The contribution of Earth observation technologies to the reporting obligations of the Habitats Directive and Natura 2000 network in a protected wetland

Submitted 22 January 2018 Accepted 6 March 2018 Published 21 March 2018 Corresponding author Adrián Regos, [email protected] Academic editor Louise Willemen Additional Information and Declarations can be found on page 14 DOI 10.7717/peerj.4540 Copyright 2018 Regos and Domínguez Distributed under Creative Commons CC-BY 4.0 OPEN ACCESS The contribution of Earth observation technologies to the reporting obligations of the Habitats Directive and Natura 2000 network in a protected wetland Adrián Regos1,2and Jesús Domínguez1 1Departamento de Zooloxía, Xenética e Antropoloxía Física, Universidade de Santiago de Compostela, Santiago de Compostela, Spain 2Predictive Ecology Group, Centro de Investigacão em Biodiversidade e Recursos Genéticos da Universidade do Porto, CIBIO/InBIO, Vairão, Portugal ABSTRACT Background. Wetlands are highly productive systems that supply a host of ecosystem services and benefits. Nonetheless, wetlands have been drained and filled to provide sites for building houses and roads and for establishing farmland, with an estimated worldwide loss of 64–71% of wetland systems since 1900. In Europe, the Natura 2000 network is the cornerstone of current conservation strategies. Every six years, Member States must report on implementation of the European Habitats Directive. The present study aims to illustrate how Earth observation (EO) technologies can contribute to the reporting obligations of the Habitats Directive and Natura 2000 network in relation to wetland ecosystems. Methods. We analysed the habitat changes that occurred in a protected wetland (in NW Spain), 13 years after its designation as Natura 2000 site (i.e., between 2003 and 2016). For this purpose, we analysed optical multispectral bands and water-related and vegetation indices derived from data acquired by Landsat 7 TM, ETM+and Landsat 8 OLI sensors. To quantify the uncertainty arising from the algorithm used in the classification procedure and its impact on the change analysis, we compared the habitat change estimates obtained using 10 different classification algorithms and two ensemble classification approaches (majority and weighted vote). Results. The habitat maps derived from the ensemble approaches showed an overall accuracy of 94% for the 2003 data (Kappa index of 0.93) and of 95% for the 2016 data (Kappa index of 0.94). The change analysis revealed important temporal dynamics between 2003 and 2016 for the habitat classes identified in the study area. However, these changes depended on the classification algorithm used. The habitat maps obtained from the two ensemble classification approaches showed a reduction in habitat classes dominated by salt marshes and meadows (24.6–26.5%), natural and semi-natural grasslands (25.9–26.5%) or sand dunes (20.7–20.9%) and an increase in forest (31– 34%) and reed bed (60.7–67.2%) in the study area. Discussion. This study illustrates how EO–based approaches might be particularly useful to help (1) managers to reach decisions in relation to conservation, (2) Member States to comply with the requirements of the European Habitats Directive (92/43/EEC), and (3) the European Commission to monitor the conservation status of the natural habitat types of community interest listed in Annex I of the Directive. How to cite this article Regos and Domínguez (2018), The contribution of Earth observation technologies to the reporting obligations of the Habitats Directive and Natura 2000 network in a protected wetland. PeerJ 6:e4540; DOI 10.7717/peerj.4540 Nonetheless, the uncertainty arising from the large variety of classification methods used may prevent local managers from basing their decisions on EO data. Our results shed light on how different classification algorithms may provide very different quantitative estimates, especially for water-dependent habitats. Our findings confirm the need to account for this uncertainty by applying ensemble classification approaches, which improve the accuracy and stability of remote sensing image classification. Subjects Conservation Biology, Natural Resource Management, Environmental Impacts, Spatial and Geographic Information Science Keywords Environmental monitoring, Habitat mapping, Wetland conservation, Remote sensing, Supervised classification, Landsat satellite imagery, Water-related indices, Conservation European directives, Ensemble classification