The use of remotely sensed data for forest biomass monitoring : a case of forest sites in north-eastern Armenia

i THE USE OF REMOTELY SENSED DATA FOR FOREST BIOMASS MONITORING A case o o es si es in No h-Eas e n A menia Man el Khudinyan ii THE USE OF REMOTELY SENSED DATA FOR FOREST BIOMASS MONITORING A case o o es si es in No h-Eas e n A menia Disse a ion supe ised by: Joel Sil a, PhD. In o ma ion Managemen School, No a Uni e si y o Lisbon, Lisbon, Po ugal Goha Ghaza yan, MSc. Cen e o Remo e Sensing o Land Su aces (ZFL), Uni e si y o Bonn, Bonn, Ge many Ignacio Gue e o, PhD. Uni e si a Jaume I, Cas ellón de la Plana, Spain Feb ua y, 2019 iii ACKNOWLEDGMENTS I am especially indeb ed o my supe iso PhD Joel Sil a o he g ea suppo and echnical guidance he has o e ed. I would also like o pa icula ly hank my co- supe iso PhD Goha Ghaza yan o he con inuous suppo and cons uc i e commen s. I am also e y hank ul o co-supe iso P o . Ignacio o he ime he has de o ed o my hesis. This wo k would no ha e been possible wi hou P o . D . Ma co Painho, P o . B ox and e e y membe o s a a UNL and i gi who I ha e o hank o my wonde ul s udy expe ience. I would especially like o hank P o . Ho ik Sayadyan who p o ided he ield da a (no da a – no hesis), o suppo ing my ideas and o he encou agemen . I am also g a e ul o PhD Go ik A e isyan and A man Kanda yan o he da a hey p o ided, as well as my iends, o es e s PhD Vahe Ma i osyan and PhD Vahe Ma sakyan o he consul a ions on he o es s and hei help. Las bu no leas I would like o hank my amily, my iends and my classma es o hei cons an suppo and ca e. i THE USE OF REMOTELY SENSED DATA FOR FOREST BIOMASS MONITORING A case o o es si es in No h-Eas e n A menia ABSTRACT In ecen yea s he e has been an inc easing in e es in he use o syn he ic ape u e ada (SAR) da a and geospa ial echnologies o en i onmen al moni o ing․ Pa icula ly, o es biomass e alua ion was o high impo ance, as o es s ha e a c ucial ole in global ca bon emission. Wi hin his s udy we e alua e he use o Sen inel 1 C-band mul i empo al SAR da a wi h combina ion o Alos Palsa L-band SAR and Sen inel 2 mul ispec al emo e sensing (RS) da a o mapping o es abo eg ound biomass (AGB) o d y sub opical o es s in moun ainous a eas. Field obse a ion om Na ional Fo es In en o y was used as a g ound u h da a. As he SAR da a su e s g ea ly by he complex opog aphy, a simple app oach o aspec and slope in o ma ion as o es y ancilla y da a was implemen ed di ec ly in he eg ession model o he i s ime o mi iga e he opog aphy e ec on ada backsca e ing alue․ Dense ime-se ies analysis allowed us o o e come he SAR sa u a ion by he o es phenology and selec he op imal C-band scene. Image ex u e measu es o SAR da a has been s ongly ela ed o he biomass dis ibu ion and has obus ly con ibu ed o he p edic ion․ Mul ilinea S epwise Reg ession allowed o selec and e alua e he mos ele an a iables o AGB. The p edic ion model combining RS wi h ancilla y da a explained he 62 % o a iance wi h oo -mean-squa e e o o 56.6 ha¯¹. The s udy also e eals ha C-band SAR da a on o es biomass p edic ion is limi ed due o hei sho wa eleng h. Fu he , he moun ainous condi ion is a majo cons ain o AGB es ima ion. Addi ionally, his esea ch demons a es a posi i e ou come in o es AGB p edic ion wi h eely accessible RS da a. KEYWORDS Syn he ic Ape u e Rada Fo es Abo eg ound Fo es Biomass Mul ilinea S epwise eg ession Sen inel 1 Sen inel 2 ALOS PALSAR GLCM i ACRONYMS AGB – Abo eg ound Biomass SAR – Syn he ic Ape u e Rada REDD+ – Reducing Emission om De o es a ion and o es Deg ada ion UN – Eu opean Union NDVI – No malized Di e ence Vege a ion Index NDMI – No malized Di e ence Mois u e Index RVI – Simple Ra io Vege a ion Index FAO – Food and Ag icul u e O ganiza ion o he Uni ed Na ions NFI – Na ional Fo es In en o y ESA – Eu opean Space Agency SRTM - Shu le Rada Topog aphy Mission SWR – S ep Wise Reg ession JAXA – Japanese Ae ospace Explo a ion Agency DBHOB – Diame e on B es Heigh O e Ba k GEE – Google Ea h Engie UNDP – Uni ed Na ion De elopmen P og am GEF – Global En i onmen al Facili y ii INDEX OF THE TEXT Page ACKNOWLEDGMENTS .......................................................................................... iii ABSTRACT ................................................................................................................. i KEYWORDS ................................................................................................................ ACRONYMS ............................................................................................................... i INDEX OF THE TEXT ............................................................................................. ii INDEX OF TABLES .................................................................................................. ix INDEX OF FIGURES ................................................................................................. x 1.1 Theo e ical F amewo k ........................................................................................... 1 1.2 S a emen o he P oblem ........................................................................................ 3 1.3 Aims and Objec i es ............................................................................................... 5 1.4 Ou line ..................................................................................................................... 6 LITERATURE REVIEW ........................................................................................... 7 2.1 In oduc ion o Abo eg ound Fo es Biomass ...................................................... 7 2.2 SAR and Op ical emo e sensing da a o AGB es ima ion .................................. 12 2.2.1 Combina ion o Mul isou ce Da a o AGB Moni o ing ................................... 14 2.2.2 SAR D awbacks ............................................................................................................ 15 2.2.3 SAR Dense Time-Se ies Analysis ........................................................................... 17 2.3 Fo es AGB p edic ion models: Model diagnos ics .............................................. 18 DATA AND STUDY AREA ..................................................................................... 21 3.1 In oduc ion o he S udy A ea .............................................................................. 21 3.2 In oduc ion o Da a .............................................................................................. 23 3.2.1 Field Da a ........................................................................................................................ 23 3.2.2 Sen inel 1 Da a ............................................................................................................... 24 3.2.3 Sen inel 2 Da a ............................................................................................................... 25 3.2.4 Alos Palsa Da a ............................................................................................................ 26 3.2.5 Fo es y Ancilla y Da a ............................................................................................... 27 APPROACHES, METHODOLOGY, DATA PREPARATION ........................... 28 4.1 App oach and Gene al Me hodology .................................................................... 28 iii 4.2 Tools ..................................................................................................................... 31 4.3 Field Da a Calcula ion .......................................................................................... 32 4.4 Sen inel 1 P e-p ocessing ...................................................................................... 34 4.5 Palsa P e-p ocessing ............................................................................................ 36 4.6 Sen inel 2 P e-p ocessing ...................................................................................... 36 4.7 Ancilla y Da a P e-p ocessing .............................................................................. 37 4.8 GLCM ex u e analysis ......................................................................................... 38 DATA PROCESSING․ FOREST BIOMASS MONITORING ............................. 41 5.1 Fo es /Non-Fo es classi ica ion ........................................................................... 41 5.2 Resul s o Reg ession analysis .............................................................................. 44 5.2.1 Model Compa ison and P edic o E alua ion....................................................... 46 5.2.2 Model Compa ison on Si e Le el ............................................................................ 49 5.3 Discussion ............................................................................................................. 52 5.4 Limi a ions and Recommenda ions o Fu u e Resea ch ....................................... 56 CONCLUSION .......................................................................................................... 58 BIBLIOGRAPHIC REFERENCES ......................................................................... 59 ix INDEX OF TABLES Table 1. Bene i s and limi a ions o a ailable me hods o es ima e na ional-le el o es ca bon s ocks (da a sou ce: Gibbs e al., 2007) ....................................................... 8 Table 2. Common ada bands used in Remo e Sensing (da a sou ce: Pohl. 2017) .... 12 Table 3. Field sampling da a pe s a a ........................................................................ 23 Table 4. Tools and ela ed p ocessing .......................................................................... 31 Table 5. AGB allome ic equa ions o dominan ee species .................................... 32 Table 6. Fo mulas o GLCM ex u e measu es .......................................................... 