Monitoring recent changes of vegetation in Fildes Peninsula (King George Island, Antarctica) through satellite imagery guided by UAV surveys

Journal Pre-proofs Monitoring recent changes of vegetation in Fildes Peninsula (King George Island, Antarctica) through satellite imagery guided by UAV surveys Vasco Miranda, Pedro Pina, Sandra Heleno, Gonçalo Vieira, Carla Mora, Carlos E.G.R. Schaefer PII: DOI: Reference:

S0048-9697(19)35287-8 https://doi.org/10.1016/j.scitotenv.2019.135295 STOTEN 135295

To appear in:

Science of the Total Environment

Received Date: Revised Date: Accepted Date:

17 July 2019 28 October 2019 29 October 2019

Please cite this article as: V. Miranda, P. Pina, S. Heleno, G. Vieira, C. Mora, C. E.G.R. Schaefer, Monitoring recent changes of vegetation in Fildes Peninsula (King George Island, Antarctica) through satellite imagery guided by UAV surveys, Science of the Total Environment (2019), doi: https://doi.org/10.1016/j.scitotenv.2019.135295

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2019 Published by Elsevier B.V.

Monitoring recent changes of vegetation in Fildes Peninsula (King George Island, Antarctica) through satellite imagery guided by UAV surveys Vasco Mirandaa, Pedro Pinaa,*,[email protected], Sandra Helenoa, Gonçalo Vieirab, Carla Morab, Carlos E.G.R. Schaeferc

aCentro

de Recursos Naturais e Ambiente, Instituto Superior Técnico (CERENA/IST), Universidade de Lisboa, 1049-001 Lisboa, Portugal

bCentro

de Estudos Geográficos, Instituto de Geografia e Ordenamento do Território (CEG/IGOT), Universidade de Lisboa, 1600-276 Lisboa, Portugal

cDepartamento

de Solos, Universidade Federal de Viçosa, MG-36571-000, Brazil

*Corresponding author.

Graphical Abstract

Highlights 1

Antarctic vegetation is organized in relatively small and sparse patches. Novel methodology tested in Fildes Peninsula, King George Island. Satellite-based mapping guided by UAV imagery and derived elevation data. Achievement of very high classification performances. Usnea and moss formations lost about 10% of their area in the period 2006-2017.

Abstract Mapping accurately vegetation surfaces in space and time in the ice-free areas of Antarctica can provide important information to quantitatively describe the evolution of their ecosystems. Spaceborne remote sensing is the adequate way to map and evaluate multitemporal changes on the Antarctic vegetation at large but its nature of occurrence, in relatively small and sparse patches, makes the identification very challenging. The inclusion of an intermediate scale of observation between ground and satellite scales, provided by Unmanned Aerial Vehicles (UAV) imagery, is of great help not only for their effective classification, but also for discriminating their main communities (lichens and mosses). Thus, this paper quantifies accurately recent changes of the vegetated areas in Fildes Peninsula (King George Island, Antarctica) through a novel methodology based on the integration of

2

multiplatform data (satellite and UAV). It consists of multiscale imagery (spatial resolution of 2 m and 2 cm) from the same period to create a robust classifier that, after intensive calibration, is adequately used in other dates, where field reference data is scarce or not available at all. The methodology is developed and tested with UAV and satellite data from 2017 showing overall accuracies of 96% and kappa equal to 0.94 with a SVM classifier. These high performances allow the extrapolation to a pair of previous dates, 2006 and 2013, when atmospherically clear very high-resolution satellite imagery are available. The classification allows verifying a loss of the total area of vegetation of 4.5% during the 11year time period under analysis, which corresponds to a 10.3% reduction for Usnea sp. and 9.8% for moss formations. Nevertheless, the breakdown analysis by time period shows a distinct behaviour for each vegetation type which are evaluated and discussed, namely for Usnea sp. whose decline is likely to be related to changing snow conditions.

Keywords: Vegetation Mapping; Change Detection; UAV; Satellite; Object-based Classification; Permafrost, Antarctica.