approach, Protected areas INTRODUCTION Wetlands are highly productive systems that provide a host of ecosystem services and benefits, including local climate regulation, erosion control, recreational fishing, flood control and long-term supply of good quality ground water, storage of pollutants, rare species habitat, and cultural heritage and educational value (De Groot et al., 2006;Horwitz & Finlayson, 2011). Nonetheless, wetlands have been perceived as a source of vectors of waterborne infectious diseases, and historically considered worthless and an impediment to development. Consequently, wetlands have been drained and filled to provide sites for building houses and roads or for establishing farmland, with an estimated worldwide loss of 64–71% of wetland systems since 1900 (Davidson, 2014). Protection for wetlands can come in many forms, ranging from local practices and national legislation to international recognition through inscription on the Ramsar List and/or the World Heritage List (Thorsell, Levy & Sigaty, 1997). In Europe, the Natura 2000 network is the cornerstone of current environmental conservation strategies. This network includes Special Protection Areas for wild birds (SPAs), designated by the Member States under the Birds Directive (2009/147/EC) with the aim of conserving the habitats of particularly threatened species and migratory species. It also includes Special Areas of Conservation (SACs), designated for other taxa and habitats under the Habitats Directive (92/43/ EEC). Every six years, Member States must report on implementation of the measures taken under these European Directives. This report must include information on the conservation measures concerning the natural habitat types listed in Annex I of the Habitats Directive (Art. 6), as well as evaluation of the impacts and surveillance (Art. 2) of those measures in relation to their conservation status, with particular regard to priority natural habitat types and priority species. Earth observation (EO) technologies have made significant contributions to nature conservation in the last few decades (Muchoney, 2008;O’Connor et al., 2015 and reference therein). Increasingly large amounts of geospatial information are being provided by satellite and aerial image processing and analysis—also known as remote sensing (RS)—which has enormous potential for conservation applications (Leyequien et al., 2007;Alcaraz-Segura et al., 2009;Petrou, Manakos & Stathaki, 2015;Skidmore et al., 2015; Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 2/19 Adamo et al., 2016, among others). Access to EO data has improved greatly in recent years, and many aerial and satellite data are now freely available (Turner et al., 2015). Despite the above-mentioned progress, the lack of a single, unifying habitat feature as well as the highly dynamic nature of wetlands (which may lead to highly variable spectral signatures) and their steep environmental gradients (which often produce narrow ecotone areas) may constrain and overwhelm the capacity of current remote sensors (Gallant, 2015). Recent advances in computing and the development of image classification techniques have made RS-based land-cover mapping easier, faster and more widely available for use in both conservation and applied ecology (Khatami, Mountrakis & Stehman, 2016). Faced with this wide range of techniques, many researchers have focused on comparing the image classification performance of land-cover mapping or other applications (e.g., Hubert-Moy et al., 2001;Cracknell & Reading, 2014;Regos et al., 2015). One effective solution for dealing with the uncertainty arising from the use of a wide range of techniques is to generate a classification ensemble by combining some individual classifiers. This is referred to as a multiple classification system or ensemble classification approach (for a review, see Du et al., 2012). The ensemble classification approach, recently applied by the remote sensing community, is viewed as an effective way of improving the classification performance of remotely sensed imagery (Briem, Benediktsson & Sveinsson, 2002;Lu & Weng, 2007). The main goal of the present work is to illustrate how EO technologies may contribute to the reporting obligations of the Habitats Directive and Natura 2000 network regarding wetland ecosystems. We analysed the habitat changes that have taken place in a protected wetland (in NW Spain), 13 years after its designation as Natura 2000 site (2003–2016). For this purpose, we analysed optical multispectral bands and water-related and vegetation indices derived from data captured by Landsat 7 TM, ETM+and Landsat 8 OLI sensors. To quantify the uncertainty arising from the algorithm used in the classification procedure and its impact on the change analysis, we compared the habitat change estimates obtained using 10 different classification algorithms and two ensemble classification approaches. MATERIAL & METHODS Study site The study area is a coastal wetland included in the Natura 2000 network in 2003 and designated as Special Area of Conservation (SAC) and Special Protection Area (SPA) for wild birds. The site covers an area of 984 ha, corresponding to the boundaries of the ‘‘Dunas de Corrubedo e lagoas de Carregal e Vixán’’ Natural Park (Fig. 1). The international importance of the wetland was recognised when it was designated a Ramsar site, in 1993. This wetland includes one of the largest dune systems in the NW Iberian Peninsula, with extensive stretches of sand (Ladeira, Ferreira and Vilar beaches) flanked by large dune and coastal lagoon ecosystems (Lagunas de Carregal and Vixán), together with an adjacent dune system, and an embryonic shifting dune (1-km long, 200–250 m wide and 12–15 m high) (Vázquez-Paz & Pérez-Alberti, 2002). The dune system, comprising a sandy barrier, has favoured the creation of an interior sedimentary area composed of fixed dunes (‘grey dunes’), marshes, sandy and muddy intertidal zones, as well as two coastal lagoons Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 3/19 Figure 1 Location of the study area and protected-area systems. Ramsar wetland (dashed-dotted line), Natural park and SAC (black dashed line) and SPA (filling lines). Full-size DOI: 10.7717/peerj.4540/fig-1 with very different aquatic characteristics: (1) the Carregal lagoon covers an irregular, delimited space between the marsh and the dune system. The area adjacent to the coastline corresponds morphologically to an estuarine channel covered by sandy deposits where flooding depends on the tidal cycle (Fig. 1); (2) the Vixán lagoon, located in the area distal to the coastline, has a dense reed bed (Phragmites australis) that occupies most of the eulittoral and supralittoral environments (Fig. 1). In the area adjacent to the coast, the reed bed is replaced by bulrushes (Typha latipholia) and, to a lesser extent, by wet grasslands. The drainage channel zigzags through the dune system until reaching the beach (Ramil-Rego, 2007). Pre-processing EO data We used satellite remote sensing imagery to map and monitor the habitat changes that have taken place between 2003 and 2016. We analysed optical multispectral bands (Path/Row: 205/30) derived from four cloud-free images acquired by NASA’s Landsat missions on 20 March (Landsat 7 ETM+) and 6 October 2003 (Landsat 5 TM) and on 2 May and 23 September 2016 (Landsat 8 OLI) (detailed information available for each band is available at: http://landsat.usgs.gov/band_designations_landsat_satellites.php). Landsat scenes captured in spring and autumn (e.g., in May and September) were analysed to take into account seasonal differences in vegetation phenology (e.g., common reed grass). The images are all available free of charge from the US Geological Survey (USGS) Centre for Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 4/19 Earth Resources Observation and Science (EROS) and were obtained by direct download from the GloVis facility (http://glovis.usgs.gov). All downloaded images were L1T (a processing level that includes a geometric correction performed with ground control points and the use of a digital elevation model) and projected in the UTM coordinate system (WGS 84 datum, UTM projection, Zone 29 North). Digital numbers (DNs) were converted to top-of-atmosphere radiance and physically meaningful units by radiometric calibration and application of sensor- and bandspecific calibration parameters. The classification process was based on the radiometric information obtained from reflective bands and two multispectral indices for each image: (1) the Normalized Difference Vegetation Index (NDVI; Rouse et al., 1974) and (2) the Normalized Difference Water Index (NDWI; Gao, 1996). This procedure enhanced the spectral separability of vegetation associated with aquatic and halophilic environments. Classification procedure Supervised classification of the remotely-sensed data was carried out using the following 10 classification algorithms available in the R-based package Caret and implemented