40 Table 7. Valida ion con usion ma ix o o es non- o es classi ica ion ................... 42 Table 8. Inpu a iables o eg ession analysis ........................................................... 44 Table 9. E alua ion o s a is ics o eg ession analysis o each model ...................... 47 6 1.4 Ou line The s uc u e o he disse a ion is as ollows: Chap e 2 p esen s he e iew o exis ing li e a u e on he opic o his esea ch. Chap e 3 desc ibes he condi ions and speci ica ion o he s udy a ea and he da a used wi hin his s udy. Chap e 4 discusses he app oaches and me hodology, also desc ibes he da a p e-p ocessing s eps in e y de ails. Chap e 5 p esen s and discusses he esul s o o es /non- o es classi ica ion and o eg ession analysis, also ema ks he limi a ions and u u e possible esea ches. Sec ion Conclusion sums up he esul s o his esea ch and e alua es i s con ibu ions. 7 LITERATURE REVIEW 2.1 In oduc ion o Abo eg ound Fo es Biomass Be o e s a ing he analysis o emo ely sensed da a o o es abo eg ound biomass es ima ion, i is essen ial o unde s and he concep o o es AGB, he adi ional echniques and me hods o es ima ion, which gi es be e unde s anding and ision o he pa ame e s and condi ions should be conside ed be o e applying emo e sensing me hods. Abo eg ound li e biomass includes he li ing biomass o ees, sh ubs and he bs abo e he soil including s em, s ump, b anches, ba k, lea es (UNDP-GEF 00091048, 2015), and ep esen s he la ges pool o ca bon s ock (Gibbs e al., 2007). AGB is widely used o co ela ions and es ima ions o ca bon s o age in some o he o he pools such as oo biomass and consequen ly he ca bon s ocks (app oxima ely 20 % o AGB (Gibbs e al., 2007)) and dead wood o li e ca bon pool (app . 10-20 % in na u al o es s (UNDP-GEF 00091048, 2015)). Thus, he o es AGB es ima ion is an essen ial s ep in quan i ying he ca bon s ocks in he o es s. The e a e se e al me hods de eloped o AGB es ima ion wi h di e en demands and le el o accu acy. The mos accu a e and s aigh o wa d me hod o quan i y he o es biomass is es ablishing sampling si es on he ield and ha es ing all he ees, d ying hem (in pa icula , o ca bon s ock es ima ion) and weigh ing he biomass (Gibbs e al., 2007). This me hod is e y expensi e and has des uc i e a ec ion on he expe imen al si e. I is p ecise o ha one loca ion only, and imp ac ical in o he egions o coun y le el analysis (Chen e al., 2018). Such me hods canno be applied ac oss he landscape. Consequen ly, many in es men s we e pu in o de elopmen o models ha a e able o “scale up” he ield measu emen esul s o e la ge a eas. Gibbs did sum up in glance he main me hods used o o es AGB es ima ion wi h hei ad an ages and limi a ions (Table 1) (Gibbs e al., 2007). 8 Me hod Bene i s Limi a ions Unce ain y Biome a e ages • Immedia ely a ailable could inc ease accu acy • Da a e inemen s could inc ease accu acy • Globally consis en • Fai ly gene alized • Da a sou ces no p ope ly sampled o desc ibe la ge a eas High Fo es in en o y • Gene ic ela ionships eadily a ailable • Low- ech me hod widely unde s ood • Can be ela i ely expensi e as ield- labo is la ges cos • Gene ic ela ionships widely unde s ood • Low- ech me hod inexpensi e as • Can be ela i ely ield-labo is la ges cos Low Op ical emo e senso s • Sa elli e da a ou inely a ailable a global scale • Globally consis en • Limi ed abili y o de elop good models o opical • Spec al indices sa u a e a ela i ely low ca bon s ocks • Can be echnically demanding High Rada emo e senso s • Sa elli e da a a e no always ee •New sys ems launched a e expec ed o p o ide imp o ed da a • accu a e o young o spa se o es • Less accu a e in ma u e o es s because signal sa u a es also inc eases e o s • Moun ainous e ain also inc eases e o s • Can be expensi e Medium Table 1. Bene i s and limi a ions o a ailable me hods o es ima e na ional-le el o es ca bon s ocks (da a sou ce: Gibbs e al., 2007) Bou e and Gibbs a e di iding biome-a e age me hod in o wo main s eps (Bou e e al., 2018; Gibbs e al., 2007). Fi s , da a ga he ing is being done h ough NFIs. Un il now he bes esul s a e gained by ield calcula ions using expe imen al plo s, ha es ing and ac ual o es olume es ima ion (Kachamba e al., 2016). In second s ep allome ic models a e applied o calcula e a e age biomass o a sampling plo loca ed wi hin a ce ain s a um. La e , na ional le el o ca bon s ock can be p edic ed by applying he a e age biomass and ca bon densi y alues (de Bad s, 2002) o e he egion o he same o es s a a (Bou e e al., 2018; Kachamba e al., 2016). Unlike he high le el o unce ain y compa ed o he o es ha es ing me hod, his me hod can be immedia ely a ailable and p o ide in o ma ion abou AGB on a wide scale landscape. G ound-based o es in en o y ocused on ield campaigns and o es in en o y measu emen s o es ima ion o o es AGB. Measu emen s include diame e a b eas heigh (DBH), a e age ee heigh and ee ype, and h ough which o es biomass is being calcula ed using allome ic ela ionships (Cha e e al., 2005; Gibbs e al., 2007). Only DBH is desc ibing AGB wi h 95 % accu acy (Cha e e al., 2005). Zhang o e s o collec da a abou he ege a ion ype and soil ype as well (Zhang e al., 2018). As he AGB consis o s em, ba k, b anches and lea es o needles depending on he o es 9 ype, he assumes o o ganize ield da a acquisi ion di e en ly o coni e ous and deciduous o es s and exclude lea e biomass o win e calcula ions in deciduous o es s (Cle ici e al., 2016). Fo ield da a collec ion he e a e se e al app oaches o sampling poin s selec ion. Sys ema ic and andom sampling designs a e he mos used ones. Sys ema ic sampling uses? egula g id o plo loca ion selec ion, and he andom selec ion is applied o andomly alloca ed sampling poin s. Bo h o he app oaches a e no conside ing o es ype s a i ica ion and o es dis ibu ion. This, bo h schemes can unde - o o e -sample as a as he pa e ns in na u e ha e na u ally clumpy and andom dis ibu ion is less likely (Gibbs e al., 2007). One o he unce ain ies o AGB es ima ion by con e ing ee measu emen s da a is he lack o s anda d me hod. Li e a u e e iew shows di e en app oaches by di e en au ho s o AGB es ima ion e en o he same ype o o es s (Cha e e al., 2005; Cle ici e al., 2016; Gibbs e al., 2007). Gibbs p esen s a comple e o e iew o o es AGB es ima ions by di e en au ho s o he same coun ies, and as he a icle poin s ou , o Sh i-Lanka he AGB es ima ion by di e en me hods can al e in he ange o 138–509 Mg/m2. The unce ain y occu s when shi ing om one o es ype o ano he . In opical o es s 1 ha o es a ea can con ain 300 di e en ee species. This means ha one canno use species-speci ic eg ession model, which is common o apply while wo king wi h empe a e o bo eal o es s. In such condi ions be e esul s a e achie ed by mixed-species eg ession model (Cha e e al., 2005). Amongs exis ing numbe o s a is ical me hods o AGB acquisi ion based on sampled allome ic measu emen s, Cha es sugges s “biome-diame e -heigh eg ession” and “biome-diame e eg ession” models o hei simplici y and wide applica ion (Cha e e al., 2005). Biome-diame e -heigh eg ession includes in o ma ion abou ee heigh (H), diame e (D), and wood speci ic g a i y (ρ) (Dawakin’s eg ession model): ln(𝐴𝐺𝐵) = α + ln(D²Hρ) (1) Bo h o he eg ession and he es a e applied using linea models. Field sampling da a does no always include he ee heigh in o ma ion. In his case Cha es assumes ela ionship be ween loga i hm o heigh , In(H), and he loga i hm o diame e , In(D). A polynomial ela ionship be ween log. heigh and log. diame e 10 gi es a easonable gene aliza ion o he powe -low model. Hence, he model will ha e ollowing equa ion: ln(𝐴𝐺𝐵) = a + b*ln(D²Hρ) + c*ln(D²) + d*ln(D³)+ β*3ln(ρ) (2) The powe -low is pa ame ized as c=d=0 (Cha e e al., 2005). All hose con en ional models a e accu a e o he loca ion o sampling si e bu a e di icul o ex end o e la ge a ea and a e being done no mo e han once o ce ain a ea because o pa amoun esou ce demand. Ins ead, emo ely sensed da a p o ides use ul o AGB es ima ion da a o he whole globe and wi h empo al esolu ion, bu canno measu e he biomass di ec ly, hus equi e g ound- u h da a (Gibbs e al., 2007; Kuma e al., 2015). Ano he impo an ad an age o emo e sensing is he abili y o easily collec da a on he places whe e g ound access is limi ed. Op ical emo e sensing da a om di e en senso s (see chap e 1.1) ha e been implemen ed ( ied) o o es AGB indi ec es ima ion ia s a is ical ela ionship be ween ield measu ed da a and sa elli e-obse ed ege a ion indices (Fassnach e al., 2014). As he op ical senso eco ds he in o ma ion e lec ed om he o es canopy and is dependen on he lea s uc u e, pigmen a ion and mois u e and ha e week pene a ion possibili ies, hus he unce ain y is high, especially, in dense opical o es s wi h big amoun o biomass (Chen e al., 2018; Joshi e al., 2016; T euha e al., 2017). In he con a y, compa a i ely low unce ain y and highe co ela ion be ween senso -da a and ield-da a a e eco ded in bo eal o es s whe e he hie a chical s uc u e is absen in he o es s and indi idual ees can be ixed on he pho os (Gibbs e al., 2007; Kachamba e al., 2016). Cle ici a i ms ha e y high- esolu ion (VHR) op ical image y imp o es he AGB es ima ions in o es s (Cle ici e al., 2016) using a io ege a ion index (RVI), no malized ege a ion index (NDVI), ans o med ege a ion index (TVI), while he Vege a ion Index Numbe (VIN) eco ded he wo s pe o ming. Kuma o e s a use o hype spec al senso da a, pa icula ly unde lying he impo ance o mid-in a ed (MIR) e lec ance as he bes desc ibe o o es biomass. He s a es ha MIR has ad an ages o e isible and nea - in a ed (NIR) e lec ance (L. Kuma e al., 2015). Signal sa u a ion in he o es s is he main cons ain o op ical emo e sensing o AGB es ima ion and almos all he e iewed a icles de ec ed o be su e ing wi h his issue. 