3

1. Introduction The vegetation associated to the active layer and the underlying permafrost (Cannone et al., 2006; Oliva et al., 2018) is a key environmental component of the terrestrial ecosystems in the Antarctic Peninsula (Guglielmin et al., 2014) and in all ice-free areas of the Antarctic continent (Lee et al., 2017). The rapid warming in the Antarctic Peninsula in the second half of the 20th century (Turner et al., 2005) and the cooling, or at least the absence of warming, in the first decade and a half of the 21st century (Turner et al., 2016; Oliva et al., 2017) associated to permafrost warming (Vieira et al., 2010; Biskaborn et al., 2019) has had a direct impact on vegetation growth/decline and its spatial distribution (Amesbury et al., 2017; Sancho et al., 2017). The increasing expansion of ice-free areas till the end of the current century (Lee et al., 2017) will provide new habits for colonization by native organisms but also very likely for invasive species (Siegert et al., 2019). This expansion in area of the terrestrial regions will be accompanied by a decrease in their number due to the coalescing of isolated areas and, possibly, in the decrease of biodiversity (Lee et al., 2017). The expansion of non-native species, which seems already to be higher in areas closer to human activity (Duffy and Lee, 2019), should also be linked to the spatial and temporal patterns of paraglacial activity (Ruíz-Fernandez et al., 2019), as to provide a more comprehensive evaluation. The potential use of biodiversity as an indicator of the effects of climate change (Green et al., 2011; Green et al., 2012; Robinson et al., 2018; Sancho et al.,

4

2019) requires the mapping and monitoring of vegetation to be not only extensive but also accurate (Cannone, 2004; Pereira et al., 2018). Spaceborne remote sensing is the only practical way to do it in the whole Antarctic Peninsula due to the repeatable extended coverage of the surfaces with multispectral imaging (Fretwell et al., 2011). The availability of satellite remotely sensed datasets in this region since the late 1970s also makes it the only possible way to create a comprehensive baseline describing the location and extent of vegetation in multitemporal evaluations. Some methodologies to classify vegetated areas in Antarctica based on remotely sensed imagery of multiple satellite platforms, mainly optical imagery but also radar, have already been developed. These supervised classifiers are methodologically diverse and, among others, are based on vegetation indexes (Fretwell et al., 2011), hierarchical procedures (Lim et al., 2012), spectral mixture analysis (Shin et al., 2014), statistical and machine learning methods (Vieira et al., 2014; Jawak et al., 2016; Schmid et al., 2017; 2018), matched filtering (Casanovas et al., 2015) or on object-based image analysis (Andrade et al., 2018). Although the spatial resolution of the images used in each situation is varied, from the higher and metric scales of more recent satellites (Ikonos, QuickBird or WorldView, for instance) to the lower 30 m/pixel of the most long-lasting satellite series (Landsat), all these methodological approaches share a common feature: the classifiers perform well only when the spectral mixing in each image pixel is low. Although the spectral signatures of the different Antarctic vegetation classes are distinct (Lovelock et al., 2002, Malenovský et al., 2015; Calviño-Cancela and Martín-Herrero, 2016), their nature of spatial occurrence, mainly constituted by relatively small and sparse patches of lichens and mosses (Convey et al., 2014), may easily provide observations in satellite imagery with undesired degrees of

5

spectral mixing with other classes (soils, rocks, water, snow, ice) and therefore being established in many locations with low levels of certainty. Although the importance of this issue decreases with the increase of spatial resolution, it is generically verified at all scales of observation. This uncertainty is obliging algorithms to be very conservative to minimize false identifications and, consequently, to incorrectly designate many vegetated areas as unvegetated (Casanovas et al., 2015). Moreover, most of these procedures are validated with data from few and specific sites with local oriented purposes, preventing robust and integrated extrapolations in space but also in time. In particular, the monitoring of the vegetation abundance and biodiversity, which is crucial for establishing plant growth rates together with associated meteorological and micro-climate data (Sancho and Pintado, 2004; Sancho et al., 2007; Li et al., 2014) could be much better performed if an intermediate scale of observation between field and satellite imagery is available. A solution to overcome this issue and consequently produce more reliable thematic maps may be provided through the incorporation in the classification procedure of intermediate scales of observation between ground and satellite levels. This integration surely allows understanding the spatial characteristics of the vegetation in a more continuous way and how it relates for instance to topography and geomorphology (Ruiz-Fernández et al., 2017) and how it changes between scales of observation in relatively extended areas. Besides the traditional aerial surveys with airplanes or helicopters, Unmanned Aerial Vehicles (UAV) are becoming the strongest alternative, due to its portability, ease of use and comparatively low cost. The ultra-high resolution up to few mm/pixel of the images captured and the fairly large coverage that they can provide (when compared to traditional field observations), permits distinguishing many details related to the vegetation not evident in satellite imagery, namely about their communities (for instance, lichen or moss) but also about their 6