in the RStoolbox package, version 0.1.5 (Kuhn, 2017;Leutner & Horning, 2017): amdai (Adaptive Mixture Discriminant Analysis), avNNet (Model Averaged Neural Network), gbm (Stochastic Gradient Boosting), knn (k-Nearest Neighbours), mda (Mixture Discriminant Analysis), pls (Partial Least Squares), rf (Random Forest), svmPoly (Support Vector Machines with Polynomial Kernel), svmLinear (Support Vector Machines with Linear Kernel) and svmRadial (Support Vector Machines with Radial Basis Function Kernel). In addition, two ensemble procedures were performed: (1) a simple voting system (‘Ens_SV’; the so-called ‘majority voting’ and ‘select all majority’ system, sensu Bauer et al., 1999), considering each habitat map as an equally weighted vote; and (2) a weighted voting approach (‘Ens_WV’), using overall accuracy obtained by individual classifiers as weights (Du et al., 2012). Eight habitat classes, defined as areas with common ecological and biophysical characteristics and, therefore, with a homogeneous spectral signature, were identified in the study area. For these habitat classes, we adopted the terminology used in the Annex I of the Habitats Directive. These eight habitat classes correspond with 23 specific habitats listed in this Annex I in our study area (Table 1). The study area is very well described, and the whole list of habitats is already defined in previous reports (see e.g., Ramil-Rego et al., 2008). Training and validation areas for each habitat class were established by on-screen digitizing in QGIS software, and consisted of a set of pixels identified over well-known homogeneous areas in each Landsat image, thus providing a reference spectral signature for each class. In particular, we applied a stratified random design as sampling strategy, with a total of about 259–346 training and validation areas proportionally distributed throughout the entire study area for each year (Table 2;Dataset S1). Specifically, for 2003 we used different Red-Green-Blue (RGB) composites from the Landsat bands and digital orthophotos in natural colours at a scale of 1:18,000 obtained from the Plan Nacional de Ortofotografía Aérea (PNOA) for 2004, while for 2016 we used digital orthophotos from 2014. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 5/19 Table 1 List of broad habitat classes used in the change analysis and their correspondence with the natural habitats (and codes) listed in the Annex I of the Habitats Directive. Asterisk indicates habitats with highest priority for conservation according to the Habitats Directive. Habitat class Natural habitats listed in the Annex I of the Habitat Directive Sand dunes 1110 Sandbanks which are slightly covered by sea water all the time. 1140 Mud flats and sandflats not covered by sea water at low tide. 1210 Annual vegetation of drift lines. 2110 Embryonic shifting dunes. 2120 Shifting dunes along the shoreline with Ammophila arenaria (‘white dunes’). Tidal areas 1130 Estuaries. 1150* Coastal lagoons. 1160 Large shallow inlets and bays. 1170 Reefs. Forest – Reedbed – Sea dunes of Atlantic coast 2130* Fixed coastal dunes with herbaceous vegetation (‘grey dunes’) 2150* Atlantic decalcified fixed dunes (Calluno-Ulicetea). 2190 Humid dune slacks. 2230 Malcolmietalia dune grasslands. 2260 Cisto-Lavenduletalia dune sclerophyllous scrubs. Natural and semi-natural grasslands 6220* Pseudo-steppe with grasses and annuals of the Thero-Brachypodietea 6410 Molinia meadows on calcareous, peaty or clayey-silt-laden soils (Molinion caeruleae). 6420 Mediterranean tall humid grasslands of the Molinio-Holoschoenion. 6430 Hydrophilous tall herb fringe communities of plains and of the montane to alpine levels. 6510 Lowland hay meadows (Alopecurus pratensis,Sanguisorba officinalis) Salt marshes and meadows 1310 Salicornia and other annuals colonizing mud and sand. 1330 Atlantic salt meadows (Glauco-Puccinellietalia maritimae). 1420 Mediterranean and thermo-Atlantic halophilous scrubs (Sarcocornetea fruticosi). Burned areas – Table 2 Total number of training and validation areas considered in the supervised classification for each habitat class and year. Habitat class Training Validation Training Validation 2003 2016 Sand dunes 38 30 35 29 Tidal areas 48 29 49 31 Forest 64 61 76 43 Reedbed 21 22 20 23 Sea dunes of Atlantic coast 46 29 56 29 Natural and semi-natural grasslands 31 55 44 39 Salt marshes and meadows 34 33 33 26 Burned areas 0 0 39 23 TOTAL 282 259 346 243 Validation procedure The accuracy of habitat maps was assessed from confusion matrices based on the number of pixels correctly (and incorrectly) classified per class, and by comparing the results obtained from different classification algorithms. The main quality parameters were the overall accuracy (%), the producer’s and user’s accuracies, and the Kappa index of agreement. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 6/19 We used McNemar’s tests to evaluate statistical significance of the difference in accuracy between each pair of algorithms. This is a non-parametric test that is based on confusion matrices collapsed to two by two contingency tables (Foody, 2004;De Leeuw et al., 2006). P-values from McNemar’s tests were represented with heatmaps to help visualizing statistical significance of the difference between all possible comparisons. These p-values were used to support the selection of algorithms for the subsequent ensemble procedures. Thereby, classification algorithms with statistically lower accuracies were not included in the ensemble procedures (McNemar’s tests, p<0.05). Data importation, pre-processing, spectral indices, image classification and graphical display were performed using the toolset available in RStoolbox package, version 0.1.5 (Wegmann, Leutner & Dech, 2016;Leutner & Horning, 2017) (see http://rpubs.com/ ARegos/359655 for R code and formatted outputs). Change analysis We quantified the spatial extent (in ha) of each habitat class per year (2003 and 2016) from each classification algorithm and ensemble approach. Boxplots were constructed using the R package ggplot2 (Wickham, 2009). The contribution of each habitat class to the habitat change (i.e., conversion from one habitat class to another) was showed through a transition matrix obtained by cross-tabulation of the habitat maps derived from the two ensemble classification approaches. Transition matrices were computed with the R package lulcc v.1.0.2 (Moulds, 2017) (see http://rpubs.com/ARegos/359655 for R code and formatted outputs). RESULTS Accuracy assessment The habitat maps with the highest accuracy (up to 95%) in 2003 were obtained using support vector machines and discriminant analysis, with the ‘amdai’ classifier providing slightly better results (Fig. 2). For 2016, the highest accuracy was obtained by applying support vector machines with linear kernel (Fig. 2). However, McNemar’s test did not show statistical significance of the difference in accuracy between individual classification algorithms (p>0.05; Fig. 3), except for ‘pls’, ‘avNNet’ (p<0.01; Fig. 3), ‘gbm and ‘svmRadial’ (p<0.05; Fig. 3). These algorithms showed limitations for specific habitat classes that have led to under- and overestimations of their extent (Fig. 4). For instance, ‘pls’ showed very low user’s accuracies for the thematic class ‘forest’, while ‘svmRadial’ markedly overestimated the habitat class ‘tidal area’ (see low user’s accuracy and high producer’s accuracy values, i.e., low omission errors and high commission errors in Fig. 4). Thereby, ‘pls’ and ‘avNNet’ for both years and ‘gbm’ and ‘svmRadial’ for year 2016 were finally not considered during the ensemble procedures. The habitat maps derived from the ensemble approaches (majority and weighted vote) showed an overall accuracy of 94% for the 2003 data (Kappa index of 0.93) and of 95% for the 2016 data (Kappa index of 0.94) (Fig. 2) with no statistical significance of the difference between them (p>0.05). Change analysis was therefore performed using the two ensemble methods. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 7/19 Accuracy Kappa 2003 2016 2003 2016 0.6 0.7 0.8 0.9 Land cover class Year alg amdai avNNet Ens_SV Ens_WV gbm knn mda pls rf svmLinear svmPoly svmRadial year 2003 2016 Figure 2 Accuracy of habitat maps (overall accuracy and Kappa coefficient) per year and classification method. amdai (Adaptive Mixture Discriminant Analysis), avNNet (Model Averaged Neural Network), gbm (Stochastic Gradient Boosting), knn (k-Nearest Neighbours), mda (Mixture Discriminant Analysis), pls (Partial Least Squares), rf (Random Forest), svmPoly (Support Vector Machines with Polynomial Kernel), svmRadial (Support Vector Machines with Radial Basis Function Kernel), svmLinear (Support Vector Machines with Linear Kernel), simply voting (‘Ens_SV’) and weighted voting (‘Ens_WV’) ensemble approaches. The boxplots display the median, the 50% (box) and 95% (whiskers) confidence intervals. Full-size DOI: 10.7717/peerj.4540/fig-2 Change analysis The change analysis revealed important temporal dynamics between 2003 and 2016 for the habitat classes identified in the study area (Fig. 5,Table 3). However, the changes depended on the classification algorithm used (Fig. 5). For example, values for water-dependent habitat classes ranged from around 60 ha with most of the classification algorithms, to almost 7 times this value with the ‘svmRadial’ classifier, clearly indicating overestimation of this unit (Figs. 5–7). The coverage estimated for habitat class dominated by salt marshes and meadows in 2016 ranged from values close to 52 ha with the ‘svmRadial’ classifier to more than 260 ha with the ‘gbm’ classifier (Fig. 5). The habitat maps obtained using the two ensemble classification approaches show a reduction in habitat classes dominated by salt marshes and meadows (24.6–26.5%), natural and semi-natural grasslands (25.9–26.5%) or sand dunes (20.7–20.9%) and an increase in forest (31–34%) and reed bed (60.7–67.2%) in the study area (Fig. 5). In particular, the Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 8/19 1 0.68 1 1 1 1 0.45 0.22 0.1 0.22 1 0.68 0.25 0.62 0.72 0.01 0.01 0.01 0.01 0.01 0.01 0.37 0.68 1 0.62 0.1 0.13 0.01 0.68 1 1 1 0.18 0.37 0.01 1 0.01 0.01 0.01 0.01 0.1 0.05 0.01 0.01 0.01 svmPoly svmRadial svmLinear rf knn gbm pls mda amdai svmRadial svmLinear rf knn gbm pls mda amdai avNNet 0.00 0.25 0.50 0.75 1.00 McNemar Test (p−value) Year 2003 0.13 1 0.05 1 0.05 0.48 0.05 1 1 0.22 1 0.1 0.1 0.05 0.01 0.01 0.01 0.01 0.01 0.01 1 0.13 1 1 0.48 0.22 0.01 0.62 0.37 0.25 0.25 0.13 0.68 0.01 0.48 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 svmPoly svmRadial svmLinear rf knn gbm pls mda amdai svmRadial svmLinear rf knn gbm pls mda amdai avNNet 0.00 0.25 0.50 0.75 1.00 McNemar Test (p−value) Year 2016 A) B) Figure 3 P-value from McNemar’s tests for each pair of classification algorithm. amdai (Adaptive Mixture Discriminant Analysis), avNNet (Model Averaged Neural Network), gbm (Stochastic Gradient Boosting), knn (k-Nearest Neighbours), mda (Mixture Discriminant Analysis), pls (Partial Least Squares), rf (Random Forest), svmPoly (Support Vector Machines with Polynomial Kernel), svmRadial (Support Vector Machines with Radial Basis Function Kernel), svmLinear (Support Vector Machines with Linear Kernel). White colours indicate p-values lower than 0.01, blue colour intensity increases with the p-value. Full-size DOI: 10.7717/peerj.4540/fig-3 Sand dunes Tidal areas Forest Reedbed Sea dunes Grasslands Salt marshes Burned areas 2003 PrA 2016 PrA 2003 UrA 2016 UrA 0.85 0.90 0.95 1.00 0.6 0.7 0.8 0.9 1.00.4 0.6 0.8 1.0 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.6 0.8 1.0 0.00 0.25 0.50 0.75 1.00 0.7 0.8 0.9 1.0 amdai avNNet gbm knn mda pls rf svmPoly svmRadial svmLinear Ens_SV Ens_WV amdai avNNet gbm knn mda pls rf svmPoly svmRadial svmLinear Ens_SV Ens_WV amdai avNNet gbm knn mda pls rf svmPoly svmRadial svmLinear Ens_SV Ens_WV amdai avNNet gbm knn mda pls rf svmPoly svmRadial svmLinear Ens_SV Ens_WV Accuracy Classification method alg amdai avNNet Ens_SV Ens_WV gbm knn mda pls rf svmLinear svmPoly svmRadial Figure 4 Producer’s (PrA) and user’s (UrA) accuracy per year, habitat class and classification method. amdai (Adaptive Mixture Discriminant Analysis), avNNet (Model Averaged Neural Network), gbm (Stochastic Gradient Boosting), knn (k-Nearest Neighbours), mda (Mixture Discriminant Analysis), pls (Partial Least Squares), rf (Random Forest), svmPoly (Support Vector Machines with Polynomial Kernel), svmRadial (Support Vector Machines with Radial Basis Function Kernel), svmLinear (Support Vector Machines with Linear Kernel), simply voting (‘Ens_SV’) and weighted voting (‘Ens_WV’) ensemble approaches. See Table 1 for habitat classes. Full-size DOI: 10.7717/peerj.4540/fig-4 Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 9/19 Cracknell MJ, Reading AM. 2014. Geological mapping using remote sensing data: a comparison of five machine learning algorithms, their response to variations in the spatial distribution of training data and the use of explicit spatial information. Computers and Geosciences 63:22–33 DOI 10.1016/j.cageo.2013.10.008. Davidson NC. 2014. How much wetland has the world lost? Long-term and recent trends in global wetland area. Marine and Freshwater Research 65:934–941 DOI 10.1071/MF14173. Delalieux S, Somers B, Haest B, Kooistra L, Mücher CA, Vanden Borre J. 2010. Monitoring heathland habitat status using hyperspectral image classification and unmixing. In: 2nd Workshop on hyperspectral image and signal processing: evolution in remote sensing, WHISPERS 2010–workshop program. 