11 Wi h he launch o di e en SAR senso s wi h eely dis ibu ed da a, he e is inc easing in e es in he use o hese obse a ions o o es AGB es ima ion. This in e es s a ed in 2000s when Te aSAR-X, Alos, Palsa L-band we e launched, and become mo e popula , when in 2014 i s Sen inel 1 C-band da a became eely a ailable. Fo ins ance, he majo i y o he s udies o SAR in AGB es ima ions we e published in he las 3-5 yea s (Be engue e al., 2018; Be ninge e al., 2018; Cle ici e al., 2016; Joshi e al., 2016; Kuma e al., 2015; Na a o e al., 2016; Reiche e al., 2018; San i e al., 2017; San o o & Ca us, 2018; Va aei e al., 2018; Yu & Saa chi, 2016). Gene ally, he mos egula ly used me hods o o es AGB es ima ion using SAR da a can be classi ied in o se e al g oups. The simples and subsequen ly he mos used me hod is he measu ing backsca e ing coe icien om Pola ime ic SAR (PolSAR) o e eal he oughness o he su ace. This is called 2D PolSAR me hod (Zhang e al., 2018). As i was men ioned be o e, he bes esul s a e ob ained wi h c oss-pola iza ion dual pola iza ion which is mo e sensi i e o he o es AGB. Mos o hem applied loga i hm o biomass and used he backsca e ing coe icien o o es biomass p edic ion (Lau in e al., 2018; Zhang e al., 2018). This me hod s ill has some limi a ions such as sa u a ion, which depends on he wa eleng h and he incidence angel (Joshi e al., 2016). The capabili ies, ad an ages and d awbacks o SAR backsca e da a in o es AGB es ima ion, will be discussed in de ails in he nex subchap e s In e e ome ic SAR (InSAR) o Pola ime ic InSAR (PolInSAR) by using in e e ome ic phase o cohe ence omog aphy o InSAR a e able o eco d he ele a ion o he g ound and he op o he o es , and hus, hey can de i e he o es heigh which hen can be ans o med in o o es biomass by allome ic equa ion models. This model is p omising and has mo e po en ial, and i can educe he sa u a ion in some ex an (Chen e al., 2018). A numbe o s udies show imp o ed AGB es ima ion esul s by using InSAR da a (Huang, Zini i, To bick, & Ducey, 2018; San o o & Ca us, 2018; Zhang e al., 2018). The e a e also o he me hods o AGB es ima ion h ough Lase scanning and LiDAR scanning (Joshi e al., 2016), which a e no i ing in he scopes o his esea ch, he e o e hey will no be discussed in de ails. 12 2.2 SAR and Op ical emo e sensing da a o AGB es ima ion SAR a e ac i e mic owa e ada s ope a ing in be ween 1 cen ime e and se e al me e s o elec omagne ic spec um. A pa o he impulse ene gy om ada , once mee ing he Ea h’s su ace, i is being sca e ed back owa ds he senso . This is wha he Rada eco ds and measu es. Depending on he su ace ype and s uc u e, di e en amoun o ene gy is sca e ed back and hus, di e en pa ame e s o he su ace a e eco ded on he image. The image in ensi ies depend on he backsca e ed signal cha ac e is ics such as wa eleng h, incidence angle, signal pola iza ion, scan di ec ion, in addi ion o such pa ame e s as su ace oughness, mois u e, dielec ic p ope ies, geome ic shape, o ien a ion (Pohl C., 2017). The equency, incidence angle, pola iza ion and scan di ec ion a e p ope ies o sys em and he dielec ic cha ac e is ics, o ien a ion, su ace oughness and he mois u e a e a ge p ope ies (Pe iasamy e al., 2018). Those a e he pa ame e s ha a e in luencing on he backsca e ed signal and hus, desc ibing he su ace objec s, hey a e co e o ada image p ocessing. SAR images o he same a ea a y as he sys em pa ame e s change. Fo ins ance, he pene a ion capabili y o he signal is dependen on he wa eleng h: The longe wa eleng h he la ge objec s hey pene a e. Acco ding o he ule o humb he pene a ion leng h is he hal o he wa eleng h. Acco ding o he same ule, highe he backsca e in ensi y he oughe he su ace ha is being imaged (CRISP 2001, n.d.). Rada Band Wa eleng h (cm) F equency (GHz) X 2.5–3.75 8–12 C 3.75–7.5 4–8 S 11.11–7.69 2.7–3.9 L 15–30 1–2 P 100 0.3 Table 2. Common ada bands used in Remo e Sensing (da a sou ce: Pohl. 2017) Depending on he used wa eleng h, SAR senso s a e di ided in o se e al bands (namely: K, X, S, C, L, P, VHF) (Timo hy e al., 2016). Table 2 desc ibes he wa eleng h and app op ia e equencies o he SAR bands mos commonly used o land co e su eys. L and P- band (15-100 cm oge he ) ha e p i ileges o e C-band (3.75-7.5 cm) in e ms o o es e ical alues desc ip ion. C-band can pene a e only lea es and 13 needles and small b anches, while L and P-bands p o ide s onge backsca e o m unk and la ge b anches (Figu e 1) (Joshi e al., 2016). This leads o a di e en esul when calcula ing o es AGB ough abo e-men ioned SAR bands. Acco ding o he e iewed li e a u e L-band analysis eco ds highe co ela ion wi h g ound da a compa ed wi h he C-band (Bou e e al., 2018; Lau in e al., 2016, 2018; Me moz e al., 2015; Yu & Saa chi, 2016). Figu e 1. Vege a ion pene a ion capabili ies o X, C and L ada bands (da a sou ce: Pohl. 2017) Pola ime ic SAR senso s a e capable o ansmi and de ec he e ical (V) and ho izon al (H) componen s o he backsca e ed adia ion (Pohl C., 2017). Hence, he e a e ou possible pola iza ion con igu a ions: co-pola iza ion- HH, VV, c oss- pola iza ion- HV, VH, and hei combina ions (Aula d-Macle M., R. Ba s ow, 2011). Pola ime ic SAR backsca e ed ene gy is di ec ly depending on he physical p ope ies o he ege a ion elemen s and in luences on backsca e ing mechanism (Pohl C., 2017). The ough su ace signi ican ly depola izes he signal in a g ea e magni ude while smoo he su ace depola izes he signal a he lowe magni ude. Thus, deg ee o pola iza ion can p o ide aluable s uc u al in o ma ion abou he o es s (Pe iasamy, 2018). Lau in and San o o in hei a icles a i m ha c oss- pola iza ion has p io i ies o e co-pola ized backsca e in o es biomass in o ma ion e ie al (Lau in e al., 2018; San o o & Ca us, 2018). A numbe o s udies poin ou he VH backsca e om Sen inel 1 C-band and HV backsca e o m Alos Palsa and Alos2 Palsa 2 L-band SAR o ha e mo e eliable co ela ion wi h AGB da a o e he co-pola iza ion and o he pola iza ion con igu a ions (Huang e al., 2018; Lau in e al., 2016) Li e a u e e iew shows se e al a emp s o image ex u e cha ac e is ics implemen a ion o biomass es ima ion using op ical image y back in he pas (Ecke , 14 2012). In he mo e esen li e a u e e iew by San o o & Ca us (2018) he in es iga ion o ex u e cha ac e is ics and pola iza ion decomposi ions o SAR image y a e b inging smalle e ie al e o s in o es AGB es ima ion compa ed o only backsca e alues. Pape s published by Be ninge , Thapa and Huang depic imp o ed co ela ions be ween image ex u e (namely, homogenei y, con as , en opy, olume sca e ing) and he ield da a om allome ic calcula ions (Be ninge e al., 2018; Huang e al., 2018; Thapa e al., 2016). In case o SAR da a, he ex u e is a measu e o he spa ial homogenei y o he backsca e ing and desc ibes he p ope ies o su ace, such as smoo hness, egula i y and onal a ia ion, and so, should enclose in o ma ion abou o es s uc u e (San o o & Ca us, 2018). Chen e al. also show ha Sen inel 1 image ex u es a e he mos ela i e and impo an p edic o s o AGB es ima ion compa ed o he o iginal backsca e ing da a (Chen e al., 2018). This e iew shows ha he implemen a ion o ex u e me hods o o es biomass e ie al a e unde s udied ye and seem o ha e high po en ial in imp o ing he es ima ion accu acy. 2.2.1 Combina ion o Mul isou ce Da a o AGB Moni o ing Combina ion o SAR and op ical da a is being in ensi ely implemen ed mos ly in he las couple o yea s and shows high po en ial o imp o e he o es biomass p edic ion accu acy (Joshi e al., 2016; Kuma e al, 2016; Lau in e al., 2018). Joshi e al 2016, in his e iew o exis ing li e a u e inds ha di e en SAR da a wi h conjunc ion o op ical da a ha e compa a i ely highe accu acy han SAR and op ical da a alone (Joshi e al., 2016). Besides he p edic ion accu acy imp o emen , combina ion o globally a ailable op ical senso s such as Landsa 8, Sen inel-2, and SAR senso s like Palsa -2 and Sen inel-1 wi h high empo al esolu ion, is becoming a e y impo an ool in a way ha ensu es ope a ional and con inuous global o es co e moni o ing in consis en and obus manne (Reiche e al., 2016). Joshi e al. (2016) and Reiche e al. (2106) assume ha ope a ing on di e en physical p inciples, hence, he SAR and op ical senso s p o ide syne gis ic in o ma ion on he o es p ope ies. As we al eady saw, he i s is dependen on he size, o ien a ion and dielec ic p ope ies, densi y, while he second is dependen on he lea s uc u e, mois u e and pigmen a ion. This le s hem conside ha combina ion o hose di e en sa elli e image y should be 15 p omising app oach o inc easing he accu acy o AGB es ima ion in o es s (Joshi e al., 2016; Reiche e al., 2016). The e is g owing ocus on his app oach and he majo i y o he esea ches e iewed a e om 2018 (Chen e al., 2018; Kuma e al., 2016; Lau in e al., 2018; Pe iasamy, 2018; Va aei e al., 2018). The mos o en used op ical senso is Sen inel-2 because o he high esolu ion, ee a ailabili y and he global co e wi h high equency e isi cycle. Lau in e al. in hei pape (2018) showed ha combina ion o Sen inel-1 and Sen inel-2 s ongly imp o ed he o es AGB es ima ion in Medi e anean sh ublands (a ound 14% wi h espec o he senso s sepa a ely). As an inpu o he p edic ion model all he bands o m Sen inel 2 as di e en a iables we e used. Chen e al. (2018) gene a ed numbe o ege a ion indices and biophysical a iables as inpu a iables o he p edic ion models. The esul s show ha he ege a ion biophysical a iables a e ou s anding compa ed o he o he Sen inel-2 p oduc s while combining wi h Sen inel-1 p oduc s (Chen e al., 2018). Lau in e al (2018) and Chen e al (2018) highligh ha combina ion o mul i-sou ce sa elli e image y imp o es he sa u a ion le el o o es AGB es ima ion. Lau in e al in hei esea ch (2018) go up o 400 ones/ha accu a e es ima ion combining Sen inel 1,2 and Palsa 2 sa elli e image y. Zhang assumes ha in e e ome ic phase and cohe ence me hods ha e po en ial o o e come he sa u a ion in some ex ends (Zhang e al., 2018). 