spatial organization (texture, density, level of mixture with other surfaces) and the exact limits of occurrence. The incorporation of this local but very detailed information, together with the one collected at ground level (Pina et al., 2016), can greatly contribute to increase the level of detection in spaceborne imagery and produce more accurate and reliable mappings of the vegetation. Therefore, the use of small UAV in terrestrial regions of Antarctica is increasing steadily, due to the possibility of gathering data of less accessible regions with unprecedented detail, although field operations are normally complicated due to weather conditions. The application topics are varied in a multitude of objectives, namely for magnetic anomalies detection (Funaki et al., 2014), atmospheric observations (Cassano et al., 2014), glacial retreat quantification (Pudełko et al., 2018), moraine sedimentological characterization (Westoby et al., 2015), periglacial landform mapping (Dabski et al., 2018), patterned ground identification (Pina et al., 2019), wildlife inventorying and monitoring (Goebel et al., 2015; Krause et al., 2017; Mustafa et al., 2018; Pfeifer et al., 2019, Korczak-Abshire et al., 2019) and monitoring coastal sea-ice (Li et al., 2019), among others. In what concerns vegetation studies in Antarctica, UAVs are being extremely helpful in providing additional details for mapping procedures (Turner et al., 2014), in obtaining the micro-topography of moss beds (Lucieer et al., 2014), in assessing the stress (Malenovský et al., 2017) and health states of plants (Turner et al., 2018) and in their mapping in sites located some kilometres away from the operation centre (Zmarz et al., 2017). All these studies refer to diverse topics related to vegetation in relatively small areas with none being related to satellite imagery.

7

The increase of the level of detail that UAV datasets can provide in relatively large areas of the surface, when compared to traditional field observations, at a ultra-high resolution of few cm/pixel, allow the unequivocal discrimination of the main vegetation types (mosses from lichens) and their densities and textural arrangements to establish the amount and degree of mixing with background soil and rock. The main objective of this work is to detect accurately recent changes on vegetation covers in Antarctica through remote sensing which, due to its nature of occurrence in small and sparse areas, requires a novel methodological approach. The procedure is guided by the details extracted from higher resolution images (UAV) to build and calibrate a robust classification model of the lower resolution images (satellite).

2. Study Area Fildes Peninsula (62°12′S, 58°58′W), located in King George Island (South Shetlands, Maritime Antarctica), is the selected study area (Figure 1). It is one of the largest ice-free regions of the South Shetlands with about 29 km2 (Peter et al., 2008), being characterized by high biodiversity (Braun et al., 2012). Geologically, the peninsula consists of a sequence of Late Cretaceous volcanoclastic rocks (basalts, andesite to dacite), overlaid by limestones, tuffaceous conglomerates, sandstones and clays of early-mid-Eocene–Oligocene age (Birkenmajer et al., 1997; Michel et al., 2014). The relief is dominated by two high structural volcanic platforms (Meseta Sur, 167 m asl and Meseta Norte, 155 m asl), with low lying erosional platforms between and around them, mainly with altitudes below 50 m asl. The Meseta Norte, where detailed field work was performed, is a mesa-like plateau bounded by steep slopes and with a slightly depressed central area at 100–120 m asl, with sub-

8

horizontal volcanic strata giving origin to a series of plateaus and scarps (López-Martínez et al., 2012). Several small lakes occur in the interior of the meseta, which, besides scarce perennial snow patches (Mora et al., 2017), stays almost snow free by the end of the summer. The climate is polar oceanic in the Köppen scheme, characterized by a mean annual temperature of −2.3°C and a mean precipitation between 350 and 500 mm per year (Oliva et al., 2017). Wind is frequent and typically strong, with predominant directions from north, northwest and west. Cloudiness is permanent, with occasional days (or partial days) of clear skies, with a weather dominated by the continuous passage of frontal systems (Kejna et al., 2013). Vegetation communities in Fildes Peninsula are dominated by moss and lichen formations (Lindsay, 1971; Simonov, 1977; Øvstedal and Lewis-Smith, 2001): moss communities with about 40 species dominate at moist sites in slopes or in depressions, while lichens with more than 170 species are widespread in exposed, well-drained gravels, boulders and rock outcrops. The most dominant lichens are Usnea aurantiacoatra (Jacq.) Bory and Usnea antarctica Du Rietz (Li et al., 2014). The most commonly occurring moss formations are Andreaea, Bartramia patens and Polytrichastrum alpinum (Hu, 1998; Olech, 2002). The area selected for making the surveys with the UAV, due to the significant vegetation and geomorphological diversity, is mostly located within Meseta Norte but also extends along its gentle northern slope (Figure 1).

3. Remotely sensed imagery

9

The remotely sensed data used in this study consists of four recent datasets of optical imagery captured between 2006 and 2017 through satellite and UAV platforms, as detailed in Table 1. The spaceborne datasets are constituted by very high resolution (VHR) imagery covering the whole Fildes Peninsula from satellites QuickBird (QB) in 2006 and WorldView-2 (WV2) in 2011 and 2017 (Figure S1). QB contains 4 multispectral bands (MS: B-blue, G-green, R-red and NIR-near-infrared) of 2.4 m/pixel and 1 panchromatic band (PAN) of 0.6 m/pixel. WV2 imagery is constituted by 8 MS bands, the same 4 of QB satellite (B, G, R and NIR1) and 4 additional ones (C-Coastal, Y-Yellow, RE-red edge and NIR2) of 2 m/pixel and 1 PAN band with 0.5 m/pixel of spatial resolution. Although the equivalent 4 MS bands of QB and WV2 are not exactly the same, since QB spectral ranges are a bit wider than WV2, they are centred at about the same location in the electromagnetic spectrum and this way can be considered as equal. All three image scenes were acquired at about the same local time in the early afternoon but in different months, QB image in February 2006 and WV2 images in April 2013 and March 2017. The WV2 image dataset of 2013 was acquired at 27° off-nadir angle and the other two at nadir. The images are cloud free for two dates (2006 and 2013), and with a small cloud cover fraction of 14% in 2017 mostly located over uninteresting surfaces for this study, namely, the sea and the northern glaciated area of Fildes (Figure S1c). The UAV dataset is constituted by an image mosaic (orthorectified) built from 3063 individual images captured in 10 adjacent flights with a DJI Phantom 3 equipped with a RGB camera during field surveys developed between 20 and 30 January 2017, a couple of months before the acquisition of WorldView-2 image in 19 March 2017. The survey flights 10

were performed in contiguous regions in a double-grid mode at heights of 60 m above ground, with front and lateral overlapping of 70% between adjacent images. The area surveyed in each flight of about 250 m x 250 m was fixed by the duration of each individual battery in a conservative procedure, so no more than 17 minutes/flight were allowed although the batteries are said to last 23 minutes. This caution is due to the environmental conditions that can change rapidly, especially wind speeds. Thus, the aerial surveys were performed whenever the weather conditions allowed, that is, with no precipitation (rain or snow), wind speeds below 9 m/s and normally between 10:00 and 16:00 local time (GMT minus 3). Coloured markers were placed on the ground before each flight, being its precise location measured afterwards with a D-GPS system (Leica GS10). A global orthorectified mosaic with a spatial resolution of 2.1 cm/pixel (Figure 2a) and the respective Digital Elevation Model (DEM) with 8.2 cm/pixel (Figure 2b) are obtained using SIFT-Scale Invariant Features Transform and SfM-Structure from Motion based techniques (Carrivick et al., 2016). The software Agisoft Photoscan is used in this processing in a straightforward procedure. A total of 15 ground-control points was added to provide accurate georeferencing of both the image mosaic and respective DEM. Finally, the less precise irregularities along the borders of the mosaic and DEM are clipped to obtain a rectangular area of 534,296 m2, where two small portions (at bottom-centre and top-left edges) were not covered with acceptable imagery (Figures 2a and 2b).

4. Methodology A methodology was developed to obtain accurate maps of vegetation that may be used with high certainty in multitemporal change detection assessments. It is constituted by two

11

sequential phases: ‘Phase I’ consists of the calibration of a classification model of satellite imagery that greatly benefits from the integration of synchronous UAV imagery, followed by ‘Phase II’ that, based on the calibrated model, classifies satellite images of other dates where such detailed data is not available. 4.1. Phase I: Calibration of the classification model Two parallel processing lines of synchronous datasets constitute this first phase (Figure S2), one based on the UAV-derived imagery and the other on the satellite datasets, interacting with each other along three main tasks: pre-processing, segmentation and classification.

4.1.1. Pre-processing The pre-processing task consists mainly on geometric correction operations, namely the orthorectifications of satellite and UAV imagery and their co-registration. The WV2 satellite image of 2017 is orthorectified with the help of a DEM with horizontal resolution of 5 m (Figure 1c) built from the IGM-INACH project (Barriga et al., 2001), while the UAV mosaic is, as already mentioned, orthorectified with the help of the DEM of 8.2 cm (Figure 2b) obtained in the same surveys. Both datasets are co-registered with high accuracy with 15 ground-control points and a root-mean square error of 0.85 pixels. The satellite digital numbers (DN) values were converted to ToA-Top of Atmosphere reflectance values using the calibration coefficients (gains and offsets), solar irradiance and sun azimuth provided as metadata by Digital Globe. This is an important step to perform as to correct for varying illumination related to the time of acquisition of the satellite images. For the atmospheric corrections, although the low atmospheric temperatures result in low atmospheric moisture with few impacts on the imagery (Fretwell et al., 2011) some 12

preliminary tests were performed with 6S and FLAASH methods. No noticeable improvements were detected. In addition, since the supervised classification to perform in each image is solely based on spectral samples from that specific date with the purpose of identifying the presence or absence of vegetation (and to discriminate the main two types) and not some of its absolute spectral characteristics, no atmospheric corrections were performed in the satellite dataset. The most dynamic surface covers like snow, water and shadows, naturally exhibit distinct cover in the imagery in different time periods. During the summer season the snow cover can change rapidly due to melting, while the smaller and shallower lakes that may drain can exhibit marked differences on their sizes and shapes. Topographic shadows are also different according to the azimuth and elevation of the Sun at the time of acquisition of the images. Therefore, the masking of these three classes is important, so to exclude them in the multitemporal analysis of the vegetation. This procedure is separately performed on the satellite and UAV datasets, followed by the creation of a single mask resulting from their union to guarantee that the exact same region is analysed on both scales. 4.1.2. Segmentation The segmentation of the UAV mosaic consists on the precise manual delineation of the vegetation patches, whose centimetric resolution allows the clear delimitation of their limits and specifically a well-founded differentiation between mosses and lichens (Figure 3). The segmentation of the masked satellite image is performed by the watershed transform (Beucher and Lantuéjoul, 1979; Soille, 2004), a very efficient tool to compute the homogeneous regions or objects of an image. It is based on the analogy of flooding a surface from its minima and on the detection of points between adjacent basins (the 13

watershed or crest lines) where the water was about to merge (Soille, 2004). This transform is applied to the magnitude of a gradient image to obtain the location of the sharper transitions (or edges) between distinct surface classes. The magnitude of the gradient, computed with Sobel operator, is applied to the NIR band, since it is the most discriminative wavelength available in the imagery set of bands between vegetated and not vegetated surfaces. In addition, due to the existence of many patches of mosses and lichens of small dimension (few pixels at satellite scale), it was decided to not perform a filtering to smooth the image before computing the watershed nor a post-processing to merge adjacent objects. These procedures, commonly used to reduce the over-segmentation of the image, would merge many of these small adjacent areas belonging to different classes, resulting in an under-segmentation of the image as verified by Pina et al. (2016). The over-segmentation of the image is not an issue in these images since similar adjacent regions separated by the segmentation are very likely to be assigned with the same label afterwards by the subsequent classification procedure. An example of the segmentation performed is shown in Figure S3.

4.1.3. Classification The procedure designed to evaluate the classification consists on splitting the UAV area in two disjoint equal-sized regions (0.27 km2 each, excluding masked areas), one for training the other for testing (Figure 2). A meridian line divided this site into the sectors UAV-west and UAV-east since, besides containing the same class occurrences, is the only possibility for having similar reliefs on both sides, namely from the Meseta Norte and the northern slope. These disjoint regions will also permit evaluating the classification in a cross-validation procedure (train the classifier in one region and test it on the other, and vice-versa) and 14

therefore better infer about the ability of this model to generalize. This is a crucial issue, since the main objective of this methodology is to build and calibrate a robust model that can not only be extended to the whole Fildes Peninsula in imagery of the same period but also from previous dates where such detailed surface data is occasional or not available at all.

A total of 52 features per each object are extracted. They refer to five geometric properties (area, perimeter, convexity, circularity and elongation), four spectral properties (average, minimum, maximum and standard deviation) and four textural properties that describe the spatial arrangement of the intensities within each object (range, mean, variance and entropy) in each of the same four bands used in each date (R, G, B and NIR). Two robust supervised classifiers, often applied in remote sensing applications, are tested: Support Vector Machines (SVM) and k-Nearest Neighbour (kNN). SVM is a supervised kernel method (Vapnik, 1995) that can successfully handle data with unknown statistical distributions and with small training sets. The decision boundary between regions is obtained through kernel functions that map the training data into a higher-dimensional space in which the classes can be linearly separated by a hyperplane. Different transformation kernels were tested, being RBF-radial basis function the one achieving the best results. Two parameters are associated to this kernel: gamma parameter defines the radius of influence of each training sample while C controls how many examples located on the wrong side of the decision region are acceptable. kNN is a non-parametric classifier which assigns an object of unknown class to the most frequented class among the ‘k’ labelled training objects that are close to it (Altman, 1992). Closeness is defined in terms of a Euclidean distance with the input variables as the axes of feature space. Each object is 15

selected by a majority vote of its neighbours, being the object assigned to the most common of its ‘k’ nearest neighbours. The quality of each classification is evaluated through two of the most common metrics used in remote sensing, the overall accuracy (oa) and kappa index of agreement (kappa) (Congalton, 1991).

4.2. Phase II: Classifications with the calibrated model This phase is practically the same as in the satellite processing line of Phase I but using now the features of the calibrated model to perform the classifications (Figure S4). In this phase, QB and WV2 images of 2006 and 2013, respectively, are also orthorectified with the 5 m DEM of Fildes Peninsula. This procedure is especially relevant for the WV2 image of 2013 since it was acquired with an off-nadir angle of 27°. The improvements obtained with the orthorectification, followed by an accurate co-registration of each image with the WV2 image of 2017 allow obtaining aligned datasets that can be adequately used in multitemporal change detection. The masking of the same previous surfaces (snow, water and shadows) is performed in each image. The segmentation of each image is performed as before with the watershed transform on the gradient of the NIR band and with no previous filtering and postprocessing. The selection of the objects for training the classifier in each date is carefully developed since almost no contemporary ground-truth data is available. For 2013, this construction is guided by selecting objects with similar spectral properties for each class to those on 2017, together with ground-truthing points obtained during January 2012 (Vieira et al., 2014); 16

these points that are located within homogeneous surface areas of 1 m2 and were measured in the field with a D-GPS. For 2006, no contemporary ground-truthing is available, so the construction of the training samples relies mainly in selecting objects spectrally identical to those of 2017.

5. Results The details of the application of the two phases of the methodology and the results obtained are presented in the following, being the experimental procedures done using ENVI software from Harris Geospatial. 5.1 Calibration of the classification model The classification task starts by the selection of the training/testing objects on the satellite imagery for three mains classes (Usnea sp., mosses and bare soil/rock) with the help from the UAV delineated patches. Although at times different lichen formations are present in the area, the study site is very largely dominated by Usnea sp. with tiny expressions of other lichens, making us not to separate the different lichen formations into distinct classes. The selection of samples is randomly performed on both areas, attempting to build sets as balanced as possible (Table 2). Although the region UAV-west is a bit less vegetated than UAV-east and the snow masking is also a bit larger, reflected in the dimension of the samples (1058 objects on UAV-west and 1449 on UAV-east), the proportional sampling by class on each side keeps the same natural order of occurrence and can be considered adequately balanced to the classification procedures: bare class in area (in pixels) is about half of the selected samples (55% on UAV-west and 45% on UAV-east), followed by Usnea sp. with about one-third of the samples (31% on west and 38% on east) with the smaller sample area for mosses (14% and 15% on west and east respectively). 17

Different parameterizations of each model are this way tested. For SVM classifier, the combination of a pair of values is tested: γ=0.03 and 0.1, C=100 and 1000. For kNN classifier, four numbers of even neighbours (to avoid ties) are tested: 1, 3, 7 and 9. The spectral mixing at the pixel scale of WV2 imagery is high in some situations and difficult to establish, meaning that ambiguous decisions can be taken. Therefore, a confidence threshold on the classification is used and fixed at 40% for all tests, meaning that a decision is not taken for objects below that value (they remain unclassified). A total of 16 tests are performed in the UAV site in a rotating procedure: 8 classifications on the west side with training samples from the east side (4 tests with SVM and 4 tests with kNN) and vice-versa. The performances obtained in all these tests are presented in Table 3. In this cross-validation procedure, both methods perform well but SVM clearly outperforms kNN in all tests. Both methods also perform better on the UAV-west region, most likely related to the larger dimension of the training dataset of UAV-east. SVM performances are consistently high and in the range 94-97% (oa) and 0.89-0.95 (kappa) while kNN values are still very good but all into a lower range, namely, 86-91% (oa) and 0.73-0.84 (kappa). Although the highest SVM performances in each area are achieved with different parameters of the model, the selection of the optimal pair of parameters can be done without great ambiguity selecting the most balanced west-east classifications, which are γ =0.03 and C=1000 (Figure 4c). It is worth analysing in detail the confusion matrix for this best classification model for better understanding the errors between the different surfaces (Table 4), where commission errors (com.) indicate the amount of false detections and omission errors (omi.) the quantity of missing detections in that class. The classification of the two vegetation 18

classes is very clearly achieved with almost no confusion between Usnea sp. and mosses (marginal values below 1% on both regions), which somehow reflects their distinctive spectral behaviour. It should also be added that the different pigmentation of the vegetation is not an issue for the classifier, since it is the NIR wavelength the most responsible for such discrimination and not any of the visible wavelengths (R, G, or B). But although a bit of confusion exists between vegetated and unvegetated surfaces, namely within each pair mosses-bare and Usnea sp.-bare, it is always with one single percentual digit. These confusions are related to areas of less dense Usnea sp. and to small/elongated patches of mosses, where the amount of bare surface in each pixel (4 m2) predominates and therefore makes the decision of the classifier likely to be assigned to the most common cover within it (bare). Nevertheless, the average total errors for each vegetated class are low with values of around 5% for Usnea sp. and 6% for mosses.

5.2 Classifications with the calibrated model The classification of 2006 and 2013 images is performed with the best configuration of the SVM classifier selected in the 2017 procedure of Phase I, that is, with RBF kernel of parameters γ=0.03 and C=1000, following the same masking procedure as before for snow, water and shadows. The respective thematic maps are shown in Figure 4a and Figure 4b. At a first glance, it can be observed that the spatial coherence of the different surfaces in the three maps is high, being the multitemporal differences shown at small scale areas. 5.3 Multitemporal change detection The area classified is the same in each year, which corresponds to 385,412 m2 or 72% of the whole study site, as the other 28% (148,884 m2) are masked for the surfaces of snow, water

19

and shadows. The classification in each of the three dates provided the distribution by class shown in Table 5, which indicates that bare surfaces dominate in each date while for the vegetation surfaces the proportion of Usnea sp. to moss formations is about 3:1. The main conclusion to derive from the multitemporal classifications (Table 5) is the loss of 4.5% of vegetated areas during this evaluation period of 11 years, from 44.3% to 39.8% of the whole area. This decrease is mostly explained by the decrease of Usnea sp. (3.5%), almost entirely in the period 2006-2013 with little change between 2013-2017. The contribution of moss formations to this reduction numbers is smaller (1%): an increase of 0.9% between 2006-2013 and a decrease of 1.9% between 2013-2017. The area variation by surface type relatively to the beginning of the evaluation period is also analysed (Table 6): Usnea sp. loses a total of 10.3%, however almost entirely in the period 2006-2013 (10.1%) being almost stable over the second evaluation period (small decrease of 0.2%), while the distinct behaviour of moss formations show first an increase in area of 9.2% (2006-2013) and then a large loss of 17.4% of surface area in the more recent period (20132017). Overall, this translates to an increase of 8.1% of bare surfaces during this 11-year period (2006-2017) faster during the first interval (4.5% in 2006-2013) than in the most recent one (3.4% in 2013-2017). The classified images allow the identification of changes per individual pixel at different times and between classes. The exchanges between each pair of surfaces are computed and presented in Table 7 as a percentage of the total area of the study site. The geographic location of the changes can be also observed in Figure 5. First, it can be derived that 79.3% of the total area did not change during the 11-year period (78.4% during the first period and 81.4% in the second one). Second, the more relevant exchanges between surfaces are those 20

between Usnea sp. and bare (with values between 2.7% and 9.0%) and mosses and bare (from 1.9% to 3.8%), since the ones occurring between Usnea sp. and mosses have a low expression (from 0.2% to 0.6%).

6. Discussion It is important to be aware that although the classification procedure is robust and achieves high performances, it contains some uncertainty related to each surface class. This issue, already evaluated in detail in the design and calibration of the classifier, can be analysed with more detail by taking into consideration the geometry of the areas that changed between two consecutive dates. The maps of multitemporal changes (Figure 5), that show a very high spatial coherence between dates, reinforce the robustness of the classification procedure. Most of the core area of each surface type is kept unchanged with the modifications occurring at their edges in the transitions between surface types and also in isolated and small groups of pixels. Therefore, two main types of geometries can be considered (Figure S5): compact, and thin and elongated shapes. Isolated and small groups of pixels are included in this last type of geometry. For instance, it can be assumed that the shapes with a more compact geometry, like those related to the losses of area of Usnea sp. in the period 2006-2013 (Figure S5a) or the area of mosses in the period 2013-2017 (Figure S5b), have a lower uncertainty, and therefore be associated to true changes. On the contrary, when the changes correspond to thin and elongated shapes, the uncertainty of changes is likely to be higher and therefore more prone to errors. For instance, the changes detected in a small area in 2013-2017 (Figure S5c) 21

made along a thin transition between classes can show both exchanges: Usnea sp. to bare (light yellow) and bare to Usnea sp. (brown). The areas of Usnea sp. in 2006 that turned to bare ground in 2013 are mostly located on the lower limits of Usnea sp. patches that prevail in terrain convexities, as shown in Figure 6. The absence of warming in the Antarctic Peninsula since the beginning of the 21st century (Turner et al., 2016) has induced the prevalence of snow for longer periods of time during the summer, as shown in the nearby Byers Peninsula in Livingston Island (de Pablo et al., 2017) and Deception Island (Ramos et al., 2017). Consequently, snow patches have increased their dimension in area (and in volume) during summertime, which seems to be responsible for the mortality of Usnea sp. and the respective decrease in area, as shown in Livingston Island (Sancho et al., 2017). The origin of the multitemporal variations of the areas of mosses is more difficult to detect with only the remotely sensed data used in this study.One may speculate that is related to the longer snow cover duration during the summer, possibly giving origin to a longer melt water availability and ponding of the moss areas. But that is something requiring additional analysis with other environmental variables. Finally, the changes detected in the vegetated surfaces in the study site, which may derive directly from recent variations in the weather conditions like the longer prevalence of snow cover in summer, are clear. Nevertheless, although the vegetation types and distribution can be considered representative of the whole Fildes Peninsula (Schmitt et al., 2017; Andrade et al., 2108) and, very likely, of the neighbouring terrestrial regions of the South Shetlands (Peat et al., 2007), they should be used carefully when discussing regional climate impacts. The surveyed area of about 0.5 km2 is totally adequate for setting the current 22

methodology but relatively small for extrapolations, being needed to be enlarged and diversified in order to include further controls (e.g. topography, lithology, etc.), together with the acquisition of additional data before any extended analysis can be performed.

7. Conclusions The inclusion of an intermediate scale of observation between ground and satellite is a major contribution for guiding the supervised decision procedure and consequently obtaining robust classification performances of the vegetation in spaceborne imagery of an ice-free area in Fildes Peninsula. The centimetric resolution provided by the UAV-based products, unveiling many details in the vegetation organization and in their clear discrimination between lichen and moss formations, is also a point to highlight. It plays a key role to calibrate a classification model that is able to achieve high performances in synchronous UAV and VHR satellite datasets and therefore capable to be applied in previous dates where ground-truth data is scarce or not available. The ultra-high-resolution image mosaic and digital elevation model provided by the UAV surveys allowed differential detection of changes in vegetation during two distinct periods in about one-decade period with a detail seldom seen in the literature and the possible cause for the decline of Usnea sp., which seem to be related to the changing snow conditions. In addition, all these detailed datasets constitute a baseline for forthcoming evaluations about vegetation surfaces behaviours. For future work, it is envisaged to apply the methodology to the whole Fildes Peninsula and to several other terrestrial regions of the Antarctic Peninsula, collecting more ground-truth data in UAV-based surveys. These surveys should include technical enhancements on 23

equipment, namely a multispectral camera to improve the accuracy on vegetation detections, and operational improvements, like a Real Time Kinematic navigation to accelerate the development of field work. In addition, since current classifications of vegetation are being developed at the 2 m resolution of WorldView and QuickBird satellites, it is also intended to evaluate the classification of vegetation surfaces with the change of scale, namely to those of Sentinel-2 (10 m) and Landsat (30 m) datasets, whose availability allows coverage of much larger terrestrial regions of Antarctica over longer timescales. Acknowledgments The research presented in this paper was funded by the Portuguese Polar Program (PROPOLAR) through research campaign projects CIRCLAR in 2016-2017 and SNOWCHANGE in 2011-2012 and the pluriannual funding projects of CERENA (UID/EEA/50009/2019) and CEG (UID/GEO/00295/2019) funded by FCT-Fundação para a Ciência e a Tecnologia. The support at Escudero station and field logistics in Fildes Peninsula provided by Instituto Antártico Chileno (INACH), and the logistical support by the Portuguese Polar Program (PROPOLAR) are warmly thanked.

24

References Altman, N.S., 1992. An introduction to kernel and nearest-neighbor nonparametric regression. The American Statistician, 46, 175–185. Amesbury, M.J., Roland, T.P., Royles, J., Hodgson, D.A., Convey, P., Griffiths, H., Charman, D.J., 2017. Widespread biological response to rapid warming on the Antarctic Peninsula. Current Biology, 27, 1616–1622. Andrade, A.M., Michel, R.F.M., Bremer, U.F., Schaefer, C.E.G.R., Simões, J.C., 2018. Relationship between solar radiation and surface distribution of vegetation in Fildes Peninsula and Ardley Island, Maritime Antarctica. International Journal of Remote Sensing, 39, 2238-2254. Barriga, R., Montero, J., Villanueva, V., Klotz, J, Bevis, M., 2001. Geodesy and digital cartographic survey in Fildes Peninsula, Rey Jorge Island, Antarctica. Geo-spatial Information Science, 4, 25-31. Beucher, S., Lantuéjoul, C., 1979. Use of watersheds in contour detection. International Workshop on Image Processing, Real-Time Edge and Motion Detection, 2.1–2.12, Rennes. Birkenmajer, K., 1997. Geology of the northern of King George Island, South Shetland Islands (West Antarctica). In: Birkenmajer, K. (Ed.), Geological Results of the Polish Antarctic Expeditions. Studia Geologica Polonica, 110, 7–26. 25

Biskaborn, B.K., Smith, S.L., Noetzli, J., Matthes, H., Vieira, G., Streletskiy, D.A., Schoeneich, P., Romanovsky, V.E., Lewkowicz, A.G., Abramov, A., Allard, M., Boike, J., Cable, W.L., Christiansen, H.H., Delaloye, R., Diekmann, B., Drozdov, D., Etzelmüller, B., Grosse, G., Guglielmin, M., