remote-sensing optical patterns is detected. This spectral region ranges from 470 to 500 nm and coincides with a spectral region in which several accessory pigments have relevant spectral features (see Fig. 1.3). The dendrogram 50
2.3 Summary and conclusions obtained from cluster analysis as applied to derivative Rrs(λ) spectra over this region is displayed in Fig. 2.11. The analysis based on this limited spectral information provide a very good performance, improving signicantly the results obtained based on the complete spectral range (see Fig. 2.9). In this sense, scenarios corresponding to all groups of phytoplanktonic dominance are clearly identied and grouped in separate clusters, showing a much higher degree of separation between them. The methodology proposed, however, still presents some limitations when the analyzed underwater scenarios are characterized by a very low concentration level of phytoplankton (i.e., 0.01 mg/m3 ). Regardless the dominant phytoplankton group, the optical patterns corresponding to these scenarios are too similar and therefore, are grouped in a unique cluster. This cluster also includes the scenarios dominated by cyanophyceae at dierent levels of concentration (in blue color), which in fact seems logical since it is the phytoplankton group that presents a weaker absorption pattern (see Fig. 2.5). Figure 2.11: Cluster analysis based on second derivative of hyperspectral Rrs(λ) spectra limited to the spectral range from 470 nm to 500 nm (see enlarged region in Fig. 2.8). 2.3 Summary and conclusions The eectiveness of hyperspectral optical information, in particular remote sensed observations, for assessing phytoplankton diversity in the ocean has been explored in this chapter. As a primary approach, the feasibility of hyperspectral oceanographic 51
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean sensing reectance. A brief description of the measurements is given in the following subsections. More methodological details which are beyond the scope of this study for the radiometric and backscattering measurements are in Stramski et al. (2008). Figure 3.2: Map depicting the location of full stations sampled along the north-tosouth ANT-XXIII/1 cruise track in the eastern Atlantic during October and November, 2005. Each full station consisted of in situ optical measurements accompanied by discrete water sample analyses. Stations chosen for use in this chapter are identied by lled circles and labeled. Figure 3.3: (left panel) Multisensor Datalogger System (MDS) used to measure in situ several IOPs of seawater. (right panel) CTD/rosette system used to collect water samples, which led to laboratory measurements of other IOPs. See Fig. 3.4 for details. Source : D. Stramski and R. Reynolds. 58
3.1 Reconstruction of hyperspectral reectances 3.1.1.1 Pigment analysis Concentrations of chlorophylla and accessory pigments in phytoplankton were measured on surface water samples from each station using HPLC techniques. As a standard security protocol, two sets of replicate samples were collected in the ship and analyzed at two laboratories, the Center for HydroOptics and Remote Sensing (CHORS) laboratory at San Diego State University (California, USA) and the GKSS Research Centre in Geesthacht (Germany). The CHORS analysis was based on a method described in Heukelem and Thomas (2001), and the GKSS samples were analyzed following the method of Zapata et al. (2000). However, the CHORS analysis included identication and quantication of more pigments (27) than the GKKS method (23) including alternative forms of chlorophylla , and therefore was chosen in this study as the primary pigment data set for identifying phytoplankton assemblages using cluster analysis (see Section 3.3.1). Throughout the rest of this chapter, it is used the CHORS values of the total chlorophylla (TChl a) as a measure of chlorophylla , which is dened as the summed contributions of concentrations of monovinyl chlorophylla (MVChl a ), divinyl chlorophylla (DVChl a ), chlorophyllidea (Chlide), and the allomeric and epimeric forms of chlorophylla . Following completion of these analyses, data quality problems in the CHORS analytical procedures were identied which suggest the potential of errors in the determination of some pigment concentrations in the samples (Hooker and Heukelem, 2009). Although it was applied an independently-derived correction factor to a few individual pigments, the nature of the methodological issues precludes development of a general correction scheme applicable for all pigments. In the Appendix B, these issues are discussed in more detail and the results of analyses are summarized, which indicate that use of the CHORS data provides nearly identical results with respect to station clustering and classication as the GKSS HPLC data. Because of this independent corroboration, it was continued to use the CHORS pigment data as the reference data set in this study. Because of a large range of pigment compositions across dierent phytoplankton classes, determination of phytoplankton composition from HPLC pigment data is not straightforward (e.g., Jerey et al. 1999). Whereas certain diagnostic pigments can serve as unambiguous markers for some phytoplankton classes (e.g., peridinin in dinoagellates, alloxanthin in cryptophytes), many important pigments are shared by more than one algal taxa (e.g., fucoxanthin in diatoms, haptophytes, chrysophytes, and raphidophytes). Nevertheless, because many of the classes have distinctive suites of marker pigments, HPLC data can be useful for indicating their presence and abundance in a mixed phytoplankton population. Specically, a useful indication of contributing phytoplankton classes can be obtained from the ratios of the concentrations of specic pigments to chlorophylla or the ratios of specic diagnostic pigments to the sum of these diagnostic pigments, because these ratios can dier between phytoplankton groups (Mackey et al. 1996; Vidussi et al. 2001; Wright et al. 1996). 59
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean For each of the nine stations selected in this study, it was calculated a set of pigment ratios for subsequent use in the cluster analysis (see Section 3.3.1). This set of pigment ratios consisted of ratios of the concentration of each individual pigment to the TChl a concentration, as obtained from HPLC measurements at the CHORS laboratory. The following 24 pigments were included in these calculations: monovinyl chlorophylla , divinyl chlorophylla , chlorophyllidea , chlorophylla allomer, chlorophylla epimer, monovinyl chlorophyllb , divinyl chlorophyllb , chlorophyllc 2, chlorophyllc 3, α - carotene, β - carotene, alloxanthin, diadinoxanthin, diatoxanthin, fucoxanthin, 19' - hexanoyloxyfucoxanthin, 19' - butanoyloxyfucoxanthin, neoxanthin, prasinoxanthin, violaxanthin, zeaxanthin, peridinin, pheophorbidea , and lutein. The CHORS pigment data set also included a few additional pigments (chlorophyllc 1, gyroxanthindiester, and pheophytina ) which were below the detection level for the nine stations. They were not included in the analysis as they would have no eect on the results. 3.1.1.2 Inherent optical properties The inherent optical properties (IOPs) of seawater were measured in situ along with laboratory analyses of water samples through the use of dierent instruments (Fig. 3.3). They have a two-fold application in this study. First, the absorption, scattering, and backscattering coecients are used to dene IOP inputs to radiative transfer simulations that generate hyperspectral data of remote-sensing reectance (see Fig. 3.4 and Section 3.1.2). Secondly, the total and phytoplankton absorption coecients, a(λ) and aph(λ) , are utilized directly in the cluster analysis (Section 3.3.2). As part of the work performed during the eld campaign in 2005, the spectral absorption coecients of particles, ap(λ) , and colored dissolved organic matter (CDOM), acdom(λ) in m −1 , were determined at 1-nm intervals from high spectral resolution measurements on freshly-collected discrete water samples with a point-source integrating cavity absorption meter (PSICAM) over the range 350750 nm (Röttgers et al. , 2005; Röttgers and Doerer, 2007). As the PSICAM did not provide data below 350 nm, the ap(λ) values within the 300 to 350 nm spectral range were obtained from lter pad measurements on discrete water samples collected on glass ber lters (GF/F) and frozen in liquid nitrogen until analysis with a dual-beam spectrophotometer (Lambda 18, Perkin Elmer). The lter pad measurements were made with the transmittancereectance (T-R) technique of Tassan and Ferrari (1995, 2002) using a correction for the path length amplication factor from Stramska et al. (2006). In this study, it has been chosen to use the PSICAM data of ap(λ) over the majority of the spectrum because the PSICAM technique involves a direct measurement of absorption on particle suspension with minimal scattering artifacts, which is expected to be generally superior to the lter pad measurements. The data of acdom(λ) below 350 nm were obtained from an exponential t to the PSICAM-measured acdom(λ) . A null point correction based on wavelengths in the far red or near-infrared was applied to all ap(λ) and acdom(λ) spectra. The total spectral absorption coecient, a(λ) , was determined as the sum of ap(λ) , acdom(λ) , and the pure water component, aw(λ) . 60
3.1 Reconstruction of hyperspectral reectances The latter was obtained from Pope and Fry (1997) for the spectral range 380727 nm and from Fry et al. (2006) for the range 300379 nm. It is important to note that the primary interest is in the spectral information contained at wavelengths longer than 350 nm extending throughout the visible part of the spectrum up to 725 nm where most phytoplankton pigments exhibit signicant absorption features. However, data at wavelengths shorter than 350 nm are useful for this analysis, especially in the context of derivative spectra whose discrete values at specic wavelengths were calculated in this study using data covering a bandwidth on the order of 1030 nm. The spectra of the phytoplankton absorption coecient, aph(λ) , were determined as a dierence between the absorption coecient of particles, ap(λ) , and the non-phytoplankton component of particulate absorption, ad(λ) , which is commonly referred to as detrital absorption. These determinations were based on the T-R lter pad measurements, in which the ad(λ) spectra were measured on GF/F sample lters following treatment with sodium hypochlorite [NaOCl] (Ferrari and Tassan, 1999). In this treatment, the particles on the sample lter were exposed to a small amount of a 2% NaOCl solution for several minutes to bleach phytoplankton pigments. Figure 3.4: Diagram illustrating the undertaken process to determine each IOP and generate each modeled hyperspectral Rrs(λ) spectra. 61
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean The spectral beam attenuation coecient of particles and CDOM, cp,cdom(λ) in m −1 , was determined at each station from in situ measurements with two singlewavelength C-Star transmissometers (488 and 660 nm; WET Labs, Inc.). Note that the C-Star derived values represent the total beam attenuation, c(λ) , with pure seawater contribution, cw(λ) , subtracted, i.e. cp,cdom(λ) = c(λ)−cw(λ) = ap(λ) + acdom(λ) + bp(λ) , where bp(λ) is the spectral scattering coecient of particles and assuming that dissolved matter makes negligible contribution to scattering. In this study (see Fig. 3.4), the values of bp(λ) at 488 and 660 nm were thus calculated from C-Star attenuation and PSICAM absorption measurements as bp(λ) = cp,cdom(λ)−ap(λ)−acdom(λ) . A power function t was then applied to these values to produce the spectral data of bp(λ) over the range 300750 nm with a 1 nm resolution. The pure seawater scattering coecient, bw(λ) , was calculated using the Buiteveld et al. (1994) equations with measured water temperature and salinity (see Twardowski et al. , 2007 and Stramski et al. , 2008). The total scattering, b(λ) , was obtained as a sum bw(λ) + bp(λ) . The spectral backscattering coecient, bb(λ) in m−1 , was determined in this study (see Fig. 3.4) by combining in situ measurements with three instruments, a Hydroscat6 and two aβ eta sensors (HOBI Labs, Inc.), to yield a total of eight spectral bands: 420, 442, 470, 510, 550, 589, 620, and 671 nm. Because bb(λ) was generally observed to be a smooth monotonic function of wavelength within the spectral range of the measurements (Stramski et al. , 2008), the experimental data were tted to a power function to obtain hyperspectral resolution over the 300750 spectral range. The spectral backscattering coecient of particles, bbp(λ) in m−1 , was determined as a dierence between the total and pure seawater backscattering coecients, bb(λ)−bbw(λ) , in which the pure seawater component, bbw(λ) , was previously calculated as indicated: 0.5 bw(λ) . From the values of bp(λ) and bbp(λ) , it was calculated the particle backscatter fraction Bp(λ) = bbp(λ)/bp(λ) necessary for this study (see Fig. 3.4). These data were then tted to a power function, Bp(λ) = Bp(λ0)·(λ0/λ)m , where λ0 is the reference wavelength 550 nm. The backscattering fraction Bp(λ0) at the reference wavelength and the exponent m represent the best t parameters of the linear regression analysis performed for the log-transformed data of Bp(λ) vs. λ for each station. The parameters of the power function t of Bp(λ) were used as input to radiative transfer simulations (see Section 3.1.2). 3.1.1.3 Multispectral remote-sensing reectance Values of multispectral remote-sensing reectance, Rrs(λ) in sr−1 , were estimated at each station at 13 wavelengths from in situ measurements of underwater vertical proles of spectral nadir upwelling radiance, Lu(λ, z) in Wm−2sr−1nm−1 , and spectral downwelling plane irradiance, Ed(λ, z) in Wm−2nm−1 , where z is depth. These measurements were made with a freefall spectroradiometer (see Figure 3.5), the SeaWiFS Proling Multichannel Radiometer (SPMR, Satlantic, Inc.). The wavelengths 62