2.2.2 SAR D awbacks The e a e numbe o challenges while analyzing and in e p e ing SAR images o land applica ions and pa icula ly o o es y. Th ee main d awbacks o SAR da a can be sepa a ed ha se e ely a ec he measu emen accu acy, namely, speckle noise, he bias on he backsca e alue because o he moun ain elie and sa u a ion o elec omagne ic signal. Speckle noises: The majo p oblems a e he speckle noises on SAR images ha migh cause o poo classi ica ion (Joshi e al., 2016). Unlike he op ical emo e sensing, ada scanning is cohe en in e ac ion o he signal wi h he su ace objec s. As he esul o cohe en summa ion o he signal sca e ed o m, he g ound sca e e s has andom dis ibu ion wi hin each pixel and is called speckle noise (CRISP 2001, n.d.). 22 co e ed (UNDP-GEF, 2015). Annual p ecipi a ion is 500-540 mm and he clima e is desc ibed wi h wa m and d y summe s and empe a e win e s (FMP, 2018). Figu e 2. Loca ion o he s udy a ea. a) Loca ion o he Ta ush p o ince, b) o es en e p ises in Ta ush p o ince, c) "Noyembe yan" o es en e p ise This esea ch ocuses on he o es si es loca ed in Ta ush p o ince (A s abe d, Ije an, Se qa , Noyembe yan and “Dilijan” Na ional Pa k), The main ocus is Noyembe yan o es en e p ise o se e al easons: • The pilo p ojec o na ional o es ca bon in en o y has been implemen ed in Noyembe yan o es en e p ise and can be as a g ound base o e alua ion o esul s, • The ield sampling densi y is much highe in Noyembe yan o es en e p ise (55 sampling poin s ou o 115), • Some o emo e sensing da a has huge dis o ions o e he o he o es si es which p ac ically makes hose da a no use ul o o es moni o ing on he whole a ea, • All he o es s in he NE A menia ha e e y simila cha ac e is ics in e ms o o es ype, s uc u e, amoun o biomass, e ain condi ions, and so, we do no pu 23 unce ain y while using he ield sampling da a om one si e o model aining and applying on he o he one. Fo s a is ical analysis and model aining in o ma ion was collec ed om all he 5 o es sigh s bu he p ojec ’s main ocus a ea is conside ed he o es s o Noyembe yan. The e o e, all he objec i es and hypo hesis a e applied and he inal maps and e alua ion a e done o Noyembe yan o es s only. The o es s in Noyembe yan en e p ise is 29334 ha and is loca ed in be ween 600- 1850 MASL al i udes and he dis ibu ion is as ollowing: <800m – 4.2%, 801-1200- 47.9%, 1201-1600-37.8%, 1601> - 10.1%. Fo es s uc u e, quali y and dominan species a y based on he al i ude - sp ead low densi y in he low al i udes o high densi y in he middle. The opog aphy is highly agmen ed and ema kable wi h s eep slopes and aspec composi ion a ia ions. This s ongly a ec s he o es dis ibu ion. The e a e mo e o es s loca ed on he no h slopes (58.6%) han on he sou h slopes (41.4%) (FMP, 2018). 3.2 In oduc ion o Da a 3.2.1 Field Da a Field da a a e gained om he UNDP-GEF ongoing p ojec . The ini ial numbe o plo s we e 115, ou o which 55 we e om Noyembe yan en e p ise collec ed in 2017, and he es a e om he o he o es si es o Ta ush p o ince collec ed in he beginning o 2018. The ield sampling plo s’ loca ions a e chosen based on a sys ema ic sampling plo design (Figu e 3). S a a No o samples Beech 18 Ho nbeam 22 Oak 18 Pine 12 O he 19 Dis u bed 26 To al 115 Table 3. Field sampling da a pe s a a 24 Figu e 3. Field sampling plo s dis ibu ion The e was also used a o es s a i ica ion map o ensu e ha he sampling plo s include all he o es ypes (Table 3). Fixed size o 0,1 ha ci cula plo s we e es ablished wi h 17.84 m adius (depending on he slope deg ee i migh change in o de o ensu e a 0,1 ha plo size when p ojec ed on a plane). All he la ge ees (>8 cm in diame e ) a e assessed o he ollowing p ope ies: species, DBHOB, ee s a us, decay class, c own class and many o he pa ame e s, ha a e no ele an o his s udy, hus will no be epo ed. 3.2.2 Sen inel 1 Da a Fo his s udy Sen inel 1A and Sen inel 1B C-band SAR da a om Cope nicus p ojec o ESA was used. The da a is in e e ome ic wide swa h (IW) scanning mode, wi h 25 250 km swa h wid h. We used Le el-1 G ound Range De ec ed (GRD) p oduc which is al eady Mul i-looked (one look in ange and i e in azimu h), Geocoded and i is a ailable wi h 10x10m pixel spacing bo h in Cope nicus Open Access Hub¹, ( he online sys em o he ESA), and in Google Ea h Engine Ja aSc ip API² (GEE). Bo h sa elli es o bi in nea pola , sun-synch onized o bi a 693 km al i ude and in he same o bi al plane (To es e al., 2017). Sen inel 1 C-band SAR has 5.405 GHz equency (co esponding o a wa eleng h o ~5.6 cm) and p o ides images in wo pola iza ion modes: VV co-pola iza ion and VH c oss-pola iza ion. The incidence angle is be ween ~31 and ~46 deg ees. The esolu ion was se o 10 m. Each image con ains 3 bands: wo o backsca e ing in ensi y (VV, VH) and one o incidence angle. Because o he e ogenei y in he pa ame e s o a ailable images he e is a need o il e o down he da a o a homogeneous subse . In o de o ha e he same incidence angle o all he images he same ela i e o bi numbe 152 was selec ed. The e we e 72 scenes acqui ed o he whole yea o 2017 (30 om Sen inel 1A, 42 om Sen inel 1B), 6 images pe mon h in a e age. Fo Sen inel 1 da a a ailable in GEE pla o m he backsca e ing coe icien om na u al alues o sigma naugh (σ◦) is con e ed in o dB alues. σ◦ = 10⁎log10 𝐷𝑁 (3) Whe e he DN is he digi al numbe o he na u al alues. (σ0 (dB) = 10*log10(absolu e (σ0))). 3.2.3 Sen inel 2 Da a The Sen inel 2 image y was ob ained om 2 sa elli es: Sen inel 2A and Sen inel 2B. Da a is a ailable ee o cha ge in Cope nicus Open Access Hub and GEE pla o m. Sen inel 2 da a a e cha ac e ized by13 spec al bands wi h 10-m, 20-m, and 60-m spa ial esolu ion and a adiome ic esolu ion o 12 bi . Two o m o da a a e a ailable: Le el 1C (L1C)– op-o -a mosphe ic e lec ance p oduc , and Le el 2A (L2A) – bo om-o -a mosphe ic e lec ance p oduc . All he acqui ed images a e om 2017: June and July (16 L1C p oduc s wi h cloud co e less han 15%), Sep embe (8 L1C p oduc s wi h cloud co e less han 25%) and Oc obe (2 L1C p oduc s wi h cloud co e less han 10%). Only 10m and 20m 26 esolu ion bands we e used o he analysis: B2 (490nm), B3 (560nm), B4 (665nm), and B8 (842nm) 10m spa ial esolu ion bands, B5 (705 nm), B11 (1610nm), and B12 (2190nm) 20 m spa ial esolu ion bands (ESA, 2018). 3.2.4 Alos Palsa Da a Global Palsa -2/ Palsa L-band SAR da a we e accessed a a 25 m scale in GEE pla o m. The da ase is gene a ed by applying Japanese Ae ospace Explo a ion Agency’s (JAXA) p ocessing and analysis echnique o a lo o images ob ained wi h Japanese (Palsa and Palsa -2) ada s on Ad anced Land Obse ing Sa elli e (ALOS and ALOS-2) sa elli es/ca ie s (JAXA, 2018). Palsa / Palsa -2 images a e L-band SAR (~23.5 cm wa eleng h) and he images a e acqui ed in Fine Beam Dual pola iza ion (FBD, 70 km swa h wid h) mode: HH- ho izon al ansmi , ho izon al ecei e, HV- ho izon al ansmi , e ical ecei e. The incidence angle is be ween 28.6 deg ee and 32.9 deg ee (CEOS, 2016; JAXA, 2018). The empo al in e al o he images con ained by he mosaic is gene ally 1 yea , and no in o ma ion is a ailable abou he scanning da e o a pa icula a ea. Acco ding o he u o ial o Global 25 m Palsa p oduc , he images a e selec ed aking in o conside a ion he wea he in o ma ion, in o de o a oid ada backsca e ing sa u a ion e ec caused by he mois u e (JAXA, 2018). O ho- ec i ica ion and opog aphic co ec ions on he SAR da a a e applied using 90m SRTM DEM, which is no p e e able o his esea ch, as ou s udy a ea is highly agmen ed and his can cause o impo an in o ma ion lose. Backsca e ing da a a e s o ed in digi al numbe (DN) o unsigned 16 bi . These alues can be con e ed o gamma naugh (γ◦) na u al alues in decibel uni (dB) by he ollowing equa ion: γ◦ = 10⁎log10(𝐷𝑁²)−𝐶𝐹 (dB) (4) Whe e CF is he calib a ion ac o and o p oduc o Palsa / Palsa -2 is measu ed o be ~83.0 dB. 27 3.2.5 Fo es y Ancilla y Da a The o es y ancilla y da a used in he scopes o his esea ch a e he o es ype in o ma ion om o es s a i ica ion map, and he opog aphic pa ame e s o he sampling plo s, namely, he aspec and he slope in o ma ion. Though his in o ma ion is a ailable om he o es in en o y da a, we used he necessa y equi alen in o ma ion e ie ed om DEM. The SRTM DEM which was gained om U.S. Geological Su ey po al (USGS, Sep embe , 2018) ee o cha ge. DEM da a is low spa ial esolu ion (30 m). 28 APPROACHES, METHODOLOGY, DATA PREPARATION 4.1 App oach and Gene al Me hodology Di e en pixel-based me hods we e used in o de o answe he esea ch ques ions and achie e he objec i es. The pixel size is chosen 30m quad a o be equal o he SRTM DEM p oduc , which was used o e ain co ec ions. Also, ce ain ci cums ances we e conside ed: he minimum mapping uni (MMU) is 0,09ha, which is compe i i e wi h he a ea o ield sampling plo s (0,1 ha), and a e applying a scaling ac o o 0,9 on he ield ABG es ima ion, we eplaced he ci cula sampling plo s wi h squa ed pixels. P ojec ion sys em o he whole p ojec was selec ed WGS- 84 38N local o he s udy a ea p ojec ion. All he as e s we e aligned in o de o ma ch he pixels o as e analysis. Field da a was ca e ully examined as ce ain condi ions should ha e been ensu ed be o e using SAR da a o eg ession analysis. E en i a pa o he ield sampling da a we e om he beginning o 2018, we kep using he emo ely sensed da a om 2017. As he annual g ow h o wood o he o es s in A menia is es ima ed as 1.4 m³/ha (Rio+20, 2012), which we conside o be no signi ican o his s udy. Li e a u e e iew shows, ha wo king wi h SAR da a demands deepe unde s anding o he physics behind i and mo e ca e ul p e-p ocessing and p ocessing o he da a, which is c ucial o imp o ing he quali y and e ie ing he ele an in o ma ion om SAR backsca e ing. Implemen a ion o some ypical p ocessing s eps, such as adiome ic co ec ion and speckle il e ing a e equi ed e en i we a e wo king wi h images o m he same senso bu wi h di e en ime s amps (Na a o e al., 2016). E en mo e, wi hin he scope o his s udy we combined da a om di e en senso s, which assumes mo e p e-p ocessing s eps such as iden ical backsca e ing naugh e ie al. As