7–10 DOI 10.1109/WHISPERS.2010.5594895. De Groot R, Stuip M, Finlayson M, Davidson N. 2006. Valuing wetlands: guidance for valuing the benefits derived from wetland ecosystem services. Ramsar technical report no 3. Ramsar Convention Secretariat, Gland. De Leeuw J, Jia H, Yang L, Liu X, Schmidt K, Skidmore AK. 2006. Comparing accuracy assessments to infer superiority of image classification methods. International Journal of Remote Sensing 27:223–232 DOI 10.1080/01431160500275762. Domínguez J, Otero M, Vidal M. 2006. Guía de las aves del Parque Natural do complexo dunar de Corrubedo e lagoas de Carregal e Vixán. Santiago de Compostela: Consellería de Medio Ambiente e Desenvolvemento Sostible. Du P, Xia J, Zhang W, Tan K, Liu Y, Liu S. 2012. Multiple classifier system for remote sensing image classification: a review. Sensors 12:4764–4792 DOI 10.3390/s120404764. Foody GM. 2004. Thematic map comparison: evaluating the statistical significance of differences in classification accuracy. Photogrammetric Engineering & Remote Sensing 70:627–633 DOI 10.14358/PERS.70.5.627. Gallant AL. 2015. The challenges of remote monitoring of wetlands. Remote Sensing 7:10938–10950 DOI 10.3390/rs70810938. Gao BC. 1996. NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment 58:257–266 DOI 10.1016/S0034-4257(96)00067-3. Gon¸ calves J, Henriques R, Alves P, Sousa-Silva R, Monteiro AT, Lomba Â, Marcos B, Honrado J. 2016. Evaluating an unmanned aerial vehicle-based approach for assessing habitat extent and condition in fine-scale early successional mountain mosaics. Applied Vegetation Science 19:132–146 DOI 10.1111/avsc.12204. Gonzalez-Martínez XI. 2014. Nuevos datos sobre xenoflora en el término municipal de Ribeira (A Coruña). Nova Acta Cientifica Compostelana 21:65–76. González-Martínez XI. 2017. Contribución al conocimiento de la flora alóctona de Galicia (NO Península Ibérica, España). Botanica Complutense 41:53–67. Horwitz P, Finlayson CM. 2011. Wetlands as settings for human health: incorporating ecosystem services and health impact assessment into water resource management. BioScience 61:678–688 DOI 10.1525/bio.2011.61.9.6. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 16/19 Hubert-Moy L, Cotonnec A, Le Du L, Chardin A, Perez P. 2001. A comparison of parametric classification procedures of remotely sensed data applied on different landscape units. Remote Sensing of Environment 75:174–187 DOI 10.1016/S0034-4257(00)00165-6. Khatami R, Mountrakis G, Stehman S. 2016. A meta-analysis of remote sensing research on supervised pixel-based land-cover image classification processes: general guidelines for practitioners and future research. Remote Sensing of Environment 177:89–100 DOI 10.1016/j.rse.2016.02.028. Kuhn M. 2017. caret: classification and regression training. R package version 6.0-76. Available at https://CRAN.R-project.org/package=caret. Kvist L, Ponnikas S, Belda EJ, Encabo I, Martínez E, Onrubia A, Hernández JM, Vera P, Neto JM, Monrós JS. 2011. Endangered subspecies of the Reed Bunting (Emberiza schoeniclus witherbyi and E. s. lusitanica) in Iberian Peninsula have different genetic structures. Journal of Ornithology 152:681–693 DOI 10.1007/s10336-011-0646-9. Lehmann JRK, Prinz T, Ziller SR, Thiele J, Heringer G, Meira-Neto JAA, Buttschardt TK. 2017. Open-source processing and analysis of aerial imagery acquired with a low-cost unmanned aerial system to support invasive plant management. Frontiers in Environmental Science 5:1–16 DOI 10.3389/fenvs.2017.00044. Leutner B, Horning N. 2017. RStoolbox: tools for remote sensing data analysis. R package version 0.1.8. Available at https://CRANR-project.org/package=RStoolbox. Leyequien E, Verrelst J, Slot M, Schaepman-Strub G, Heitkönig IMA, Skidmore A. 2007. Capturing the fugitive: applying remote sensing to terrestrial animal distribution and diversity. International Journal of Applied Earth Observation and Geoinformation 9:1–20 DOI 10.1016/j.jag.2006.08.002. Lu D, Weng Q. 2007. A survey of image classification methods and techniques for improving classification performance. International Journal of Remote Sensing 28:823–870 DOI 10.1080/01431160600746456. Marcaccio JV, Markle CE, Chow-Fraser P. 2015. Unmanned aerial vehicles produce high-resolution, seasonally-relevant imagery for classifying wetland vegetation. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 40:249–256 DOI 10.5194/isprsarchives-XL-1-W4-249-2015. Moulds S. 2017. lulcc: land use change modelling in R. R package version 1.0.2. Available at https://cran.r-project.org/package=lulcc. Muchoney DM. 2008. Earth observations for terrestrial biodiversity and ecosystems. Remote Sensing of Environment 112:1909–1911 DOI 10.1016/j.rse.2008.01.003. Ng WT, Rima P, Einzmann K, Immitzer M, Atzberger C, Eckert S. 2017. Assessing the potential of sentinel-2 and pléiades data for the detection of prosopis and vachellia spp. in Kenya. Remote Sensing 9(1):74 DOI 10.3390/rs9010074. O’Connor B, Secades C, Penner J, Sonnenschein R, Skidmore A, Burgess ND, Hutton JM. 2015. Earth observation as a tool for tracking progress towards the Aichi Biodiversity Targets. Remote Sensing in Ecology and Conservation 1:19–28 DOI 10.1002/rse2.4. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 17/19 Petrou ZI, Manakos I, Stathaki T. 2015. Remote sensing for biodiversity monitoring: a review of methods for biodiversity indicator extraction and assessment of progress towards international targets. Biodiversity and Conservation 24:2333–2363 DOI 10.1007/s10531-015-0947-z. Ramil-Rego P. 2007. Evaluación del estado de conservación de las lagunas costeras de Galicia: propuestas de conservación y uso sostenible. Santiago de Compostela: Fundación Arao. Ramil-Rego P, Rodríguez-Guitián MA, Ferreiro da Costa J, Rubinos-Román M, Gómez-Orellana Rodríguez L, De Nóvoa Fernández B, Hinojo-Sánchez BA, Martínez-Sánchez S, Cillero-Castro C, Alberto Díaz-Varela R, Rodríguez-González PM, Muñoz-Sobrino C. 2008. Os Hábitats de Interese Comunitario en Galicia. Fichas descritivas. Monografías do Ibader. Lugo: Universidade de Santiago de Compostela. Regos A, Ninyerola M, Moré G, Pons X. 2015. Linking land cover dynamics with driving forces in mountain landscape of the Northwestern Iberian Peninsula. International Journal of Applied Earth Observation and Geoinformation 38:1–14 DOI 10.1016/j.jag.2014.11.010. Romão C. 1996. Interpretation manual of European Union habitats. Version EUR 15. Brussels: European Commission. Available at http://aei.pitt.edu/33244/1/A3.pdf . Rouse JW, Hass RW, Schell JA, Deering DH, Harlan JC. 1974. Monitoring the vernal advancement and retrogradation (green wave effect) of natural vegetation. Greenbelt: NASA/GSFC. Skidmore AK, Pettorelli N, Coops NC, Geller GN, Hansen M, Lucas R, Mücher CA, O’Connor B, Paganini M, Pereira HM, Schaepman ME, Turner W, Wang T, Wegmann M. 2015. Environmental science: agree on biodiversity metrics to track from space. Nature 523:403–405 DOI 10.1038/523403a. Stellmes M, Röder A, Udelhoven T, Hill J. 2013. Mapping syndromes of land change in Spain with remote sensing time series, demographic and climatic data. Land Use Policy 30:685–702 DOI 10.1016/j.landusepol.2012.05.007. Stratoulias D, Balzter H, Sykioti O, Zlinszky A, Tóth VR. 2015. Evaluating sentinel- 2 for lakeshore habitat mapping based on airborne hyperspectral data. Sensors 15:22956–22969 DOI 10.3390/s150922956. Thorsell J, Levy R, Sigaty T. 1997. A global overview of wetland and marine protected areas on the World Heritage list. Gland: IUCN. Available at https://www.iucn.org/ content/global-overview-wetland-and-marine-protected-areas-world-heritage-list. Turner W, Rondinini C, Pettorelli N, Mora B, Leidner A, Szantoi Z, Buchanan G, Dech S, Dwyer J, Herold M, Koh LP, Leimgruber P, Taubenboeck H, Wegmann M, Wikelski M, Woodcock C. 2015. Free and open-access satellite data are key to biodiversity conservation. Biological Conservation 182:173–176 DOI 10.1016/j.biocon.2014.11.048. Valkama E, Lyytinen S, Koricheva J. 2008. The impact of reed management on wildlife: a meta-analytical review of European studies. Biological Conservation 141:364–374 DOI 10.1016/j.biocon.2007.11.006. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 18/19 Vázquez-Paz M, Pérez-Alberti A. 2002. Análisis dinámico y medioambiental de un espacio protegido en la costa atlántica gallega (Parque Natural de Corrubedo, A Coruña). Xeográfica 2:155–176. Wegmann M, Leutner B, Dech S. 2016. Remote sensing and GIS for ecologists: using open source software. Exeter: Pelagic Publishing Ltd. Wickham. 2009. ggplot2: elegant graphics for data analysis. New York: Springer-Verlag New York. Zorrilla-Miras P, Palomo I, Gómez-Baggethun E, Martín-López B, Lomas PL, Montes C. 2014. Effects of land-use change on wetland ecosystem services: a case study in the Doñana marshes (SW Spain). Landscape and Urban Planning 122:160–174 DOI 10.1016/j.landurbplan.2013.09.013. Regos and Domínguez (2018), PeerJ, DOI 10.7717/peerj.4540 19/19