3.1 Reconstruction of hyperspectral reectances for these measurements are 339, 380, 412, 443, 470, 490, 510, 532, 554, 589, 619, 666, and 683 nm. The radiometric measurements and data processing were consistent with methods recommended in NASA protocols (Mueller et al. , 2003b). The deployment of the spectroradiometer occurred simultaneously with water sample collection from the CTD/rosette, and immediately before or after deployment of the instrument package containing attenuation and backscattering sensors. Figure 3.5: SeaWiFS Proling Multichannel Radiometer (SPMR) utilized to measure upwelling radiances, Lu(λ, z) , and downwelling plane irradiances, Ed(λ, z) . Source : D. Stramski and R. Reynolds. 3.1.2 Modeled hyperspectral remote-sensing reectance Because hyperspectral radiometric measurements were not conducted during the ANT-XXIII/1 cruise and the primary interest is in the analysis of hyperspectral optical data, numerical simulations of radiative transfer (RT) were performed to estimate the hyperspectral remote-sensing reectance, Rrs(λ) , for each of the nine selected stations (Fig. 3.4). The radiative transfer model Hydrolight/Ecolight version 5.0 was used (Mobley and Sundman, 2008). An important prerequisite for undertaking these RT simulations was the availability of a comprehensive suite of IOPs for each station for use as input to the simulations (see Section 3.1.1.2), and also the availability of the multispectral Rrs(λ) derived from in situ measurements for use in validating the simulated hyperspectral Rrs(λ) . The RT calculations were carried out within the spectral region from 300 nm to 725 nm with high spectral resolution (1 nm). Similarly to the absorption, the main interest is in the reectance data at wavelengths longer than about 350 nm. However, in addition to the requirements associated with derivative calculations, the RT simulations below 350 nm are needed to account for Raman scattering contributions observed at λ > 350 nm. The ocean was assumed to be innitely deep and optically homogeneous, and the simulations included the Raman scattering and uorescence by colored dissolved organic matter within the ocean. The sea surface boundary conditions were estimated from observations of wind speed and sky conditions (cloudiness) at each station site, and the solar zenith angle was calculated for the corresponding date, time, and geographic coordinates. The inherent optical properties of the 63
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean water column required as input to the simulations were derived from the IOP measurements in the surface waters and additional relevant determinations of a(λ) and b(λ) as described in Section 3.1.1.2. The selection of the particulate scattering phase function, which is also part of IOP inputs to the RT simulations, was based on the particle backscatter fraction Bp(λ) . The FournierForand phase functions (Fournier and Forand, 1994; Fournier and Jonasz, 1999) were used, which are parameterized in terms of Bp(λ) and are built into the Hydrolight/Ecolight model. Fig. 3.6 compares the modeled hyperspectral Rrs(λ) with the measured multispectral Rrs(λ) for two representative stations. The model results compare reasonably well with measurements, which lends condence to the use of hyperspectral Rrs(λ) in the cluster analysis. This level of consistency between model and measurements suggests that the suite of parameters used as input to the RT simulations realistically represents the actual eld conditions. The ability to dene realistic inputs derives, in turn, from a comprehensive suite of IOP measurements carried out in the eld. Figure 3.6: Hyperspectral (1 nm) determinations of the remote-sensing reectance Rrs(λ) obtained from radiative transfer simulations (solid line) compared with in situ multispectral measurements at 13 discrete bands (solid circles). Each panel illustrates a dierent station location. 3.2 Cluster analysis and similarity indices between dendrograms A hierarchical cluster analysis (HCA) was used to classify the 9 selected stations into distinct groups on the basis of several types of input data vectors (or objects), which included the HPLC pigments and optical data derived from spectral absorption coecients and remote-sensing reectance. For a given type of data, the input to the cluster analysis consisted of 9 numerical data vectors, each representing one of the stations. For the input data representing the ratio of individual pigment concentrations to TChl a , an object for a given station is a data vector {p1, p2, p3, . . . , p24} where the consecutive elements pi represent the ratio of each of the 24 individual pigment concentrations to TChl a concentration. 64
3.2 Cluster analysis and similarity indices between dendrograms Several types of optical data vectors were used as input to the HCA analysis, including objects consisting of hyperspectral data of the remote-sensing reectance, Rrs(λ) , the phytoplankton absorption coecient, aph(λ) , the sum of pure water and phytoplankton absorption coecients, aw(λ)+ aph(λ) , and the total absorption coecient, a(λ) . The input characterizing the hyperspectral remote-sensing reectance for any given station was used in the form of the following data vector {Rrs(λ1)/Rrs(555), Rrs(λ2)/Rrs(555), Rrs(λ3)/Rrs(555), ..., Rrs(λn)/Rrs(555)} , where the consecutive elements represent the values of Rrs(λ) at successive light wavelengths normalized to Rrs at 555 nm over the spectral range from λ1 to λn . Similar input vectors were created for the dierent components of spectral absorption. Because the analysis is focused on the spectral shapes, all the optical spectra used in the cluster analysis were normalized by the value of the optical variable at 555 nm at which variations in Rrs within the open ocean are generally small. The spectra involving the absorption coecients were additionally normalized by TChl a concentration to minimize variability in absorption associated with changes in phytoplankton biomass. The rationale for selecting the data of a(λ) , aw(λ)+aph(λ) , and aph(λ) to create input data vectors for the cluster analysis stems from the fact that the variation in the spectral shape of a(λ) is typically a major determinant of the variation in the spectral shape of Rrs(λ) . In turn, the variations in the spectral shape of aw(λ) + aph(λ) or aph(λ) can be viewed as an important or dominant source of variation in the spectral shape of a(λ) in open ocean situations. Vectors from the second derivative spectra of the hyperspectral reectance and absorption objects were also created for input into the cluster analysis. The estimation of the second derivative spectra from these data was made with a nite divided dierence algorithm (see Section 2.1.1), which computes the changes in curvature of a given spectrum over a sampling interval ( 4λ ) or band separation ( BS ) dened as 4λ=λj−λi , where j > i . As described in Section 2.1.1, because the identication of spectral details in the derivative spectra depends on the selection of the band separation, it was tested the sensitivity of cluster results to the choice of BS . It was also tested the sensitivity of cluster results to dierent values of the smoothing lter window ( WS ) used prior to computation of derivative spectra. The sensitivity analysis over a range of BS and WS values allowed to achieve the best compromise between the ability to resolve ne spectral details and the reduction of noise eects in the second derivative spectra. As discussed in Section 3.3.4, the optimal values of BS and WS chosen in this study for the derivative analysis of absorption data are 9 nm. For the derivative analysis of reectance data, these values are 27 nm. Therefore, although the derivative calculations were made using data from the spectral range 300725 nm, the results from the derivative analysis for absorption will be reported between the wavelengths of λmin = 309 nm (≡300 + 9) and λmax = 716 nm (≡725 −9) . For the reectance derivative, the results will be reported between λmin = 327 nm (≡300 + 27) and λmax = 698 nm (≡725 −27) . With regard to the analysis of remote-sensing reectance, the cluster analysis was also applied to the multispectral reectance data obtained from in situ SPMR measurements at several discrete wavelengths (and described in Section 3.1.1.3). The 65
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean 3-element objects {Rrs(443)/Rrs(554), Rrs(490)/Rrs(554), Rrs(510)/Rrs(554)} were examined, which consist of band ratios that are similar to those used in current research based on satellite ocean color observations such as the Sea-viewing Wide Field-of-View Sensor (SeaWiFS; O'Reilly et al. , 2000). The vectors consisting of 13 band ratios of remote-sensing reectance with Rrs(554) in the denominator, as determined from SPMR measurements at 13 wavebands, were also examined. The HCA method, schematically presented in Fig. 3.1, was applied using the above described pigment and optical data vectors as input objects. This method utilizes the unsupervised classication algorithm dened in the previous chapter (see Section 2.1.2). Each cluster tree is obtained using a linkage algorithm (i.e., shortest distance) based on initial calculations of the pairwise distance (i.e., cosine angular distance) between all objects included in the input data set. As a novelty in this study, in order to automatically evaluate the usefulness of optical data for discriminating phytoplankton pigment assemblages, the dendrograms obtained for the dierent spectral optical data were compared with a reference dendrogram obtained using the pigment composition data (see nal step in Fig. 3.1). The pigment information has been commonly used by the scientic community as a proxy for the phytoplankton composition. For this evaluation, the use of two objective criteria of cluster similarity was proposed, which allowed to go beyond the subjective validation of performance made by visual inspection of the resulting cluster partitions. An example is shown in Fig. 3.7, illustrating how the similarity between two given dendrograms is computed using the cophenetic index and the Rand index: The cophenetic index (Sokal and Rohlf, 1962), rC , is a measure of how precisely two dendrograms preserve the pairwise distances between data objects. This index is computed from the cophenetic matrix ( C ) associated with each dendrogram. The elements of a cophenetic matrix ( ci,j ) encode the distance between two objects ( i, j ), representing in the dendrogram the height of the link at which those two objects are rst joined. This height is the distance between the two clusters that are merged by this link. The cophenetic index rC represents the correlation between two cophenetic matrices ( C1 and C2 ): rC=PiPjc1i,j −¯c1c2i,j −¯c2 sPiPjc1i,j −¯c12PiPjc2i,j −¯c22 (3.1) where ¯c1 and ¯c2 are the mean values of the elements of the matrices C1 and C2 . The Rand index (Rand, 1971), rR , provides a measure of the similarity between two hierarchical dendrograms in terms of the proportion of pairs of objects whose relationship is the same in both dendrograms. The rR value of 1 means that all pairs of objects are clustered in the same way in both dendrograms. Note that this index has to be computed using all dendrograms cut horizontally at a level (i.e., at a specic linkage distance) which yields the optimal number of clusters ( k ). Otherwise, rR would always provide a proportion of 100% because 66
3.2 Cluster analysis and similarity indices between dendrograms a complete dendrogram always decomposes the input data all the way through the lowest level (i.e., until the branches consist only of single objects). Detecting natural groupings in the dendrogram and selecting the optimal number of clusters is performed by analyzing a diagram of the increasing linkage distances along the dendrogram. Based on the points at which the linkage distances between the objects change abruptly (which is associated with a steep increase of the within cluster variance), the optimal number of clusters k is determined and all objects located below the point where the hierarchical tree is cut o are assigned to a single cluster (Salvador and Chan, 2004). Figure 3.7: Example of similarity between two hierarchical cluster trees, computed using the cophenetic and Rand indices. In the latter, an optimal number of clusters ( k ) needs to be pre-dened in the cluster tree. 67