he s udy a ea is su e ing om complex and highly segmen ed opog aphy and aking in o conside a ion he ac ha in such case e en he adiome ic e ain co ec ion is no su icien o mi iga ing he bias and he gain on he SAR backsca e ing da a caused by high opog aphic a ia ion (Pohl C., 2017), we implemen ed ancilla y opog aphic in o ma ion o he sampling si es di ec ly in o 29 eg ession model. Simple emo al o he s eep slopes and masking o he o esho ening and shadows on he SAR da a is no possible, as hose a eas a e occupying he mos pa o he s udy a ea. Besides he ee co e ed a eas, o es si es include also non- o es ed o es lands (meadows, pas u es, he bs and o es dis u bance a eas) a e some imes signi ican ly big and can mislead he o es biomass es ima ion. Fo he FNF map gene a ion in pu pose o p ecise delinea ion o o es s, Sen inel 2 image y and Random Fo es (RF) classi ica ion me hod was used. This non-pa ame ic classi ie was selec ed since i does no equi e any s a is ical analysis o inpu a iable and can handle wi h mul icollinea i y e ec . Recen S udies show success ul implemen a ion o Ha alic ex u e analysis (GLCM ma ices) o biomass p edic ion, when di e en pa ame e s o he local a iance o he pixel alues is calcula ed (Chen e al., 2018; Huang e al.,, 2018). Image ex u e analysis echnique was also applied by choosing he mo e ele an ex u e pa ame e s. Fo he biomass p edic ion and AGB mapping backwa d S epwise Mul iple Linea Reg ession (SWR) model was used o au oma e he bes explana o y a iable selec ion. E en i he s udied s a e some be e esul s o non-pa ame ic models o e he pa ame ic ones o biomass p edic ion, he use o SWR allows us o e alua e and compa e each o he inpu a iables (i.e. he inpu da a and he en i onmen al condi ions), which is one o he main ocus a eas o his esea ch. Va iables wi h pa ame e s o p- alue < 0.05 and VIF > 5 we e excluded o m eg ession model (Be ninge e al., 2018; Lau in e al., 2018). The Figu e 4 p o ides he lowcha o he o e all me hodology (The comple e low cha o he me hodology see in APPENDIX B). The low cha d is designed ollowing way: aw inpu da a (g ay), in o ma ion o unp ocessed da a (blue), da a p e-p ocessing and p ocessing s eps (whi e), p ocessed- eady o analysis da a (g een), classi ica ion and eg ession analysis (b own). Red dashed boxes a e indica ing he main sub-p ocesses which will be discussed in a de ailed manne in he nex chap e s. Those sub-p ocesses a e: • Field da a calcula ion • Sen inel 1 da a p e-p ocessing • PALSAR da a p e-p ocessing 30 • GLCM ex u e analysis • Sen inel 2 da a p e-p ocessing • Ancilla y da a p e-p ocessing • Fo es /Non-Fo es classi ica ion The eg ession analysis, model diagnos ics and e alua ion a e discussed in he Chap e 5. Figu e 4. Flow cha o gene al me hodology 31 4.2 Tools The ools used o he da a p e-p ocessing and p ocessing a e p esen ed in he Table 3. The ocus is on he open so wa e and ee o cha ge cloud en i onmen . Sa elli e image p e-p ocessing and p ocessing is mos ly ca ied ou in he Google Ea h Engine cloud-based en i onmen by using Ja aSc ip API. P ocess Tools S1 p e-p ocessing GEE Ja aSc ip API Palsa mosaic p e-p ocessing S2 ege a ion indices calcula ions Hansen Global o es map implemen a ion FNF classi ica ion wi h RF Classi ica ion accu acy assessmen GLCM ex u e analysis S2 L1C o L2A p ocessing SNAP, Sen inel Toolbox, Sen2Co plugin S2 ege a ion indices calcula ions Reg ession analysis R C oss- alida ion and model diagnos ics AGB mapping on he s udy a ea P ojec ion ans o ma ion QGIS Ras e e-scaling and aligning Field sampling da a p ocessing Shape o as e con e sions Map designing A cGIS Desk op Aspec and Slope p ocessing Field da a calcula ions MS O ice Excel Table 4. Tools and ela ed p ocessing GEE is cloud based geospa ial p ocessing pla o m and p o ides huge compu a ional powe , which made possible he p ocessing o eno mous amoun o sa elli e image y and conduc dense ime-se ies analysis. GEE is an en i onmen o plane a y-scale en i onmen al da a analysis and con ains he a chi es o many publicly a ailable emo e sensing image y (Go elick e al., 2017). 38 Because he aspec da a is ca ego ical, we combined i in o 8 segmen s, each g oup consis ing o 45 deg ees o aspec . Comple ely la a eas consis 0.8% o he o al a e, so we dis ega ded ha and included in he 1 h segmen . The numbe ing is designed s epwise. A e wa ds, we implemen ed one-ho -encoding echnique o ans o m he ca ego ical a iable in o 8 di e en inpu a iables wi h nume ical alues using R (R Co e Team, 2018). Thus, we made i possible o his in o ma ion o be used in linea eg ession analysis as an inpu a iable. “Encoding” package was used o his. Figu e 10. DEM, Aspec and Slope maps o he s udy a ea 4.8 GLCM ex u e analysis Image ex u al ea u es, de eloped by Ha alic e al. (1973), measu e he spa ial homogenei y o he backsca e ing and con ain in o ma ion abou o es s uc u e. Se e al s udies ha e p o en ha GLCM ex u e om SAR can be be e desc ibing he biomass dis ibu ion han SAR backsca e ing i sel (Chap e 2.2). GLCM analysis we e applied only on he SAR image y wi h a selec i e app oach (Figu e 11), aiming o dec ease he numbe o inpu a iables o he eg ession analysis. 39 Figu e 11. Flow cha o GLCM ex u e analysis s eps In o de o de e mine he mo e ele an SAR a iables, he eg ession model was un wi h he inpu a iables o S1 (22 a iable om ime-se ies s acks o each VV and VH pola iza ion), Palsa (2 a iable o HH and HV pola iza ion), Vege a ion indices (3 a iable – NDVI, NDMI, RVI) and ancilla y da a (3 a iable - ee species ype, aspec , slope). Mo e impo an a iables we e selec ed based on hei co ela ion signi icance acco ding o he s epwise linea eg ession es (SWR). In he i s i e a ion he mo e signi ican SAR a iable (Palsa HV) was selec ed, hen 14 ex u e measu es we e calcula ed o Palsa HV in GEE (Angula Second Momen , Con as , Homogenei y, Co ela ion, Va iance, Sum a e age, Sum a iance, En opy, Sum en opy, Di e ence en opy, Ene gy, Di e ence a iance, Di e ence en opy, Maximum co ela ion). Those masseu s as inpu a iables we e added o he same a iables o ano he i e a ion o eg ession. This ime he bes ex u e measu es wi h highes co ela ion wi h he biomass we e de e mined again based on he a iable signi icance om SWR es . Those measu es a e Va iance, En opy, Co ela ion, Sum o a e age (Table 6). GLCM ex u e analysis we e ca ied wi h GEE cloud en i onmen . Fo hese analysis SAR da a in gamma naugh (con e ed o na u al alues) a e used. To ha e he eal s a is ics o he image pixels, ex u e measu es a e calcula ed on he images be o e applying a speckle il e . 4x4 window size was se up o he calcula ions. 40 GLCM ex u e Fo mula Desc ip ion No o eq. Va iance Measu es he dispe sion (wi h ega d o he mean) o he g ay le el dis ibu ion (12) En opy Measu es he deg ee o diso de among pixels in he image; i is (app oxima ely) in e sely co ela ed wi h uni o mi y; images wi h a la ge numbe o g ay le els ha e la ge en opy (13) Co ela ion Measu es he linea dependency o g ay le els on hose o neighbo ing pixels; i p o ides a measu e simila o au oco ela ion me hods (14) Sum a e age Measu es he mean o he g ay le el sum dis ibu ion o he image (15) Table 6. Fo mulas o GLCM ex u e measu es. Fo mulas o GLCM ex u e measu es. Fo all he equa ions, p(i, j) is he (i, j)- h en y o he no malized g ay-le el co-occu ence ma ix, ha means, p(i, j) = P(i, j) / ∑P(i,j) 𝑖𝑗 , whe e P(i, j) is he (i, j)- h en y o he compu ed GLCM; 𝑁𝑔 is he o al numbe o g ay le els on he image; and 𝜇𝑥, 𝜇𝑦 and 𝜎𝑥,d𝜎𝑦 s and o he Mean and S anda d De ia ion o he ow and column sums o he GLCM, espec i ely (da a sou ce: Ha alic e al., 1973). ∑ ∑ (𝑖−𝜇)2 𝑁𝑔 𝑗=1 𝑝(𝑖,𝑗) 𝑁𝑔 𝑖=1 −∑ ∑ 𝑝(𝑖,𝑗) 𝑁𝑔 𝑗=1 log𝑝[(𝑖,𝑗)] 𝑁𝑔 𝑖=1 ∑ ∑ 𝑖𝑗𝑝(𝑖,𝑗)− 𝜇𝑥𝜇𝑦 𝜎𝑥𝜎𝑦 𝑁𝑔 𝑗=1 𝑁𝑔 𝑖=1 ∑ 𝑖𝑝𝑥+𝑦 −(𝑖) 2𝑁𝑔 𝑖=2 41 DATA PROCESSING․ FOREST BIOMASS MONITORING 5.1 Fo es /Non-Fo es classi ica ion Be o e s a ing he o es moni o ing and biomass calcula ions, i ’s signi ican o dis inguish he main concep s ega ding o es de ini ion, which will b ing be e insigh in o he ield da a we ha e, and on he ex en s o he span ha should be explo ed wi hin his esea ch. As he o es y ield in o ma ion is p o ided by each o es en e p ise (bigges o es si e uni s), i is wo h o de ine he o es en e p ise o his case. Acco ding o he Fo es Code o A menia: Fo es en e p ise – a p oduc ion uni wi h he aim o sus ainable o es managemen . As he o es en e p ise is o es -economic uni , i consis s o o es ed and non- o es ed lands. Acco ding o he same Code: Fo es Lands - o es ed lands and lands alloca ed o en isaged o lo a and auna p o ec ion, na u e p o ec ion as well as non- o es ed lands alloca ed o en isaged o he unning o o es economy. On he o he hand, ecen s udies show ha he non-homogeneous o es ed a eas can cause o SAR backsca e ing alue dis o ion on he image. To a oid om his kind o unce ain ies, we o e o delinea e he bounda ies o o es s only, inside he o es en e p ises and conside hose a eas as he limi o he ex en s o his s udy. Acco ding o he same o es Code: Fo es - in e connec ed and in e ac ing in eg i y o biological di e si y domina ed by ee-bush ege a ion and o componen s o na u al en i onmen on o es lands o o he lands alloca ed o a o es a ion wi h he minimal a ea o 0,1 ha, minimal wid h o 10 m and wi h ee c owns co e ing a leas 30% o he a ea, as well as non- o es ed a eas o p e iously o es ed o es lands (Fo es Code, 2005). Thus, wi hin he scope o his esea ch we closely aligned he o es de ini ion o he one o Fo es Code A menia, meaning, ha he o es s a e de ined as lands o mo e han 0.09 ha wi h he ee canopy co e o mo e han 30%. Fo his pu pose, Hansen o es co e p oduc o 2000 was adop ed, as well as o es /non- o es (FNF) bina y classi ica ion was pe o med using S2 p oduc and Random Fo es classi ie (Figu e 12). 