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean nm to λmax = 716 nm. This result indicates that a high degree of similarity between the second derivative spectra of a∗ n,ph(λ) and pigment composition can be obtained for many dierent combinations of spectral ranges of absorption data (i.e., dierent combinations of λmin and λmax ). Figure 3.11: Similarity indices between absorption-based and pigment-based cluster trees obtained for the 9 stations using dierent combinations of spectral range for (a, b) a∗ n(λ) , (c, d) a∗ n,w+ph(λ) and (e, f) a∗ n,ph(λ) . The y-axis indicates the lower limit of the spectral range ( λmin ) and the x-axis the upper limit of the spectral range ( λmax ) used in the cluster analysis. Left and right panels depict the cophenetic and Rand indices, respectively. 74
3.3 Results and discussion Figure 3.12: Similar to Fig. 3.11, but based on cluster trees obtained using dierent spectral range combinations for the second derivative spectra of (a, b) a∗ n(λ) , (c, d) a∗ n,w+ph(λ) and (e, f) a∗ n,ph(λ) . Optimal values for band separation and window size determined from prior analyses ( BS = WS =9 nm, see Section 3.3.4) were used in the calculation of derivative spectra. Because the cophenetic index does not require a priori selection of the optimal number of clusters, this index facilitates comparison of the results for ordinary and derivative spectra with pigment-based cluster trees shown in Figs. 3.11e and 3.12e. The improved similarity obtained by utilization of the second derivative spectra of 75
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean a∗ n,ph(λ) was evidenced as a marked increase in the values of the cophenetic index when computed for all possible combinations of spectral ranges. These increased values were associated with a broader overall spectral region, or equivalently a larger number of spectral ranges, for which the index exceeded 0.9 (see Fig. 3.13). Figure 3.13: Histograms of cophenetic indices obtained for all combinations of spectral ranges shown in Fig. 3.11e and 3.12e based on (a) the absorption of phytoplankton, a∗ n,ph(λ) , and (b) its second derivative spectra. 3.3.3 Classication of stations based on remote-sensing reectance The relationship between the spectral remote-sensing reectance, Rrs(λ) , of the ocean and phytoplankton pigment composition is less direct and far more complicated than that for the spectral phytoplankton absorption coecient, mainly due to the presence of many optically signicant non-phytoplankton constituents in seawater. The investigation of Rrs(λ) is of particular interest, however, because information contained in this measurement provides a potential means for remote-sensing applications. Fig. 3.14 shows the Hydrolight-simulated Rrs(λ) spectra normalized at 555 nm for the nine stations identied as classes from A (station 1) through F (station 59). In general, there are signicant dierences in the UV and blue spectral regions between these normalized spectra, with the largest contrast between the classes A (station 1) and C1 (station 12). The dendrograms obtained from cluster analysis as applied to four dierent sets of input data vectors containing information about remote-sensing reectance (as described in Section 3.2) are displayed in Fig. 3.15. The limited spectral information, i.e., the three reectance band ratios used commonly in satellite ocean color applications (Fig. 3.15a) and the 13 band ratios corresponding to multispectral measurements with the SPMR instrument (Fig. 3.15b), provide a very dissimilar 76
3.3 Results and discussion classication of stations compared with the pigment-based cluster analysis (see Fig. 3.8a). All stations, with only the exception of class A, show very little separation in the cluster-tree based on multispectral reectance data. The high spectral resolution (1 nm) normalized reectance spectra over the entire spectral range 300 725 nm also produce a dendrogram (Fig. 3.15c) that is very dierent from the pigment-based cluster tree. Although the stations belonging to class C are closer to one another compared with the multispectral-based cluster tree, they are grouped together with two other stations (classes E and D). In addition, station B forms a separate singleobject cluster in Fig. 3.15c, whereas it is grouped in a single multi-object cluster together with the stations from class C in the pigment analysis. The only case when the cluster analysis of reectance data provides a high degree of similarity with pigment analysis (i.e., Rand index of 0.78) is for the second derivative of hyperspectral normalized reectance over the entire spectral range 327698 nm (Fig. 3.15d). It is important to note that the derivative reectance spectra were calculated with the parameters WS and BS of 27 nm (as supported by the sensitivity analysis discussed in Section 3.3.4). The stations A, E, and F in Fig. 3.15d form single-object clusters at a signicant distance from the remaining stations. Similarly to pigment analysis, the stations C1, C2, C3, C4, and B are grouped relatively close to one another. However, station D also belongs to that group, which is not the case in the pigment-based cluster tree. This may be attributable to the fact that Zea and DVChl a , which are the two most dominant diagnostic pigments at stations C1, C2, C3, C4, and B, also play a signicant role at station D where they are ranked as the second and third most important diagnostic pigments (see Table 3.1). Figure 3.14: Hydrolight-simulated Rrs(λ) spectra, normalized at 555 nm, computed for the nine stations using measured IOPs as input. 77
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean Figure 3.15: Dendrograms resulting from cluster analysis of the nine stations calculated using four dierent sets of input data vectors: (a) three reectance band ratios of Rrs(λ) based on 4 SeaWiFS wavebands obtained from measurements with SPMR instrument, (b) 13 band ratios corresponding to multispectral measurements of Rrs(λ) with SPMR instrument, (c) hyperspectral (1 nm) ordinary (non-dierentiated) normalized Rrs(λ) spectra computed from the Hydrolight simulations, and (d) second derivative of hyperspectral normalized Rrs(λ) spectra obtained using optimal values for band separation and smoothing lter window (i.e., BS = WS = 27 nm, see Section 3.3.4). The progression of linkage distances corresponding to the four dendrograms from Fig. 3.15 clearly illustrates the advantage of the second derivative spectra over the multispectral data or non-dierentiated spectra of reectance (see Fig. 3.16), and indicates that this approach enables better identication of the dierences in the magnitude and shape between the high resolution spectra. In contrast, in the analysis of multispectral data and ordinary spectra stations are linked at a very small distance, suggesting that these types of reectance data will be of little value for obtaining information about pigment assemblages from cluster analysis. The improvement in the similarity between the pigment-based and reectancebased classication achieved with the use of second derivative spectra as opposed to multispectral data or ordinary spectra of reectance is presented in Table 3.2. The results for the derivative analysis of hyperspectral normalized Rrs(λ) over the entire 78
3.3 Results and discussion spectral range 300725 nm show a signicant increase in both the cophenetic and Rand index when compared with multispectral and ordinary spectral data. However, the best performance is obtained when the derivative analysis is restricted to the spectral range from 435 to 495 nm (as supported by the sensitivity analysis discussed below). In this case, the similarity indices are highest. Figure 3.16: Linkage distances as a function of distance along the dendrogram for each cluster tree depicted in Fig.3.15. Table 3.2: A comparison of similarity indices between pigment-based clusters and reectance-based clusters for the dierent sources of reectance data that are depicted in Fig. 3.15. For the case of the second derivative of hyperspectral reectance, the result of computations for two dierent spectral regions is given. 79
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean Fig. 3.17 shows distributions of the cophenetic and Rand index which identify the optimal spectral ranges for the cluster analysis of the second derivative of remotesensing reectance. The cophenetic index is generally close to or slightly lower than 0.5 for most spectral ranges examined, that is for most combinations of λmin and λmax (Fig. 3.17a). In the spectral region from 435 nm to 510 nm, this index is about 0.62 and can be considered as an optimal spectral range for the application of derivative approach with potential for good similarity between the pigment-based and reectance-based cluster trees. This result is also supported by very high value of the Rand index of 0.86 in that spectral range (Fig. 3.17b). The cophenetic and Rand indices attain even higher values of about 0.65 and 1, respectively, within somewhat narrower wavelength range from λmin = 435 nm to λmax = 495 nm, which denes an alternative optimal spectral range. In addition, the cophenetic and Rand indices suggest good performance of the derivative-based analysis over a broader spectral regions including shorter and longer wavelengths of λmin . The Rand index is quite high for λmin varying between 350 nm and 450 nm, for example as high as 1 when the spectral range is from λmin = 365 nm to λmax = 480 nm. The cophenetic index is also quite high for the spectral ranges from λmin = 327 nm to λmax = 400 nm and λmin = 610 nm to λmax = 670 nm. However, because the optical roles of dierent diagnostic pigments in these regions are insignicant or certainly less important than in the blue region, the use of the spectral range 435510 nm or 435495 nm, where both the cophenetic and Rand indices are relatively high, appears to be most reasonable. Figure 3.17: Similarity indices between reectance-based and pigment-based cluster trees obtained for the nine stations using dierent combinations of spectral ranges for the second derivative of the hyperspectral normalized Rrs(λ) . Optimal values of BS = WS = 27 nm, determined from prior analyses, were used for the calculation of derivative spectra. The y-axis indicates the lower limit of the spectral range ( λmin ) and the x-axis the upper limit of the spectral range ( λmax ) utilized in the cluster analysis. Panels (a) and (b) depict the cophenetic and Rand indices, respectively. 80
3.3 Results and discussion 3.3.4 Sensitivity analysis and determination of optimal parameters for derivative analysis Although the previous results illustrate the advantages of derivative analysis, such analysis can be highly sensitive to parameters chosen for the calculation of derivative spectra, specically the size of the lter window ( WS ) used in the spectral smoothing of the ordinary spectra and the band separation ( BS ) used in the calculation of derivatives. In order to display the results from the sensitivity of cluster analysis to the selection of these parameters, a new graphical representation was also used in this study. The distribution of values of the cophenetic and Rand index is shown as a function of the band separation ( BS ) and smoothing lter window ( WS ) considered. In particular, it was computed the distribution of cophenetic index between the pigment-based and absorption-based cluster trees using the second derivative spectra of a∗ n,ph(λ) within the spectral range of 420515 nm as input (Fig. 3.19a). This spectral range is adequate for this sensitivity analysis because it showed very high values of cophenetic and Rand indices in Figs. 3.11e,f and 3.12e,f. The distribution of cophenetic index in Fig. 3.19a is shown with the smoothing parameter WS varying from 1 to 29 consecutive samples (with a step of 2 samples) along the y-axis and the band separation parameter BS varying also from 1 to 29 samples with a step of 2 samples along the x-axis. Note that the number of consecutive samples is equivalent to the wavelength interval in nanometers because the spectral data (samples) have the resolution of 1 nm. The highest values of the cophenetic index are obtained for intermediate values of WS and BS around 9 nm10 nm, and thus are the optimal values for the derivative analysis of the absorption spectra. The corresponding derivative spectra of a∗ n,ph(λ) are presented in Fig. 3.18a. This result is consistent with the general expectation that if the values of WS and BS are too small, the derivative spectra are sensitive to noise and exhibit false spectral features, and on the other hand if the WS and BS are too large, the real signicant spectral features are smoothed over and essentially removed from the analysis. As the best compromise, the cluster analyses of derivative spectra presented in Fig. 3.10 were obtained with WS and BS of 9 nm. A similar analysis was performed to determine the appropriate sizes of WS and BS for utilization of the second derivative of the reectance spectra (Fig. 3.19b). These results are shown for the derivative spectra calculated over one of the optimal spectral ranges, specically 435 495 nm (Fig. 3.18b), which showed high values for both cophenetic and Rand indices. The best similarity with cophenetic index of about 0.65 is obtained when the calculations of second derivative spectra are made with relatively large values of WS and BS . For example, a very good result is obtained if both WS and BS assume a value of 27 consecutive spectral samples (i.e., 27 nm as the resolution of the hyperspectral reectance data is 1 nm). It is recalled that this value was used to compute the results pertinent to the derivative reectance spectra presented in Fig. 3.15d. Similarly, good results are obtained with a smaller WS (~14 nm) and a larger BS (~37 nm), or vice versa. In contrast, if both WS and BS are small (less than about 10 nm) or large (above ~40 nm) the cophenetic 81
3 Hyperspectral data for discriminating phyto. pigment assemblages in the open ocean index is reduced signicantly. This sensitivity analysis, in agreement with similar analysis for absorption spectra, supports the notion of a strong dependence of the derivative-based cluster trees on the selection of derivative parameters WS and BS . Figure 3.18: The second derivative spectra at each station of the (a) chlorophyllspecic normalized phytoplankton absorption coecient, a∗ n,ph(λ) , and (b) normalized hyperspectral remote-sensing reectance, Rrs(λ)/Rrs(555) . Optimal values were used for the derivative calculations, i.e., BS =WS = 9 nm for the absorption data and BS =WS = 27 nm for the reectance data. Figure 3.19: Cophenetic indices obtained from similarity analysis between the pigment-based cluster tree with trees obtained using the second derivative of hyperspectral normalized spectra for (a) phytoplankton absorption, a∗ n,ph(λ) , and (b) remote-sensing reectance, Rrs(λ) , in which different parameter sets for the derivative analysis are considered. The y-axis indicates the size of the lter window used in smoothing of the spectra ( WS ), and the x-axis represents the band separation used in the calculation of the derivative ( BS ). The analysis was conducted using the optimal spectral regions of 420 to 515 nm for a∗ n,ph(λ) , and from 435 to 495 nm for Rrs(λ) . 82
3.4 Summary and conclusions 3.4 Summary and conclusions Through application of an unsupervised hierarchical cluster analysis to pigment and optical data from the eastern Atlantic Ocean in 2005, the potential usefulness of hyperspectral data of absorption coecients and remote-sensing reectances and their second derivative spectra for discriminating dierent phytoplankton pigment assemblages in the open ocean has been demonstrated. The ability to discriminate dierent phytoplankton pigment assemblages from the hyperspectral-based cluster approach has been optimized by selecting the optimal spectral range and the most suitable parameters used in the spectral derivative calculations (i.e., smoothing lter size and derivative band separation). The assessment of similarity between the hyperspectral data and phytoplankton pigment composition has been made using a novel validation approach which is based on a cluster algorithm and two similarity indices, cophenetic and Rand. In this chapter, it has been demonstrated that application of these indices to quantify the similarity between dierent optical-based cluster trees and the pigment-based cluster tree provides a valuable methodology for identifying optical variables and the spectral ranges most suitable for characterizing the phytoplankton pigment assemblages. In this sense, the dened reference based on the pigment composition (the sea truth) has proven valid as a proxy for phytoplankton community composition. In order to address the main objectives of this chapter, the following methodology has been applied. First, a detailed description of the eld measurements has been provided along with a model-data closure exercise for the reconstruction of hyperspectral Rrs(λ) (Section 3.1). It has been demonstrated that the modeled data set of hyperspectral Rrs(λ) is relatively consistent in comparison to measured multispectral Rrs(λ) . Regarding the pigment data, it has been shown by comparing two independent HPLC pigment data sets that the results and conclusions of this study were robust and independent of the choice of pigment data used in the creation of the reference pigment dendrogram for the cluster-based similarity analysis of pigment and optical data (see Appendix B). Secondly, a successful cluster analysis based on a small but carefully selected set of stations has been performed as a proof-of-concept study (Section 3.3). The most promising results from the cluster analysis were obtained with the second derivative spectra of the phytoplankton absorption coecient, a∗ n,ph(λ) , over the spectral range 370716 nm (or narrower spectral regions from within that range), and the second derivative spectra of the remote-sensing reectance, Rrs(λ) , over the spectral range from about 435 nm to 510 nm. In the course of this study, in addition to the second derivative spectra of a∗ n,ph(λ) and Rrs(λ) , other absorption and reectance data have been examined but they generally show either more limited value or no usefulness at all for discriminating phytoplankton pigment assemblages. For example, the cluster analysis of the ordinary (non-dierentiated) reectance spectra at 1 nm resolution or multispectal (13 wavebands) reectance data exhibited very poor similarity with pigment-based clusters. Similar results were obtained for the ordinary spectra of the total absorption coecient, a∗ n(λ) . However, the ordinary spectra of 83
4 Hyperspectral data to dene bio-optical provinces in the open ocean north-to-south structure can be distinguished when stations corresponding to each cluster are depicted on the map (see Fig. 4.1). Figure 4.2: Ratios of 23 dominant pigments to TChl a corresponding to the 48 stations. Pigment abbreviations are: Chl c 3 = chlorophyllc 3, Chl c1+c 2 = chlorophyllc 1+ c 2, Peri = peridinin, But = 19' - butanoloxyfucoxanthin, Fuco = fucoxanthin, Neo = neoxanthin, Hex = 19' - hexanoyloxyfucoxanthin, Viola = violaxanthin, Asta = astaxanthin, Pra = prasinoxanthin, Dia = diadinoxanthin, Allo = alloxanthin, Diato = diatoxanthin, Zea = zeaxanthin, Lut = lutein, DVChl b = divinyl chlorophyllb , Chl b = monovinyl chlorophyllb , DVChl a = divinyl chlorophylla , Chl a = monovinyl chlorophylla , Phytb = phaoephytinb , Phyta = phaoephytina , α - caro = α - carotene and β - caro = β - carotene. The results of this pigment-based cluster analysis are quite consistent with the classication obtained by just considering the dominant pigments in each cluster of stations, which main features are summarized in Table 4.1. For the 48 stations, there exists a larger variability in the estimate of TChl a concentration in comparison with the previous study (see Section 3.3). It ranges from about 0.104 mg m−3 at the most oligotrophic waters to 5.36 mg m−3 at stations corresponding to some bloom events. Cluster I comprises just station 1, which predominant pigments are Fuco and Chl b . It is clearly dierent from all other stations and this predominance suggests that it could be dominated by Fuco-rich diatoms, as well as Chl b -rich chlorophytes. It is noted that Fuco can also be a precursor of pigments Hex and But (Jerey and Vesk, 1997), which are specic for haptophytes and chrysophytes. Cluster II is also composed of a single station (station 2) and is associated with a local algal bloom given its high TChl a concentration (2.06 mg m−3 ). As Chl b and Hex are the predominant pigments at this station, it could have been dominanted by chlorophytes and Hex-rich haptophytes. However, a very high concentration of the pigment Peri, up to 100 times larger than any other station, also suggests an important presence of dinoglagellates. 90
4.2 Results and discussion Figure 4.3: Cluster tree obtained for the 48 stations using pigment data determined by HPLC. Main clusters are identied with symbols and roman numbers. Cluster III is geographically rather patchy but all stations are clearly dominated by the pigment Hex. Therefore, the main phytoplankton group at these stations is haptophytes, even a possible coexistence of other groups depending on the secondary pigments present in each station belonging to this cluster. Cluster IV is the biggest cluster and unies all stations with DVChl a , Zea and Hex as dominant pigments. In accordance with the results shown in Section 3.3.1, DVChl a and Zea are diagnostic of picophytoplankton that includes DVChl a - and Zea-containing prochlorophytes and Zea-containing cyanobacteria (mainly Synechococcus in the open ocean waters). The abundance of picophytoplankton groups is also accompanied by the presence of haptophytes due to that pigment Hex follows in the ranking. In Table 4.1, stations from cluster IV are indicated separately into a northern and a southern sub-clusters. This geographic division can be explained by considering the levels of TChl a and other environmental factors (e.g., temperature and salinity proles). Nevertheless, the reasons are beyond the scope of this study (further details can be found in Taylor et al. , 2011). Cluster V comprises the bloom stations (i.e., stations 21, 24 and 25) with Fuco as the dominant diagnostic pigment and very high TChl a concentrations, which 91
4 Hyperspectral data to dene bio-optical provinces in the open ocean range from 1.3 to 5.36 mg m−3 . The blooms mainly consist of Fuco-rich diatoms, haptophytes and chrysophytes. The stations belonging to cluster VI (i.e., 41-48) are clearly separated from other stations due to the predominance of the marker pigments Zea and Hex. It suggests a dominance of Synechococcus-type cyanobacteria and haptophytes. Due to their distinct pigment composition, stations 11, 12, 13, 19, 20 and 38 have been clustered away from their geographical neighbours. Stations 19 and 20 are dominated by the pigment Fuco but in a less proportion than the bloom stations in cluster V. Stations 11, 12, 13 and 38 present a dominant role of picoplankton, in particular of prochlorophytes in comparison with their geographical neighbours. This is due to the dominance of DVChl a and DVChl b . Despite the differences in pigment composition, all these six stations have been dened as outliers given that, as described below, they do not show distinct optical features. Table 4.1: Summary of cluster properties obtained for the 48 stations using the pigment information. Properties and pigments abbreviations are: TChl a = Chl a + DVChl a (Mean ± SD); Temp. = Surface temperature (Mean ± SD); Salinity = Surface salinity (Mean ± SD); MVChl a = monovinyl chlorophylla , DVChl a = divinyl chlorophylla , Fuco = fucoxanthin, Hex = 19' - hexanoyloxyfucoxanthin, But = 19' - butanoloxyfucoxanthin, Chl b = monovinyl chlorophyllb , Chl c1 = chlorophyllc1 , Chl c2 = chlorophyllc2 , Zea = zeaxanthin, and Dia = diadinoxanthin; = signicantly dierent from sub-cluster IV-S (p<0.05). (Table adapted from Taylor et al. , 2011). The cluster analysis has also been performed based on aph(λ) spectra measured for all 48 stations and shown in Fig. 4.4. Fig. 4.5 depicts the corresponding hierarchical cluster tree, including the symbols of the pigment-based cluster tree for each station. The partition provided by the cluster analysis based on the aph(λ) data is quite similar 92
4.2 Results and discussion to the partition obtained using the pigment information. This is reected in the high value of the cophenetic index ( rC ) obtained, which is equal to 0.70 considering all 48 stations and equal to 0.84 when the outliers were taken out in the analysis (i.e., stations 11, 12, 13, 19, 20 and 38). In order to optimize the performance, a sensitivity test of the cluster analysis regarding the choice of the spectral range has been needed and is discussed below (see Fig. 4.8a). In accordance with the results shown in the previous chapter (Section 3.3), the best degree of similarity between these two cluster trees (i.e. the highest cophenetic index) is obtained when the aph(λ) spectra are analyzed over a restricted spectral range from 435 to 520 nm. This is the range in which most relevant pigments show their main absorption characteristics. For aph(λ) data, derivative computations do not improve the performance. Figure 4.4: Spectral absorption of phytoplankton ( aph(λ) ) for all stations. By taking into account the optimal spectral range, clusters I, II, V and VI are exactly reproduced by the aph(λ) spectra, whereas clusters III and IV show a few dierences. In particular, stations 6 and 8, which belong to cluster IV in the pigmentbased tree, are classied with cluster III in the aph(λ) -based tree. This may be attributable to dierences in the spectral shape of the aph(λ) spectrum of these two stations around the wavelength 470 nm in comparison to other stations from group IV. Probably because its proximity in space, their shape shows a greater similarity with aph(λ) spectra from stations in cluster III. In addition, most of the outliers described in the pigment-based cluster tree (i.e. stations 11, 12, 13 and 38) are grouped in the aph(λ) -based tree together with stations assigned to cluster IV (i.e., stations indicated with a triangular symbol in Fig. 4.5). Apart from DVChla, which is a principal pigment common to all stations in cluster IV and outliers, the main reason for being singled out in the pigment-based tree is the abundance of pigment DVChl b . This pigment has its main absorption peak around 480 nm (see Fig. 1.3). However, since the pigments Zea and Hex (i.e., the dominant pigments in cluster IV) 93
4 Hyperspectral data to dene bio-optical provinces in the open ocean also absorb strongly at that wavelength range, it would explain the similarity of the aph(λ) spectra despite the dierent pigment composition. Stations 19 and 20 were also considered outliers and grouped separately in the pigment-based tree due to the dominance of the pigment Fuco in these two samples. Their location coincides with the outer rim of the diatom bloom detected in stations 21, 24 and 25 and characterized by a high concentration of the pigment Fuco. Nevertheless, stations 19 and 20 present a smaller percentage of Fuco compared to total pigment and are also characterized by the presence of DVChl a which delineated the stations before and after the bloom. Due to such dierences in pigment composition, it could be expected that these stations also partition separately in the cluster tree based on the absorption information. However, both stations 19 and 20 clustered together with stations assigned to cluster IV, as the rest of outliers. Figure 4.5: Results of cluster analysis applied to aph(λ) from the 48 stations. Symbols correspond to pigment-based clusters (see Fig. 4.3). 94
4.2 Results and discussion To explore the potential of remote sensing information in distinguishing dierent phytoplankton assemblages, the partition of the stations has also been carried out based on available Rrs(λ) spectra measured for 21 of the stations and shown in Fig. 4.6. As for the absorption data analysis, the cluster analysis based on Rrs(λ) data also yields satisfactory results (Figure 4.7). A cophenetic index equal to 0.62 is obtained for all stations and equal to 0.90 when again the outliers stations have been ruled out in the analysis. As discussed in great detail in Section 3.3, a good performance has only been achieved when considering the second derivative of Rrs(λ) spectra in this case over the spectral range of 435 to 580 nm. The ability to discriminate phytoplankton pigment assemblages from the derivative of Rrs(λ) spectra has been optimized by selecting the most suitable parameters used in the spectral derivative computations. In particular, it has been determined that the optimal values for the smoothing lter window and band separation were 9 nm for remote-sensing reectance (results from this sensitivity analysis are not shown). Due to time constraints at the start of the cruise, no Rrs(λ) measurements were collected at the stations corresponding to pigment-based clusters I and II. In any case, clusters III, V and VI are well reproduced and cluster IV shows some discrepancies similar to the analysis based on aph(λ) data. The outliers stations 12, 20 and 38 which clustered separately in the pigment-based tree (due to their high concentration of DVChl b ) are associated with stations from cluster IV in the optical data, as happened in the aph(λ) -based cluster analysis. Unexpectedly, station 6 (from cluster IV in the pigment-based tree) has been singled out in the Rrs(λ) -based tree. Its distinct Rrs(λ) spectrum, with a much lower in magnitude in comparison to other stations from cluster IV, might indicate that some problems or errors occured during the acquisition. Figure 4.6: Remote-sensing reectance spectra ( Rrs(λ) ) for the 21 stations measured. 95
4 Hyperspectral data to dene bio-optical provinces in the open ocean Figure 4.7: Cluster tree resulting from analysis of the 21 stations calculated using second derivative of hyperspectral Rrs(λ) spectra (optimal values: WS = BS = 9 nm). Symbols correspond to pigment-based clusters (Fig. 3.8). In order to optimize the results, the degree of similarity between cluster trees has been again evaluated for calculations involving dierent spectral ranges of optical data (Torrecilla et al. , 2011a). Fig. 4.8a illustrates the degree of similarity between the aph(λ) -based and pigment-based cluster trees for dierent spectral ranges of absorption data. Figure 4.8: Cophenetic indices between pigment-based and optical-based cluster trees obtained using dierent combinations of spectral range for (a) aph(λ) and (b) second derivative spectra of Rrs(λ) . Optimal values of WS =BS = 9 nm, were used for the calculation of derivative spectra. The y-axis indicates the lower limit of the spectral range ( λmin ) and the x-axis the upper limit of the spectral range ( λmax ) utilized. 96
4.2 Results and discussion According to the results, the optimal region is clearly when the phytoplankton absorption spectrum is analyzed over the spectral range approximately from λmin = 425 nm to λmax = 540 nm. Fig. 4.8b shows distributions of the cophenetic index for the cluster analysis of the second derivative of Rrs(λ) . The spectral region from 435 nm to 580 nm, which includes the narrower region from 435 nm to 475 nm, can be considered as the optimal spectral range for the application of derivative approach with potential for good similarity between the pigment-based and reectance-based cluster trees. Both optimal spectral regions identied, in particular the one based on the analysis of phytoplankton absorption data, overlap with the wavelength range where absorption characteristics of main accessory pigments appear (see Fig. 1.3). As expected, the results from this sensitivity analysis are consistent with the results shown in the previous study (Section 3.3). It is important to note that these results are also in agreement with the ones shown in the chapter 2, where the optimal range was from λmin = 470 nm to λmax = 500 nm. In that case, however, this range was narrower given that the analysis was performed considering simplied underwater optical scenarios, in which only a single phytoplankton group was present. The results obtained with aph(λ) and Rrs(λ) spectra are summarized in Table 4.2, including the improvement achieved with the optimal ranges and derivative settings. Table 4.2: Summary of results including cophenetic indices between pigment-based and optical-based cluster trees obtained using dierent spectral ranges of aph(λ) , Rrs(λ) and second derivative spectra of Rrs(λ) . In parenthesis, the cophenetic indices obtained once the outliers stations (i.e., 11, 12, 13, 19, 20 and 38) were ruled out from the analysis. 4.2.1 Bio-optical provinces and its biogeographical relevance On land, the distinction between dierent ecosystems such as the edge of a forest or the extent of a desert is generally easier compared to the marine environment. Due to the highly dynamic nature of the oceans and the lack of knowledge about many of those dynamics, these boundaries between ecological systems are much harder to observe. They are, however, no less real (Platt and Sathyendranath, 1999; Hooker et al. , 2000; Devred et al. , 2007; Longhurst, 2006). Areas in the world's oceans with 97
4 Hyperspectral data to dene bio-optical provinces in the open ocean similar physical and biological characteristics are generally referred to as ecological or biogeochemical provinces or units (e.g. Platt et al. , 2005). Several studies have used TChl a as one of the descriptors to dene marine ecological provinces (e.g. Devred et al. , 2007; Hardman-Mountford et al. , 2008). In this study, however, the global contribution of all pigment composition has been combined with hyperspectral optical data to establish a bio-optical geography of the eastern Atlantic Ocean. By applying an unsupervised HCA analysis to data from the eastern Atlantic Ocean, the potential applicability of pigment composition, phytoplankton absorption coecients and remote sensing reectances for identifying dierent biooptical provinces in the ocean has been demonstrated (see a summary in Table 4.2). The establishment of dierent bio-optical provinces has led to examination of its biogeographical relevance by comparison to ecological provinces previously proposed in the literature. In order to assess the bio-optical approach with an established biogeography, the bio-optical provinces/clusters have been compared to the widely used system dened by Longhurst (2006). Longhurst proposed a global partitioning into 4 biomes or basic vegetation types within the pelagic realm of the oceans: the Polar, Westerlies, Trades and Coastal biome. Within the biomes, 51 provinces were de- ned on the basis of a global data set including satellite images of surface chlorophyll elds, regional oceanography data and also considering many previous proposals for partitioning the oceans. Although the boundaries of the provinces were forced into a static grid as a matter of convenience, Longhurst pointed out that the position of the borders is dynamic and varies on annual, seasonal or even shorter time scales. In Figure 4.1, the 48 stations with their assigned cluster symbols are plotted on a map, in which Longhurst's provinces are also depicted (shapele from VLIZ, 2009). A wide discussion regarding the comparison between bio-optical provinces and Longhurst's provinces can be found in Taylor et al. (2011). Nevertheless, the main results with a summarized description of similarities observed are provided. Clusters I and II belong to the continental shelf of Western Europe, described by Longhurst as the Northeast Atlantic Shelves Province (NECS). This analysis separates the two stations into two clusters probably due to the fact that station 2 (cluster II) is situated in a area where tidal forces dominate the English Channel. Stations joined in cluster III do not all belong to the same geographical province. Stations 3-5 lie in the south-east corner of Longhurst's North Atlantic Drift province (NADR) and show the same bio-optical traits than stations 7 and 22. Nevertheless, oceanographic data and geography show that these two stations belong to dierent oceanic provinces. The sub-cluster IV-N lies within Longhurst's North Atlantic Subtropical Gyral Province, East (NAST(E)). However, this area is bio-optically not consistent and many outliers lie alongside the stations in this sub-cluster. These inconsistencies might be explained by typical physical processes in this region (e.g., eddy-driven vertical advection, internal waves, Mediterranean water eddies which ow out of the Strait of Gibraltar and move southwards once they have detached from the continental slope, etc.). Stations 19-25, including bloom stations (i.e., 21, 24 and 25), lie on the approximate boundary between the North Atlantic Tropical Gyral Province (NATR) and the Canary Current Coastal Province (CNRY), which comprises the southerly coastal ow of the 98
4.2 Results and discussion eastern boundary current of the North Atlantic from Cap Finisterre (Spain) to Cape Verde (Senegal). The sub-cluster IV-S including stations 26-40 is geographically a very consistent cluster and lies within Longhurst's Eastern Tropical Atlantic Province (ETRA). The boundaries of this province are clearly dened and in correspondence with the observed changes in the pigment composition and bio-optical traits from stations to the north and south. Perez et al. (2005) describe the typical phytoplankton community structure for this region as rather constant with the contribution of picophytoplankton to TChl a always exceeding 45%, a description which is corroborated by the results. Stations belonging to cluster VI are clearly separated through their bio-optical traits from the other stations. They present distinct pigment composition, aph(λ) and Rrs(λ) spectra and are clearly located in Longhurst's South Atlantic Gyral Province (SATL), which comprises the anticyclonic circulation of the South Atlantic, excluding the coastal boundary currents. The northern boundary of this province is clearly dened by the change in bio-optical data from stations 40 and 41. This coincides fairly well with the boundary set by Longhurst for the SATL province. The bio-optical provinces have been shown to agree well with the provinces established by Longhurst and thus can be used to classify areas of similar biogeography. However, in areas with high variability, the Longhurst's provinces do not harbour consistent phytoplankton assemblages. In such areas, it would be advantageous to use a more exible and dynamic approach to describe the ecosystem, in which physical and optical information could be combined (Devred et al. , 2007; Raitsos et al. , 2008). Figure 4.9 shows variability in the temperature and salinity regimes from stations selected for this study that combined with latitude and longitude information could be used to improve the ability to describe ecological complex regions. Figure 4.9: Temperature-salinity diagram for the north-south transect in the eastern Atlantic Ocean (see Fig. 4.1). Points corresponding to stations are labeled with their station number and their symbol according to the pigmentbased cluster partition (see Fig. 4.3). 99