42 Figu e 12. Flow cha o o es /non- o es classi ica ion s eps This FNF classi ica ion wi h a new da a was necessa y due o he li e a u e indica ing he scales o illegal elling (Chap e 1.3) and he Global Fo es Wa ch online pla o m (Global Fo es Wa ch), a i ming ha he e a e signi ican dis u bance spo s in he o es s and he e is a need o upda ing he Hansen map a e 17 yea s. On he o he hand, a emp s o adjus Hansen Global map wi h simple h eshold o o es non- o es delinea ion ailed o հis s udy a ea. The aining and es ing da ase s we e gene a ed and a bina y classi ica ion was ca ied ou in GEE cloud en i onmen using Ja aSc ip API and was pe o med in wo s eps. Fi s , o es canopy co e h eshold o 30% was se up using Hansen global o es co e map, as each pixel on his map ep esen s he canopy co e pe cen age. This way we de ined o es and non- o es a eas o he yea o 2000. Tha was pa icula ly done o aining da ase gene a ion. Fo ha pu pose, “s a i iedSample” ool was used o gene a e andomly dis ibu ed poin s wi hin a g id wi h a scale o 30m. 200 hund ed poin s we e gene a ed his way: 100 poin s wi hin he o es a ea and 100 poin s wi hin he non- o es a ea. Valida ion con usion ma ix P edic ion n =200 Posi i e Nega i e To al Ac ual Posi i e TP = 97 FP = 3 100 Nega i e FN = 4 TN = 96 100 To al 101 99 193 Table 7. Valida ion con usion ma ix o o es non- o es classi ica ion 43 In he second s ep Random Fo es classi ie wi h 300 ees was ained using he S2 da a om 2017 and he aining da ase . FNF classi ica ion was d i en and he allowing accu acy pa ame e s we e assessed again in GEE en i onmen (Tables 7 ). The RF pe o mance on bina y classi ica ion shows high sensi i i y: 0.96 (p opo ion o ac ual posi i es ha a e co ec ly iden i ied as such) and speci ici y: 0.97 (p opo ion o ac ual nega i es ha a e co ec ly iden i ied as such) using S2 da a. Thus, he o e all accu acy o classi ica ion is 0.965 and we use his classi ica ion esul as a base map o ou esea ch o delinea e only o es ed a eas. Kappa s a is ics o he classi ica ion was 0.93. As he esul , 86% (25184.4 ha) wi hin he s udy a ea was classi ied as o es and 14% (4149.6 ha) was classi ied as non- o es (Figu e 13). Majo i y o ing il e was applied on he ou pu as e map once. Figu e 13. Map o FNF bina y classi ica ion. Non- o es ed si e examples on he snapsho s on he igh (g een line shows he classi ied o es bo de , ed line shows he o es en e p ise bo de ). 44 5.2 Resul s o Reg ession analysis The eg ession analysis was conduc ed using AGB da a (adjus ed by scaling ac o o 0.09 ha MMU) as a dependen a iable and he es o he inpu s as independen a iables (p edic o s). The sequence o he s eps a e ollowing he gene al low cha in he sec ion o me hodology (Chap e 4.1). Di e en combina ion o p edic o a iables we e calcula ed and he eg ession model has been e alua ed o mul icollinea i y in p edic o s (VIF es ), no mali y (Shapi o-Wilk no mali y es ) and au oco ela ion (co elog am) in esiduals o each ime. Reg ession analysis we e ca ied ou wi h R p og amming language and “Akaike c i e ion” unc ion, which s ands o mul ilinea s epwise eg ession. Fo he eg ession analysis o al o 60 inpu a iables we e used (Table 8). Sou ce Numbe o a iables Explana ion Va iable To al Sen inel 1 12 VV mon hly s ack - 2017 32 12 VH mon hly s ack - 2017 4 GLCM VV10 (Co , En , Va , Sa g) 4 VH12 (Co , En , Va , Sa g) Sen inel 2 3 NDVI, NDMI, RVI 3 Alos Palsa 2 HH, HV - 2017 10 4 GLCM HH (Co , En , Va , Sa g) 4 HV (Co , En , Va , Sa g) Ancilla y da a Aspec 8 aspec ca ego ies 15 Slope 1 slope alues Fo es ype 6 dominan ee ype To al amoun o a iables 60 Table 8. Inpu a iables o eg ession analysis Ini ially, only 24 obse a ions om S1 ime se ies (VV and VH) we e used o eg ession analysis in o de o selec he bes scenes om S1 wi h highe impo ance in AGB p edic ion. This was done o wo main pu poses: Fi s , o e alua e he p e e ence o using S1 dense ime se ies ins ead o choosing one scene o o es AGB calcula ions and second, o dec ease he numbe o inpu a iables o GLCM ex u e analysis o he u he use. This was pa icula ly impo an o keeping he numbe o inpu a iables as less as possible, as he aining da ase ( ield da a) was no la ge. 45 F om he pe spec i e o he phy oclima ic seasonali y in he o es in ou s udy a ea, he co ela ion be ween ield AGB and Sen inel 1 C-band s ack ime se ies we e explo ed i s using Pea son’s co ela ion coe icien (Figu e 14). This shows he co ela ion o be almos always highe o VH c oss-pola iza ion compa ed wi h VV co-pola iza ion. Ne e heless, o bo h pola iza ion modes i is lowe han 0.5, and ge s o i s maximum o he s ack a e ages o Oc obe and Decembe (bo h o VH backsca e ). The assump ion is ha C-band backsca e ing be e ep esen s he o es biomass du ing he lea es-o season, and when he soil is no ozen. Figu e 14. Pea son’s co ela ion coe icien calcula ed be ween Sen inel 1 s ack ime se ies and AGB om he 79 plo s and o each o he ime se ies. The eg ession analysis we e epea ed using he o es ype in o ma ion as a ca ego ical a iable. The esul s showed ha among he S1 ime se ies only he s acks o VH pola iza ion om Oc obe and Decembe a e passing he minimum limi o p- alue < 0.05. Also, bo h HH and HV pola iza ion mosaics om Palsa we e chosen by eg ession model o be impo an p edic o s o AGB. The o es ype in o ma ion was ne e selec ed as impo an , he e o e was d opped om la e analysis. Hence, we deduce ha SAR backsca e ing is no sensi i e o he ee ype. Those wo s acks o m S1 and HH, HV Palsa mosaics we e chosen o GLCM ex u e measu es and o u he eg ession analysis. The GLCM analysis we e done as i is desc ibed in he Chap e 4.8. Fu he , du ing his eg ession analysis, 2 samples wi h ex eme biomass alues o he s udy a e (680 ℎ𝑎−1 and 500 ℎ𝑎−1), as well as 1 sample wi h unusual ex eme pixel alues on Palsa scenes we e excluded om he u he analysis, as hey we e causing majo unce ain y in he p edic ion. As ollows, 79 ield sampling poin s we e conside ed o he inal eg ession analysis. 46 5.2.1 Model Compa ison and P edic o E alua ion Since he objec i e o his s udy is o e alua e he di e en o igin da a o o es AGB p edic ion, 4 eg ession models we e p oposed wi h di e en p edic o combina ions (Table 10). In his sec ion we compa ed p edic ion models wi h a ocus o de ec ing ela i e di e ence in he p edic ion accu acy. Fo his pu pose, mul iple linea eg ession analysis was ca ied ou wi h di e en p edic o usion aiming o de e mine he bes models using di e en ype o da a in di e en combina ions. Acco dingly, 4 models wi h he ollowing combina ions we e de eloped: Model 1 – only SAR da a, Model 2 – SAR+op ical da a, Model 3 – SAR+ancilla y da a, Model 4 – SAR+op ical+ancilla y da a. As many o he pa ame e s a e gene a ed om a ying decomposi ions o he same da a o he same sou ce, se e al o hese me ics a e expec ed o be highly co ela ed (Appendix B). High co ela ion amongs he p edic o s causes o model o e i ing and biased R² alue. The e o e, wi h he h eshold o a iance in la ion ac o maximum o 5 was se up. Adjus ed R² is calcula ed. Fu he , Lea e-One-Ou c oss- alida ion is ca ied ou and R² as well as RMSE a e calcula ed. As he RMSE ep esen s he ela i e e o o 0,09 ha a ea (MMU), la e he absolu e RMSE was calcula ed o he hec a e and is p o ided in he Table 9. Linea Model 1 consis ing o S1 C-band and Palsa L-band SAR da a and he GLCM ex u e analysis o hose da a yielded low p edic ion esul s: Adjus ed R² equal o 0.46, LOO alida ion es ga e R² equal o 0.38 and RMSE equal o 70 pe hec a e. The bes p edic o a iables we e Palsa HH and HV pola iza ion o m he ini ial image y, and he es o he a iables we e di e en ex u e measu es: Sum o A e age (SAVG) GLCM o S1 s ack o Oc obe , SAVG o Palsa HV. I is wo h o men ion, ha none o he Sen inel ini ial images appea ed in any o he models as impo an p edic o s. Model 2 combined all he inpu s om Model 1 and he 3 ege a ion indices (NDVI, NDMI, RVI) om S2. De e mina ion coe icien o LOO R² dec eases insigni ican ly o 3.3 and he RMSE inc eases o 72.8 ha¯¹. The s epwise p ocess disca ded many o he a iables ha we e impo an o he Model 1 and educes he dimension o he p edic o s. Palsa HH and HV s ay as he mos impo an p edic o s oge he wi h En opy measu e o Oc obe s ack om S1 da a. NDVI a iable was be e sui ed o 47 AGB p edic ion amongs he ege a ion indices wi h he less signi icance o he model. Model (sou ces) P edic o s VIF (<5) p- alue<0.05 (signi icance) Adj. R² LOO R² ela i e RMSE ( /0,09 ha) RMSE ( /ha) Model 1 Palsa HH 2.26764 1.68E-05 *** 0.46 0.38 6.3 70 Palsa HV 2.211967 1.18E-05 *** S1 VH 10 sa g 1.913548 6.34E-05 *** SAR S1 VH 12 co 1.744905 0.03294 * S1 VH 10 en 2.79781 0.03224 * Palsa HH en 1.703185 0.00836 ** Palsa HV sa g 2.345768 2.16E-05 *** Palsa HH co 1.483653 0.04045 * Model 2 Palsa HH 2.703908 3.24E-06 *** 0.41 0.33 65.6 72.8 Palsa HV 2.21506 0.00045 *** SAR+ Op ical NDVI 1.104523 0.022439 * S1 VH 10 en 1.134397 0.000457 *** Palsa HH sa g 2.323204 0.020287 * Model 3 Aspec (4,6,7) 1.21354 5.00E-07 *** 0.65 0.6 50 57.7 Slope 1.283596 7.15E-07 *** SAR+ Ancilla y Palsa HV 2.930844 9.66E-06 *** Palsa HH 2.438455 0.002232 *** Palsa HV sa g 1.721521 3.65E-10 *** S1 VH 10 en 2.126821 9.23E-05 *** Model 4 Aspec (4,6,7) 1.216558 0.000842 *** 0.68 0.62 4.9 56.6 Slope 1.290634 1.90E-07 *** SAR+ Op ical+ Ancilla y Palsa HV 2.933416 4.20E-06 *** Palsa HH 2.459693 0.000854 *** Palsa HV sa g 1.122131 0.000503 *** S1 VH 10 en 1.736985 4.22E-10 *** NDVI 1.122131 0.019642 * Signi icance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 Table 9. E alua ion o s a is ics o eg ession analysis o each model The p edic ion accu acy changes eno mously when he o es ancilla y da a (namely, aspec , slope and he o es ype) we e added o he p e ious da a. Model 3 in able 9 shows he o e iew o he s a is ics o eg ession analysis wi h all he inpu s wi hou da a om op ical senso , and he Model 4 combines also he op ical da a. The LOO- alida ion R² o he Model 3 inc eases up o 0.6 and RMSE dec eases o 57.7 ha¯¹. Fo he Model 4 he p edic ion imp o es e y sligh ly wi h he implemen a ion o 54 combined in he same model o eg ession analysis. Simila ly, Huang uses 5 di e en measu es o m he same sou ce o aining he bes model (Huang e al., 2018). The co ela ion o op ical da a o o es AGB was e y weak and only NDVI amongs he es ed ege a ion indices was selec ed by SWR wi h low impo ance. The insigni ican con ibu ion o op ical da a in he AGB p edic ion can be explained wi h he ac , ha he a e age biomass o he o es s o his s udy a ea was measu ed 148 ha¯¹ while he obse ed sa u a ion le el o op ical da a in AGB p edic ion in he li e a u e al e s a ound 50 – 70 ha¯¹ biomass (Zhang e al., 2018). Al hough some o he s udies depic highe co ela ion be ween NDMI index and he biomass (li e a u e e iew by Joshi e al., 2016), his analysis showed no simila esul s. I is possible ha o NDMI calcula ion mo e humid season is p e e ed, while he e was no cloudless image o he s udy a ea om Ma ch and June. The use o ancilla y da a oge he wi h all he emo ely sensed da a b ough a g ea impac on he eg ession model accu acy imp o ing R² up o 0.62. This is he bes esul achie ed wi hin his esea ch and is e y impo an , as one o he hypo heses and he main inno a ion o e ed by his pape was he assump ion ha he gain o bias o he SAR backsca e ing alue in moun ainous a eas canno be co ec ed by he adiome ic e ain co ec ion me hods, i migh be mi iga ed by he p edic ion model when he o es y ancilla y da a on he sampling si e is gi en. Amongs he ancilla y da a he slope and aspec in o ma ion we e always in a highe co ela ion wi h he biomass while combined wi h SAR da a. Fo es ype in o ma ion did no imp o e he biomass p edic ion in con as o o he s udies (Lau in e al., 2018), which ound i an impo an componen o he p edic ion model. E en i he implemen a ion o ancilla y da a g ea ly imp o es he p edic ion model pe o mance as well as he isual obse a ion o he p oduc maps depic imp o emen in biomass dis ibu ion, mo e ield da a is needed o check he eliabili y o hose models o implemen ing he ada emo ely sensed da a in ca bon s ock moni o ing ac i i ies in moun ainous o es s. Backwa d s epwise mul iple linea eg ession echnique conside ably au oma ed he bes a iable selec ion as well as isola ed he mul icollinea i y e ec . Fo ins ance, VIF es applied a e SWR es did no de ec mul icollinea i y in he a iables. Ne e heless, no all he s udies use he same me hods and measu es o model diagnos ics, which makes he esul e alua ion and compa ison qui e p oblema ic. Fo 55 ins ance, many pape s use only adjus ed coe icien o de e mina ion (R²) o eg ession models o he e alua ion (Chen e al., 2018; Ecke , 2012; Huang e al., 2018), which is a esul s wi hou model c oss- alida ion, so we canno di ec ly compa e ou esul wi h he ones om hose wo ks. The sa u a ion le el o SAR backsca e ing as a gene al limi a ion, was ha d o de ine as he backsca e ing alue is su e ing by he opog aphy and canno se e o such calib a ion. Howe e , om he model 1 which uses SAR da a only, we can obse e le el o 150-160 ha¯¹, a e which he p edic ion line indica es mos ly unde p edic ion. This esul is simila o he one om Be ninge , 2018, e en hough he claims sa u a ion le el o L-band wi h ex u e measu es a e possible o imp o e up o 250 ha¯¹. The sa u a ion le el esea ch can be explo ed in a mo e accu a e way once mo e sampling da a is a ailable, which can make possible o sepa a e only la a eas no a ec ed by shadowing o o esho ening and do analysis wi hou applying opog aphic in o ma ion, e en hough diag am (Fig. 18) shows highe unce ain y o he low biomass p edic ion. This is likely because he measu ed ees a e only he ones la ge han 8 cm DBHOB, he e o e smalle ees a ailable in he sampling plo s can ha e signi ican impac on he Sen inel 1 C-band (3.75–7.5 cm wa eleng h) backsca e ing causing o AGB o e es ima ion which makes his ange (<100 ha¯¹) he mos in luen ial. This end o o e es ima ion o samples wi h smalle biomass alue is obse ed in all he simila esea ches (Be engue e al., 2018; Be ninge , e al., 2018; Huang e al., 2018; Lau in e al., 2018). E en i e y ew o he e iewed pape s pe o m new delinea ion o o es bounda y, he impo ance o FNF classi ica ion pe o med wi hin he ames o his pape wo k is p o en wi h wo main poin s: • FNF upda e shows signi ican (13%) non- o es ed a eas wi hin he o es en e p ise, which should be aken in o conside a ion while es ima ing he ca bon s ock o mo e ealis ic ca bon budge e alua ion. • As he Hansen map is a global model, he FNF delinea ion based on ha map was no adjus able o he s udy a ea. The e o e, he ield sampling plo s design gene a ed based on he Hansen map can mislead he e iciency o he ield wo k. As he esul showed, 26 sampling plo s ou o 115 we e loca ed ou side o he newly upda ed o es bounda ies. 56 Thus, we conclude ha he biomass da a calcula ed by s a is ical me hods using he allome ic equa ions can be signi ican ly a ec ed by he easons men ioned abo e. As his pape was aiming o imi a e he o es ca bon moni o ing ac i i ies, and pa icula ly, o es AGB measu emen s, we a e planning o compa e ou esul s wi h he one om adi ional NFI once i becomes a ailable. The inal biomass map has 30 m spa ial esolu ion, which is he i s map ha ing synop ic iew on he spa ial dis ibu ion o he biomass in NE o es o A menia, ye wi h a compa able accu acy. The de eloped model is applicable o simila o es s whe e in si u da a is a ailable. Ano he e y impo an achie emen is ha his ine esolu ion map con aining he p opo ional dis ibu ion o biomass can lead o mo e co ec ca bon es ima ion when used as sou ce o ca bon- ela ed models. The e o e his s udy can acili a e he GEF/UNDP-REDD+ ac i i ies in NE o es s in A menia o achie e hei objec i es in be e accu acy, as well as help o make o es ca bon s ock measu emen s and moni o ing cos -e ec i e. In o de o calcula e he ca bon s ock o he s udy a ea, he modeled o es s and biomass was educed by a ac o o 0.47 (UNDP-GEF 00091048, 2015). Fo he model 1 measu ed ca bon s ock in he s udy a ea was 4.1 G , and o he model 4 i was 3.5 G (Fig. 16). This da a is ye o be compa ed wi h NFI esul s once i is a ailable. 5.4 Limi a ions and Recommenda ions o Fu u e Resea ch One o he majo limi a ions o his s udy was he lack o la a eas ha could be analysed sepa a ely om hose a eas wi h complex opog aphy in o de o be able o calib a e he ue alue o SAR backsca e ing o he o es s in he s udy a ea. The nex obs acle was he lack o sampling poin s. As seen in numbe o s udies, he ou comes o such esea ch a e s ongly dependan on he sampling da a quan i y and quali y․ As o he u u e, i is planned o collec o al o 315 samples o his s udy a ea wi hin he GEF-UNDP p ojec . This amoun o da a i sel is signi ican ly bigge o model aining and can make i possible o de elop mo e sus ainable p edic ion models and ca y ou esea ch only on he ela i ely la a eas. This s udy, on he o he hand, made i ob ious ha complex opog aphy pu s high unce ain y on he s udy in he way, ha o es s uc u e and dis ibu ion changes quickly on e y sho dis ances, which is no desi able o SAR da a. This is because i 57 is no e y clea i he pixel on he SAR scene is i ing wi h he ield sampling plo e y well. Thus, he backsca e ing alue can be p esen ing no he eal biomass in o ma ion o he sampling plo . GLCM ex u e measu es can be s onge o weake co ela ed o o es biomass depending on he selec ed window size o he ope a ion (Ecke , 2012). As o his esea ch only one window size was adop ed (4x4 ke nel size), hus, we assume GLCM analysis o be no exhaus i e and we conside po en ial possibili y o imp o e p edic ion accu acy by adop ing di e en window sizes o ex u e measu es. On he o he hand, ex u e analysis should be es ed on he aw L-band ada da a in o de o keep he ue speckle dis ibu ion and backsca e ing alue s a is ics. When ield da a becomes a ailable, his me hodology o biomass p edic ion can be applied pe speci ic g oup o biomasses. We assume o ha e be e co ela ion be ween SAR backsca e ing alue and he biomass o unde 150-160 ha¯¹ whe e he sa u a ion le el is obse ed. As he ele ance o he SAR da a and i s combina ion wi h o he sou ce da a a e well explo ed and he bes inpu a iables o o es AGB p edic ion a e desc ibed, we p opose applica ion o non-pa ame ic eg esso s o gene a ing mo e accu a e biomass map, as he e iew o he la es s udies e eal hei p io i y o e pa ame ic eg esso s (Chap e 2.3). 58 CONCLUSION O e all, his s udy shows he use ulness o SAR backsca e ing da a on o es biomass mapping, which allows o ha e spa ially explici AGB dis ibu ion wi hin he o es si e. In his esea ch, we p opose a me hod based on he combina ion o mul isou ce emo ely sensed da a, image ex u e measu es and o es ancilla y in o ma ion o o es AGB es ima ion on he moun ainous a eas. Fo his pu pose, only eely a ailable da a sou ces we e used. The esul s in e he impo ance o mul i empo al C- band da a o a oid om backsca e ing sa u a ion om mois u e, also o mul i empo al speckle il e ing. Howe e , he e is a need o in eg a ion C-band wi h L-band da a o educe he unce ain y in he p edic ion. This s udy also showed ha he GLCM ex u e cha ac e is ics o SAR da a we e he mos ele an p edic o s o explaining he obse ed a iabili y o AGB in he s udy a ea. The ela ionships be ween SAR da a and s and cha ac e is ics, howe e , can be nega i ely in luenced in a eas o high opog aphic a ia ion and make he AGB modelling uly p oblema ic. The aspec and slope ancilla y da a was able o mi iga e he opog aphy e ec on SAR da a and ensu ed imp o ed eg ession analysis wi h accu acy o R² = 0.62 and RMSE = 56.6 ha¯¹. The ou pu o his esea ch will be p o ided o GEF/UNDP p ojec in NE o es s o Republic o A menia and i can ha e an inpu in achie ing hei objec i es, as well as suppo ing o es moni o ing and ca bon s ock egula ions. The achie ed esul s need o be alida ed wi h mo e sophis ica ed g ound- u h da a in o de o be implemen ed as cos -e ec i e su oga e o NFI ac i i ies. This AGB es ima ion app oach is adap able and allows modeling o biomass in o he moun ainous o es s wi h simila condi ions. Fu he , i is wo h no ing ha wo king wi h SAR da a demands p io knowledge and deep unde s anding o p e-p ocessing s eps and he physics behind hem. I should also be said, ha his s udy was ca ied ou wi h eely a ailable emo ely sensed da a, and mo e p e e ed comme cial da a a ailabili y can signi ican ly inc ease he model pe o mance accu acy. 59 BIBLIOGRAPHIC REFERENCES Aula d-Macle M., R. Ba s ow, D. R. (2011). Sen inel-1 P oduc De ini ion. Balz e , H., Cole, B., Thiel, C., & Schmullius, C. (2015). Mapping CORINE Land Co e om Sen inel-1A SAR and SRTM Digi al Ele a ion Model Da a using Random Fo es s. Remo e Sensing, 7(11), 14876. h ps://doi.o g/10.3390/ s71114876 Be engue , E., Oszwald, J., Id, I. T., Lau en , F., Hasan, A. F., & Sis , P. (2018). The Po en ial o Mul isou ce Remo e Sensing o Mapping he Biomass o a Deg aded Amazonian Fo es , 1–21. h ps://doi.o g/10.3390/ 9060303 Be ninge , A., Lohbe ge , S., S ängel, M., & Siege , F. (2018). SAR-Based Es ima ion o Abo e-G ound Biomass and I s Changes in T opical Fo es s o Kaliman an, 2. h ps://doi.o g/10.3390/ s10060831 Bou e , A., Me moz, S., Le Toan, T., Villa d, L., Ma hieu, R., Naidoo, L., & Asne , G. P. (2018). An abo e-g ound biomass map o A ican sa annahs and woodlands a 25 m esolu ion de i ed om ALOS PALSAR. Remo e Sensing o En i onmen , 206(Decembe 2017), 156–173. h ps://doi.o g/10.1016/j. se.2017.12.030 B uniquel, J., & Lopes, A. (2010). Mul i- a ia e op imal speckle educ ion in SAR image y, (No embe 2014), 37–41. h ps://doi.o g/10.1080/014311697218962 CEOS. (2016). In e p e a ion Guide o ALOS PALSAR / ALOS-2 PALSAR-2 global 25 m mosaic da a Table o Con en s, 1.1, 1–15. Re ie ed om h p://www.g oi.o g/wp-con en /uploads/2016/10/PALSAR-In e p e a ion-Guide- 1.1.pd Cha e, J., Andalo, C., B own, S., Cai ns, M. A., Chambe s, J. Q., Eamus, D., … Yamaku a, T. (2005). T ee allome y and imp o ed es ima ion o ca bon s ocks and balance in opical o es s. Oecologia, 145(1), 87–99. h ps://doi.o g/10.1007/s00442-005-0100-x Chen, L., Ren, C., Zhang, B., Wang, Z., & Xi, Y. (2018). Es ima ion o Fo es Abo e- G ound Biomass by Geog aphically Weigh ed Reg ession and Machine Lea ning wi h Sen inel Image y, 1–20. h ps://doi.o g/10.3390/ 9100582 60 CRISP 2001. (n.d.). P inciples o Remo e Sensing - Cen e o Remo e Imaging, Sensing and P ocessing, CRISP. Re ie ed Oc obe 19, 2018, om h ps://c isp.nus.edu.sg/~ esea ch/ u o ial/sa _in .h m de Bad s, E. P. J. (2002). Global Land Co e 2000: E alua ion o he SPOT VEGETATION senso o land use mapping, 49. Dos álo á, A., Wagne , W., Milenko ić, M., Hollaus, M., Dos álo á, A., Wagne , W., … Wagne , W. (2018). Annual seasonali y in Sen inel-1 signal o o es mapping and o es ype classi ica ion. In e na ional Jou nal o Remo e Sensing, 00(00), 1–23. h ps://doi.o g/10.1080/01431161.2018.1479788 Dungan, J. L. (2002). Towa d a Comp ehensi e View o Unce ain y in Remo e Sensing Towa d a Comp ehensi e View o Unce ain y in Remo e Sensing Analysis Unce ain y in Remo e Sensing Analysis. In Unce ain y in Remo e Sensing and GIS. by Foody, M. Giles A kin on, M. Pe e (pp. 25–35). Depa men o Geog aphy, Uni e si y o Sou hamp on, UK. Ecke , S. (2012). Imp o ed o es biomass and ca bon es ima ions using ex u e measu es om wo ldView-2 sa elli e da a. Remo e Sensing, 4(4), 810–829. h ps://doi.o g/10.3390/ s4040810 Englha , S., Keuck, V., & Siege , F. (2012). Modeling abo eg ound biomass in opical o es s using mul i- equency SAR da a-A compa ison o me hods. IEEE Jou nal o Selec ed Topics in Applied Ea h Obse a ions and Remo e Sensing, 5(1), 298–306. h ps://doi.o g/10.1109/JSTARS.2011.2176720 ESAa. (2019). Use Guides - Sen inel-1 SAR. Re ie ed Oc obe 25, 2018, om h ps://sen inel.esa.in /web/sen inel/use -guides/sen inel-1-sa / e isi -and- co e age ESAc, Ga i, A., Naud, C., Cas ellani, C., & Ca ie o, F. (2018). Sen inel-2 P oduc s Speci ica ion Documen . Thales Alenia Space, 1–487. Fassnach , F. E., Ha ig, F., La i, H., Be ge , C., He nández, J., Co alán, P., & Koch, B. (2014). Impo ance o sample size , da a ype and p edic ion me hod o emo e sensing-based es ima ions o abo eg ound o es biomass, 154, 102–114. h ps://doi.o g/10.1016/j. se.2014.07.028 FMP. (2018). Noyembe yan Fo es En e p ise: Fo es Managemen Plan-2018. 61 Fo es Code, A. (2005). Fo es Code o The Republic o A menia, 1–22. G.F. De G andi, M. Leysen, J.S. Lee, D. S. (2004). Rada Re lec i i y Es ima ion Using Mul iple SAR Scenes o he Same Ta ge : Technique and Applica ions., (l), 1047–1050. GEE API Tu o ial. (2019). API u o ials | Google Ea h Engine API. Re ie ed Janua y 5, 2019, om h ps://de elope s.google.com/ea h-engine/ u o ials Ghaza yan, G., Dubo yk, O., Löw, F., La eniuk, M., Schellbe g, J., & Kussul, N. (2018). A ule-based app oach o c op iden i ica ion using mul i- empo al and mul i-senso phenological me ics. Eu opean Jou nal o Remo e Sensing, 51(1), 511–524. h ps://doi.o g/10.1080/22797254.2018.1455540 Gibbs, H. K., B own, S., Niles, J. O., & Foley, J. A. (2007). Moni o ing and es ima ing opical o es ca bon s ocks : making REDD a eali y. h ps://doi.o g/10.1088/1748-9326/2/4/045023 Global Fo es Wa ch. (n.d.). Re ie ed Janua y 20, 2019, om h p://bi .ly/2R0Yg8 Goe z, S., & Dubayah, R. (2011). Ad ances in emo e sensing echnology and implica ions o measu ing and moni o ing o es ca bon s ocks and change. Ca bon Managemen , 2(3), 231–244. h ps://doi.o g/10.4155/cm .11.18 Haywood A. and Sayadyan H. and Ho hannisyan N. Implemen ing a simple design- based na ional o es ca bon in en o y in A menia (Regula pape in p epa a ion).2018. Huang, X., Zini i, B., To bick, N., & Ducey, M. J. (2018). Assessmen o Fo es abo e G ound Biomass Es ima ion Using Mul i-Tempo al C-band Sen inel-1 and Pola ime ic L-band PALSAR-2 Da a. h ps://doi.o g/10.3390/ s10091424 JAXA. (2018). Global 25 m Resolu ion PALSAR-2 / PALSAR Mosaic and Fo es / Non-Fo es Map ( FNF ): Da ase Desc ip ion. Joshi, N., Baumann, M., Ehamme , A., Fenshol , R., G ogan, K., Hos e , P., … Ryan, C. M. (2016). A Re iew o he Applica ion o Op ical and Rada Remo e Sensing Da a Fusion o Land Use Mapping and Moni o ing, 1–23. h ps://doi.o g/10.3390/ s8010070 62 Kachamba, D. J., Ø ka, H. O., Gobakken, T., Eid, T., & Mwase, W. (2016). Biomass es ima ion using 3D da a om unmanned ae ial ehicle image y in a opical woodland. Remo e Sensing, 8(11). h ps://doi.o g/10.3390/ s8110968 Kuma , K. K., Nagai, M., Wi ayangku n, A., K i iyu anan , K., & Nakamu a, S. (2016). Abo e G ound Biomass Assessmen om Combined Op ical and SAR Remo e Sensing Da a in Su a Thani P o ince, Thailand. Jou nal o Geog aphic In o ma ion Sys em, 08(04), 506–516. h ps://doi.o g/10.4236/jgis.2016.84042 Kuma , L., Sinha, P., Taylo , S., & Alqu ashi, A. F. (2015). Re iew o he use o emo e sensing o biomass es ima ion o suppo enewable ene gy gene a ion. Jou nal o Applied Remo e Sensing, 9(1), 097696. h ps://doi.o g/10.1117/1.JRS.9.097696 Lau in, G. V., Lau in, G. V., Balling, J., Co ona, P., Ma ioli, W., Papale, D., & Pule i, N. (2018). Abo e-g ound biomass p edic ion by Sen inel-1 mul i empo al da a in cen al I aly wi h in eg a ion o ALOS2 and Sen inel-2 da a, 12(1). h ps://doi.o g/10.1117/1.JRS.12 Lau in, G. V., Pi o i, F., Callega i, M., Chen, Q., Cuozzo, G., Lingua, E., … Papale, D. (2016). Po en ial o ALOS2 and NDVI o Es ima e Fo es Abo e-G ound Biomass, and Compa ison wi h Lida -De i ed Es ima es. h ps://doi.o g/10.3390/ s9010018 Liu, Z., Li, P., & Yang, J. (2017). Soil Mois u e Re ie al and Spa io empo al Pa e n Analysis Using Sen inel-1 Da a o Dah a, Senegal. Remo e Sensing, 9(11), 1–18. h ps://doi.o g/10.3390/ s9111197 Me moz, S., Réjou-Méchain, M., Villa d, L., Le Toan, T., Rossi, V., & Gou le - Fleu y, S. (2015). Dec ease o L-band SAR backsca e wi h biomass o dense o es s. Remo e Sensing o En i onmen , 159, 307–317. h ps://doi.o g/10.1016/j. se.2014.12.019 MONGABAY. (2019). Fo es da a: A menia De o es a ion Ra es and Rela ed Fo es y Figu es. Re ie ed Janua y 7, 2019, om h ps:// ain o es s.mongabay.com/de o es a ion/a chi e/A menia.h m? bclid=IwA R0aRkzhqwoCCnb5FZocBp_MumAi2b5hcqH3U4-8KIa8k9jsB_YTXRyogJ0 63 Mo eno-Sanchez, R., Sayadyan, H., S ee e , R., & Rozelle, J. (2007). The A menian o es s : h ea s o conse a ion and needs o sus ainable managemen , (Augus ). h ps://doi.o g/10.2495/ECO070121 Na a o, A., Rolim, J., Miguel, I., Ca alão, J., Sil a, J., Painho, M., & Veke dy, Z. (2016). C op moni o ing based on SPOT-5 Take-5 and sen inel-1A da a o he es ima ion o c op wa e equi emen s. Remo e Sensing, 8(6). h ps://doi.o g/10.3390/ s8060525 Pe iasamy, S. (2018). Remo e Sensing o En i onmen : Signi icance o dual pola ime ic syn he ic ape u e ada in biomass e ie al : An a emp on Sen inel-1. Remo e Sensing o En i onmen , 217(Augus ), 537–549. h ps://doi.o g/10.1016/j. se.2018.09.003 Pohl C., G. J. (2017). Remo e sensing image usion: A p ac ical guide. h ps://doi.o g/10.1117/12.226824 Po in, P., Rosich, B., Mi anda, N., & G imon , P. (2016). Sen inel-1 Mission S a us. P ocedia - P ocedia Compu e Science, 100, 1297–1304. h ps://doi.o g/10.1016/j.p ocs.2016.09.245 Po ie E, Lee, J., & Fe o-Famil, L. (2004). ADVANCED CONCEPTS IN POLARIMETRY – PART 1 (Pola ime ic Ta ge Desc ip ion, Speckle il e ing and Decomposi ion Theo ems). Na o O an, 1(1), 1–29. h ps://doi.o g/10.1.1.214.8632 R Co e Team. (2018). R: A Language and En i onmen o S a is ical Compu ing. Re ie ed Feb ua y 22, 2019, om: h ps://www. -p ojec .o g/ Reiche, J., Hamunyela, E., Ve bessel , J., Hoekman, D., & He old, M. (2017). Remo e Sensing o En i onmen Imp o ing nea - eal ime de o es a ion moni o ing in opical d y o es s by combining dense Sen inel-1 ime se ies wi h Landsa and ALOS-2 PALSAR-2. Remo e Sensing o En i onmen , 204(Ap il), 0–1. h ps://doi.o g/10.1016/j. se.2017.10.034 Reiche, J., Lucas, R., Mi chell, A. L., Ve bessel , J., Hoekman, D. H., Haa pain ne , J., … He old, M. (2016). Combining sa elli e da a o be e opical o es moni o ing. Na u e Publishing G oup, 6(2), 120–122. h ps://doi.o g/10.1038/nclima e2919 69