5 Hyperspectral data for a remote sensing assessment of phyto. communities in a estuary 5.1 Optical modeling and data analysis 5.1.1 Modeled hyperspectral remote-sensing reectance In order to examine the sensitivity of the remote-sensing reectance corresponding to shallow waters to the phytoplankton taxa, several secenarios have been simulated with the Hydrolight/Ecolight version 5.0 radiative transfer (RT) model (Mobley and Sundman, 2008). A Case 2 waters model representing a shallow estuarine environment have been parameterised by dening the number of optically active substances and providing their specic optical properties. The pure water absorption and scattering coecients have been derived from Pope and Fry (1997). The specic absorption and scattering spectral coecients corresponding to two phytoplankton groups representative for the considered area (i.e., diatoms and dinoagellates, see Fig. 5.1c) have been taken from Kim and Philpot (2006). Due to the lack of knowledge about the colored dissolved organic matter (CDOM) in the Alfacs Bay, the absorption coecient of CDOM has been calculated using a common exponential relationship included in the Hydrolight code, based on a model from Morel and Maritorena (2001): acdom(z, λ)=0.2·aph(z, 440) exp[−0.014(λ−440)] (5.1) where aph(z, 440) is the absorption coecient of the chlorophyll-bearing particles (i.e., phytoplankton) at 440 nm. The concentration of suspended sediments, SS , has been taken to vary from 0 to 6 mg/L based on eld observations (M. L. Artigas, unpublised data ). The mass-specic absorption and scattering coecients of SS have been derived from a composite of data in Bukata et al. (1995) and Gallie et al. (1992), also included in the Hydrolight code. The modeling of the remote-sensing reectances was carried out over the spectral region of 375-725 nm with 1 nm resolution and included the inelastic Raman scattering and uorescence by chlorophyll and CDOM within the ocean. The sea surface boundary conditions have been congured using the default settings in Hydrolight (i.e., wind speed equal to 5 m/s in accordance to real conditions in the area, solar zenith angle equal to 30o and clear sky condition). In order to emulate similar conditions to those found in the Alfacs Bay, the ocean has been assumed to be a nite-depth water body with a zmax equal to 6.5 m and examined at intervals 4z equal to 0.05 m. Dierent bottom types have been considered corresponding to various sediments and benthic biota, among them those ones similar to the ones found in the studied area (de Pedro Puente, 2007). Each bottom type, characterized by a specic irradiance reectance (see Fig. 5.2), was provided by the Hydrolight code. 106
5.1 Optical modeling and data analysis Figure 5.2: Bottom irradiance reectances provided by the Hydrolight code and considered for the present study. Two dierent vertical proles of chlorophyll concentration have been used in the model simulations. On the one hand, several uniform vertical proles of chorophyll concentration equal to 0.5, 1, 3, 5 and 7 mg/m3 have been used to test the sensitivity of Rrs(λ) to the phytoplankton taxa, bottom type and presence of suspended sediments (see Section 5.2.1). On the other hand, in order to examine their eect on the Rrs(λ) , dierent nonuniform vertical proles of the chlorophyll concentration have been generated (see Section 5.2.2). These vertical proles, which simulate the condition characterized by the presence of a subsurface chlorophyll maximum at a specic depth (i.e., a thin layer condition, see Fig. 5.3), have been approximated by a Gaussian function superimposed upon a constant background (Lewis et al. , 1983): Chl(z) = Chl0+Chlmax exp[−(z−zmax)2/2σ2] (5.2) where Chl0 is the background value of chlorophyll, zmax is the depth of the chlorophyll maximum, σ is the standard deviation that controls the thickness of the cholorphyll peak (i.e., 95% of the integrated biomass is located within a water layer of thickness equal to 4σ ) and Chlmax is the amplitude of the chlorophyll maximum above the value of Chl0 . According to values derived from the analysis of water samples taken in the Alfacs Bay (Artigas et al., submitted ), the calculations were made for two values of Chl0 (1 and 3 mg/m3 ), three values of zmax (3, 4, and 5 m), one value of σ (0.25 m) and three values of Chlmax (6, 9 and 12 mg/m3 ). 107
5 Hyperspectral data for a remote sensing assessment of phyto. communities in a estuary Figure 5.3: Diagram of the nonuniform vertical chlorophyll prole expressed in equation (5.2) and utilized in part of the RT simulations. The role of each optically active component for the type of environments simulated in this chapter diers from the cases simulated in the previous chapters (i.e., Case 1 waters). Figure 5.4 shows an example of the most signicant IOPs (i.e., absorption and backscattering spectral coecients) considered as the input information to calculate one of the modeled remote-sensing reectances. It is noted that the suspended matter has a higher inuence on the total absorption and backscattering coecient spectra, and hence also expected on the Rrs(λ) signal, even becoming the most dominant component when increasing its concentration. Figure 5.4: Examples of absorption (a) and backscattering (b) spectra utilized in one of the simulations for selected values of: Chl0 (1 mg/m3 , uniform vertical prole) and SS (2 mg/L ). In this case, the underwater scenario was characterized by the presence of the algal group of diatoms. 108
5.2 Results and discussion 5.1.2 Cluster analysis and Multi-dimensional Scaling A hierarchical cluster analysis (HCA) has been applied to the optical datasets of Rrs(λ) generated for the study presented in this chapter. The assessment of the impact of having dierent underwater scenarios given dierent bottom types, vertical structures of phytoplankton and abundance of suspended sediments has been carried out by analyzing the similarity between each simulated hyperspectral Rrs(λ) . In this sense, similarly to previous chapters, the HCA analysis has allowed to examine our ability to distinguish Rrs(λ) corresponding to dierent underwater conditions. The cosine distance has also been chosen as measure of similarity between spectra in order to give priority to dierences in the spectral shape of optical data rather than magnitude. It is noted in this study, however, that once a hierarchical cluster tree has been obtained corresponding to a set of underwater scenarios, a new graphical representation instead of a tree has been used to more easily examine our ability to distinguish one scenario from another. Multi-dimensional Scaling (MDS) is an ordination technique that, similar to cluster techniques, ranks objects according to their relative similarity (Shepard, 1980). In this study, it has been used as a complement of the cluster analysis for graphically representing dissimilarities in the Rrs(λ) data set. It provides an alternative visualization of the complex relations among the Rrs(λ) spectra obtained with the cluster analysis. MDS is based on the dissimilarity matrix obtained in the HCA analysis (using the cosine distance) but displays the inter-relations between each pair of objects (i.e., each pair of spectra) on a low dimensional plot (2D). The distance between objects on this biplot is an indicator of the similarity in the Rrs(λ) spectra. The MDS representation provides a reliability measure, which indicates how well the 2D-ordination plot depicts the relationships in the similarity between objects. This is indicated by the minimized stress value. 5.2 Results and discussion 5.2.1 Eect of the bottom type on the remote sensing signal The variability in remote-sensing reectance spectra due to dierent characteristics of the bottom type is signicant in optically shallow waters. In fact, it has been demonstrated useful to map bottom properties in coastal environments, in particular if observations at hyperspectral resolution are available. Figure 5.5 illustrates the inuence of the dierent bottom types considered in this study (i.e., brown algae, coral sand, green algae, red algae and seagrass) on the Rrs(λ) spectra when only pure water is assumed to be present in the modeled water body. Coral sand provides a higher Rrs(λ) spectra than algae due to its higher reectance over the visible part of the spectrum. Rrs(λ) spectra corresponding to green algae is higher than those for other algae, in accordance with its reectance spectrum (see Fig. 5.2). Rrs(λ) spectra from red, brown algae and seagrass are more similar but show signicant dierences in shape. 109
5 Hyperspectral data for a remote sensing assessment of phyto. communities in a estuary Figure 5.5: Remote-sensing reectance, Rrs(λ) , of ve dierent bottom types seen through clear water with a maximum depth equal to 6.5 m. Our ability to discriminate dierent algal communities and bottom types from Rrs(λ) information, however, can be complicated by the presence of high concentrations of phytoplankton biomass, for example in the case of existing bloom-forming or thin layer processes. The optical water clarity is degraded as the abundance of phytoplankton along the water column increases. Figure 5.6a shows the eects on Rrs(λ) spectra caused by dierent parameters of the bottom type and the level of chlorophyll concentration (i.e., 0.5, 1, 3, 5 and 7 mg/m3 ) corresponding to scenarios characterized by the presence of two algal groups: diatoms and dinoagellates. In this case, it corresponds to uniform vertical proles of chorophyll concentration. In general, the eect of increasing the chlorophyll concentration is to decrease Rrs(λ) spectra in the blue spectral region and to increase it in the green and red spectral regions. The highest Rrs(λ) spectra occur for scenarios characterized by the presence of diatoms, regardless the bottom type. Nevertheless, it is noted that bottom type has no impact on the Rrs(λ) spectra for high concentrations of chlorophyll (> 3 mg/m3 ). In this cases, the Rrs(λ) spectra corresponding to a xed chlorophyll concentration and all bottom types are seen in cyan color, since they have been plotted superimposed (i.e., it is just seen the Rrs(λ) corresponding to the last bottom type). In order to assess if dierent shallow water scenarios can still be dintinguished despite the presence of dierent algal groups and bottom type congurations (see Fig. 5.6a), the HCA cluster analysis has been applied to the dataset of hyperspectral Rrs(λ) spectra. Figure 5.6b depicts by means of Multi-dimensional scaling (MDS) all cases according to their relative similarity found in the cluster analysis. It is noteworthy that scenarios dominated by either diatoms and dinoagellates are clearly distinguished regardless the chlorophyll concentration of the uniform vertical prole (i.e., for concentrations from 0.5 to 7 mg/m3 ). In addition, for chlorophyll concentrations up to 1 mg/m3 , the analysis of hyperspectral Rrs(λ) spectra also allows to discriminate scenarios with dierent bottom characteristics. On the contrary, when 110
5.2 Results and discussion the chlorophyll increases, the information of the bottom type is reected neither in the Rrs(λ) spectra nor in the MDS plot. Figure 5.6: (a) Examples results of radiative transfer simulations of optically shallow waters, showing the Rrs(λ) spectra corresponding to uniform vertical proles of two phytoplankton groups at dierent concentrations and given dierent bottom types. (b) MDS plot visualizing similarities between the Rrs(λ) spectra shown in (a). Symbols and colors represent each algal group and bottom type, respectively. Labels indicate the chlorophyll concentration considered in each modeled underwater scenario. Monitoring of changes in phytoplankton taxa under dierent bottom conditions by remote sensing can also be complicated by the eect of another optically signicant component found in estuarine waters: the suspended matter. Figure 5.7a shows the eects on Rrs(λ) spectra caused by dierent parameters of the bottom type and the level of suspended sediment (i.e., 0, 2, 4 and 6 mg/L ). The scenarios correspond to water bodies with diatoms or dinoagellates at a xed concentration equal to 111
5 Hyperspectral data for a remote sensing assessment of phyto. communities in a estuary 0.5 mg/m3 along the water column. The eect of increasing the concentration of the suspended matter is to increase absorption spectra in the blue spectral region and hence decrease Rrs(λ) spectra in this region. Again, the highest Rrs(λ) spectra occur for the group of diatoms regardless the bottom type and the concentration of suspended matter. Nevertheless, it is noted that bottom type has no impact on the Rrs(λ) spectra for high levels of suspended sediments, cases in which Rrs(λ) spectra are seen in cyan color as occured in spectra shown in Fig. 5.6a. Figure 5.7: (a) Examples results of radiative transfer simulations of optically shallow waters, showing the Rrs(λ) spectra corresponding to uniform vertical pro- les of two phytoplankton groups at a concentration of 0.5 mg/m3 and given dierent concentrations of suspended matter and bottom types. (b) MDS plot visualizing similarities between the Rrs(λ) spectra shown in (a). Symbols and colors represent each algal group and bottom type, respectively. Labels indicate the concentration suspended sediment considered in each modeled underwater scenario. 112
5.2 Results and discussion The MDS plot shown in Fig. 5.7b better illustrates the dissimilarities between the Rrs(λ) spectra of respective phytoplankton groups and bottom types, given dierent concentrations of suspended matter. By analyzing the dissimilarities in the whole spectral range, all scenarios corresponding to dierent phytoplankton taxa and different concentrations of suspended matter have been distinguished. In addition, it is noted that for concentrations of suspended matter smaller than 2 mg/L , scenarios characterized by dierent bottom types have also been clearly discriminated. 5.2.2 Eect of the vertical structure of phytoplankton on the remote sensing signal After having assessed the inuence on Rrs(λ) spectra by the presence of dierent algal groups, levels of phytoplankton biomass, suspended matter and dierent bottom types, this section is aimed at evaluating how well dierent shallow estuarine scenarios with distinct vertical structures in terms of phytoplankton distribution and abundance can be detected through the analysis of the Rrs(λ) signal. Several underwater scenarios were generated simulating a thin layer condition taking place in a shallow estuarine area like the Alfacs Bay (see Fig. 5.3). Figure 5.8a shows the Rrs(λ) spectra modeled by considering the presence of one algal group, diatoms or dinoagellates and dierent combinations for the values of the background chlorophyll, Chl0 (1 and 3 mg/m3 ), depth of the chlorophyll maximum, zmax (3, 4, and 5 m), amplitude of the chlorophyll maximum, Chlmax (6, 9 and 12 mg/m3 ) and a xed value of the parameter related to the thickness of the chlorophyll peak, σ (0.25 m) . The level of suspended sediments was xed (i.e., 2 mg/L for all scenarios), likewise the bottom type. It is also plotted the Rrs(λ) spectra corresponding to uniform chlorophyll proles when the concentration was equal to 1 and 3 mg/m3 . It can be easily seen that uniform and nonuniform vertical proles of phytoplankton for each combination of parameters (indicated in black symbols and red lines, respectively) provide very similar, sometimes identical, Rrs(λ) spectra. The corresponding MDS plot obtained after the application of the similarity-based cluster analysis (see Fig. 5.8b) also conrms that the contribution of the thin layer to the overall Rrs(λ) spectra (red symbols) is nearly negligible when compared to the Rrs(λ) obtained for uniform vertical proles (black symbols). Only for two scenarios, corresponding to the case of having a background chlorophyll, Chl0 , equal to 1 mg/m3 and a thin layer of Chlmax equal to 9 and 12 mg/m3 at the upper level (i.e., zmax equal to 3 m), spectral discrepances allow to distinguish the nonuniform from the uniform situation. The discrimination of dierent nonuniform vertical proles of the chlorophyll concentration in shallow estuarine waters based on the analysis of hyperspectral Rrs(λ) spectra has been unsuccessful. Scenarios characterized by two algal groups at different concentrations have been discriminated. Nevertheless, the analysis of the hyperspectrally-resolved Rrs(λ) data set has not allowed to distinguish dierent vertical structures along the water column of these phytoplankton groups. Attempting to improve our results by applying derivative spectroscopy or limiting the analysis to specic spectral regions has not provided either a good performance (not shown). 113
5 Hyperspectral data for a remote sensing assessment of phyto. communities in a estuary Figure 5.8: (a) Examples results of radiative transfer simulations of optically shallow waters, showing the Rrs(λ) spectra corresponding to uniform (indicated in black symbols and lines) and nonuniform vertical proles (in red) of two phytoplankton groups. (b) MDS plot visualizing similarities between the Rrs(λ) spectra shown in (a). The results of the here presented optical modeling exercise show that the vertical structure of the distribution of the phytoplankton biomass in the water column of shallow estuarine environments has little impact on the remote-sensing reectance spectra. In this sense, as suggested by other authors (see several contributions in Babin et al. , 2008), a multidisciplinary approach based on highly resolved vertical proling systems could provide ne-scale optical and hydrographic measurements so as to complement remote-sensing capabilities. A better monitoring of coastal blooms dynamics is among the advantages of obtaining a high resolution in the vertical distribution of phytoplankton. Ideally, these observation systems should resolve vertical 114
5.3 Summary and conclusions distributions of phytoplankton in relationship to temperature, salinity and currents. Additionally, nutrient availability should also be part of the monitoring strategy. In case of using highly resolved vertical prolers including hyperspectral measurements, it is important to note that the proposed dissimilarity-based methodology could also be applicable. The underlying idea would be to investigate the suitability of this approach applied to periodical hyperspectral measurements along the water column in order to retrieve information about the temporal and spatial variability of the distribution of phytoplankton and other optically signicant substances present in the area under study. Figure 5.9 illustrates an example in which the vertical structure of phytoplankton distribution corresponding to one modeled scenario is assessed by using the cluster-based approach. In particular, the dissimilarity analysis has been applied to a set of irradiance reectance spectra obtained along the water column (i.e., every 0.05 m), allowing to detect the idealized thin layer of phytoplankton located at 4 m depth. In this example, the level of suspended sediments was equal to 2 mg/L and the bottom type was brown algae. Figure 5.9: Results of cluster analysis based on the irradiance reectance along the water column corresponding to a very shallow area characterized by the presence of a thin layer of phytoplankton at 4 m depth, obtained by means of radiative transfer simulation. 5.3 Summary and conclusions In this chapter, hyperspectral optical signatures of shallow estuarine environments in relation to dierent phytoplankton comunities have been evaluated. In particular, the Alfacs Bay, an estuarine embayment of the Ebro Delta (NW Mediterranean Sea) with a maximum depth of 6.5 m, has been considered as a case study area. The presence of several harmful algal species responsible for dierent types of toxicity in 115
6 Summary and outlook represented the actual eld conditions. It is noteworthy that the ability to dene realistic inputs derived, in turn, from a comprehensive suite of IOP measurements carried out in situ and in the laboratory. These results serve to demonstrate that the conditions under which theoretical radiative transfer calculations of hyperspectral Rrs(λ) were performed produced a close agreement with the experimental multispectral measurements of Rrs(λ) . The identication of dierent bio-optical provinces from the analysis of 48 open ocean stations in the eastern Atlantic Ocean (Chapter 4), as one of the potential applications of the proposed methodology, has led to examination of its biogeographical relevance by comparison to ecological provinces previously proposed in the literature (Longhurst, 2006). The bio-optical provinces showed to agree well with these provinces and thus could be used to classify areas of similar biogeography. In fact, the use of hyperspectral observations in combination with other physical parameters (e.g., the sea-surface temperature) is suggested. In terms of globally signicant issues such as carbon export and primary production, the hyperspectral-based approach represents a possibility for a faster and detailed assessment of the state of a temporally and spatially variable marine environment. At the moment, other approaches have already been proposed to detect simultaneously Phytoplankton Functional Types (PFTs) from satellite hyperspectral data (Bracher et al. , 2009). However, PFT identication is a slightly dierent concept than identifying dierent spatially distributed phytoplankton assemblages. The novel methodology proposed in this thesis allows to dene dierent bio-optical provinces on the basis of phytoplankton community structure and their bio-optical traits, instead of a small set of functional groups. The results from the optical modeling exercises for shallow estuarine waters have shown that scenarios dominated by either diatoms and dinoagellates were clearly discriminated regardless of the chlorophyll content, abundance of suspended matter, bottom type and vertical structure of the phytoplankton along the water column. In fact, the analysis of Rrs(λ) spectra also permitted to derive information about the bottom properties of each simulated underwater scenario as long as the chlorophyll concentration was not larger than 1 mg/m3 or the concentration of suspended matter was smaller than 2 mg/L . On the contrary, the contribution of a idealized thin layer in the water column to the remote-sensing signal showed to be nearly negligible when compared to Rrs(λ) spectra obtained for analogous uniform vertical proles. Overall, this thesis demonstrate how hyperspectral oceanographic observations can be used to better assess phytoplankton biodiversity and dynamics, furthering the understanding of the role of each algal group at a local scale or in the global marine ecosystem and biogeochemical cycles. It is noteworthy that the proposed approach is generally applicable to dierent data sets (i.e., in situ and remotely-sensed Inherent or Apparent Optical Properties) or other types of data (e.g., combination of optical and hydrographic data), gathered at the sea surface or along the water column. A similar optimization analysis can also be used to provide the best performance for a given data set. Moreover, this analysis is exible and permits dierent optimal values to be utilized in the case of other optical data with diering spectral resolutions. 122
In terms of future work , the following points seem worthwhile to pursue as part of further investigations: - The results from the assessment of phytoplankton communities in open ocean environments, even encouraging, are far from denitive. In fact, further investigation is needed with larger databases of measurements from various oceanic environments in order to determine the generality of the presented approach and the specic set of optimal parameters. The minimum data necessary for application of the dissimilaritybased proposed methodology is information regarding either the hyperspectral phytoplankton absorption coecient or the remote-sensing reectance in the blue to green region of the spectrum, which would yield an initial classication of stations based on similarities and dierences in the optical spectra or their derivative spectra. Given signicant interest in the development of the capabilities for large-scale characterization of phytoplankton biodiversity from optical measurements including remote-sensing observations, one may expect further expansion of comprehensive databases consisting of concurrent pigment and hyperspectral optical information in the near future. The availability of such a larger data sets will support validation and optimization of this dissimilarity-based approach and other classication techniques, such as unmixing or neural networks, which may exploit hyperspectral optical measurements as a source of information on phytoplankton community composition. In order to address these issues, the research project PHYTOSCOPE (PHYToplankton biOdiversity multiScale Characterization using advanced OPtical technologiEs) has recently been initiated within the framework of the Marine Technology Unit (UTM, CSIC) and in coordination with the Phytooptics Group (Alfred-Wegener Institute for Polar and Marine Research, AWI, and Institute for Environmental Physics of the University Bremen). - Concerning the characterization of more complex shallow estuarine waters, there are some important aspects that may be interesting to explore. In particular to Alfacs Bay, current uncertainties in the knowledge of marine optical characteristics and the ability to model the underwater radiation elds in such a complex region indicate that more additional research is still needed. Despite the results from the performed modeling exercise indicate that hyperspectral observations might play an important role in the characterization of optically signicant water constituents (i.e., phytoplankton, suspended matter and CDOM), there is a strong need for more in situ optical and hydrographic data. This observational eort will provide important insights into the factors regulating phytoplankton community composition in this area and will help to explore the full potential of hyperspectral observations. To overcome this challenge, the research projects: PHYTOSCOPE and MESTRAL (Modelling and advanced observational tEchnologies to link tranSport processes, opTically-active constituents, and wateR light-eld vAriability in a coastaL ecosystem) have recently been initiated within the framework of the Marine Technology Unit (UTM, CSIC). 123
Appendix A Radiative transfer equation in the ocean 1 Radiative transfer (RT) theory aims to describe the interaction of light with matter and to quantify all the processes that aect the direction and the quantity of photons. In particular, the change of the light eld under water is due to several processes and is described by the so called Radiative Transfer Equation (RTE). This appendix is devoted to explain in detail the derivation of this unique fundamental equation, which enables to link the two categories of optical properties (i.e., inherent and apparent optical properties) dened for a water body as it has been shown in Fig. 1.8. The description of these two categories (i.e., IOPs and AOPs) and the involved properties can be found in Section 1.1. The path of the radiance L(z, θ, ϕ, λ) through a thin layer of water of thickness dz = z2−z1 is shown schematically in Fig. A.1. The zenith and azimuthal angles θ0, ϕ0 and θ, ϕ denote the direction of the incoming and outgoing light beam, respectively. The processes responsible for the variation of L(z, θ, ϕ, λ) in a water body are primarily extinction processes (i.e., absorption or scattering of photons out of the light beam) and intensivation by scattering into the direction of light propagation or by emission. In this sense, it can be considered as a process of balance of light loss and gain. Figure A.1: Path of the radiance and inuences of absorbing and scattering particles in a thin layer of water ( Source : Albert, 2004). 1 The description of the basic processes provided in this appendix are described in the following textbooks: Mobley, 1994 and Kirk, 1994. 125
Appendix A The RTE is dened by the following integro-dierential equation 2 : cos θdL(z,θ,ϕ,λ) dz =−c(z, λ)L(z, θ, ϕ, λ) +´4πL(z, θ0, ϕ0, λ)·β(z;θ0, ϕ0→θ, ϕ;λ)dΩ0+S(z, θ, ϕ, λ) (1.12) This equation expresses that the change of a specic radiance L(z, θ, ϕ, λ) along an elementary distance dz/ cos θ results from: the attenuation (i.e., a loss) along the elementary slant path. It, thus, involves the beam attenuation coecient c(z, λ) . the gain by elastic scattering into the direction of interest (θ, ϕ) from all other directions about the element. It depends on the scattering phase function β(z;θ0, ϕ0→θ, ϕ;λ) 3 , which represents the scattering probability into a specic direction of the solid angle dΩ0 . the gain by inelastic scattering from other wavelengths and internal light sources into the direction of interest (θ, ϕ) . It is indicated as a source term S(z, θ, ϕ, λ) . Among all the processes involved in the RTE equation, the process of scattering can be described as the deviation of a photon from a straight trajectory in the medium through which it passes. It takes place in natural water due to the interaction of photons with water molecules and water constituents. The inuence of each constituent on the scattering process depends on wavelength, particle size, concentration and refractive index. When the total energy of the scattered photon is conserved, (i.e., it is scattered at the same wavelength, λ , as the incoming photon), it is called elastic scattering . However, when the energy is not conserved (i.e., it is scatterd at another wavelength, λ0 ), it is called inelastic scattering . In this last process, the scattered wavelength is longer than the absorbed ( λ0> λ ) and therefore some energy is transformed in contrast to elastic scattering, where only the direction of the incident radiation is changed. Regarding the source term, S(z, θ, ϕ, λ) , the rst component is related to two types of inelastic processes: Raman scattering and uorescence . The result of both processes is in essence the same: a photon with the wavelength (or frequency) dierent from that of the incident photon is produced and the molecule is brought to a higher or lower energy level. Nevertheless, the Raman eect diers from the process of uorescence. For the latter, the incident light is completely absorbed and 2 It is considered the situation of horizontally homogeneous water and time independence. The dependence on θ and ϕ is noted and combined by the symbol Ω . 3 The angle between any direction from the surrounding space and the direction of interest is denoted (θ0, ϕ0→θ.ϕ) . 126
the system is transferred to an excited state from which it can go to various lower states only after a certain resonance lifetime. Another dierence is that the Raman eect can take place for any wavelength of the incident light. In contrast to the uorescence eect, the Raman eect is therefore not a resonant eect. In practice, this means that a uorescence peak is anchored at a specic excitation wavelength, whereas a Raman peak maintains a constant separation from the excitation wavelength. Raman scattering in natural water is mainly caused by the water molecules. In contrast, uorescence is the inelastic process ocurring in pigments contained in the water constituents such as phytoplankton and CDOM. The second component of the source term S(z, θ, ϕ, λ) corresponds to the process of bioluminiscence . It is a internal light source where light is emitted by organisms in the water. The quantity of light emitted, and produced by chemical reactions of bacteria, phytoplankton and sh, is very small compared to the incident sun light. Therefore, its impact on radiance is neglected at the surface level but considered as the only natural source of light for deep waters. It is important to note that despite the compactness of the RTE equation, its solution is not straightforward. Nevertheless, mathematical and numerical techniques have been developed to derive quasi-exact solutions, which in fact have been revealed extremely accurate. For instance, the RT model Hydrolight/Ecolight version 5.0, which is described in Section 1.4 and utilized within the framework of this thesis, employs invariant imbedding techniques to solve the RTE equation. 127
Appendix B Independent corroboration for the pigment analysis The Center for Hydro-Optics and Remote Sensing (CHORS) at San Diego State University in California was the designated HPLC analytical facility for NASA-sponsored investigations during the period encompassing the ANT-XXIII/1 cruise from 2005. It was subsequently determined that the measurements performed by CHORS during this time were plagued with a number of methodological problems, which potentially led to errors in the determinations of some pigments. Because these problems and associated uncertainties changed with time, the NASA team tasked with investigating these data quality problems at CHORS deemed that retrospective corrections for the entire span of several years is not feasible (Hooker and Heukelem, 2009). Their nal recommendation to investigators was that use of the data be accompanied by a statement that These data are not validated and should not be used as sole basis for scientic results or conclusions. Independent corroborating evidence is required. Because all of the samples were analyzed as part of a single run, ad hoc corrections were developed for a subset of specic individual pigments using a combination of eld data and laboratory standards measured during the SeaHARRE-3 intercalibration experiment (C. Trees, personal communication; Hooker et al. , 2009). These pigments included both the monovinyl and divinyl forms of chlorophylla (MVChl a , DVChl a ) which are used in the calculation of total chlorophylla (TChl a ), monovinyl chlorophyllb (MVChl b ), and the accessory pigments β - carotene and alloxanthin: MV Chlacorr = 0.598 ·MV Chlauncorr + 0.0073 DV Chlacorr = 0.655 ·DV Chlauncorr + 0.0003 MV Chlbcorr = 0.5682 ·MV Chlbuncorr + 0.001 β−carotenecorr = 0.7·β−caroteneuncorr alloxanthincorr = 0.8·alloxanthinuncorr where the superscripts corr and uncorr refer to corrected and uncorrected data, respectively, and concentrations are in units of mg m−3 . Other pigments did not show consistent relationships in the intercalibration results, and no corrections were attempted for them. It is cautioned that the corrections should not be used indiscriminately with other data sets aected by the CHORS data quality problems. The values of TChl a calculated using the corrected concentrations of MVChl a and DVChl a exhibit reasonable agreement with uorometrically-derived chlorophyll measurements, and provide realistic estimates of chlorophyll-specic phytoplankton absorption coecients within the red peak of chlorophylla . In a comparison of the 129
Appendix B corrected CHORS results with independent pigment determinations done by GKSS, some dierences were observed in the concentrations of individual pigments between the two data sets. For example, the sum of MVChl a and DVChl a was generally higher (on average by 20% with a standard deviation of 23%, number of samples 25) for CHORS compared with GKSS. However, the CHORS data yielded more reasonable estimates of chlorophyll-specic phytoplankton absorption within the red peak of chlorophylla . Despite such dierences in the estimates of some individual pigment concentrations, with regard to the present application it is important to note that both laboratories provided similar characterization of samples in terms of the relative pigment composition as described by ratios of various individual pigments to TChl a . Both sets of HPLC analyses generally yielded the same dominant accessory pigments present at any given station, and similar trends in the pigment ratios among the stations. This is an essential result for the present study because in the cluster analysis only pigment ratios were utilized, and not individual pigment concentrations. A series of analyses were performed to test for any dierences in the results when either the CHORS or GKSS pigment data sets were used to generate the reference pigment dendrogram. For the nine stations selected in this part of the analysis, the cluster techniques described in Section 3.2 applied independently to the CHORS and GKSS sets of HPLC pigment ratios yielded a very similar partitioning of stations into clusters. This high degree of similarity between the two pigment-derived dendrograms exhibits a cophenetic index value of 0.92, indicating that overall pairwise distances between the data objects are well-preserved. This observation is further supported by additional cluster analyses comparing both pigment data sets with optical data such as the normalized phytoplankton spectral absorption coecient (see Fig. B.1). Figure B.1: Scatter plot of the cophenetic indices between the second derivative spectra of a∗ n,ph(λ) -based and the pigment-based cluster trees. The xand y-axis indicate the indices obtained using the CHORS and GKSS HPLC analysis, respectively. 130
Comparisons of these absorption-based clusters with pigment-based clusters were found to be similar when either the CHORS or GKSS pigment data were used. In particular, Figure B.1 depicts how the values of the cophenetic index obtained from the analysis of similarity between the cluster trees of pigments and the second derivative spectra of normalized phytoplankton absorption are closely correlated between the two pigment data sets. The correlation coecient is 0.97. The results include the cophenetic indices evaluated for all combinations of spectral ranges of the second derivative spectra of a∗ n,ph(λ) as shown in Fig. 3.12. In summary, the comparisons of two independent HPLC pigment data sets indicate that the results and conclusions of this study are robust and independent of the choice of pigment data used in the creation of the reference pigment dendrogram for the cluster-based similarity analysis of pigment and optical data. Therefore, it was chosen the one set of pigment results from CHORS for all subsequent analyses in the study. 131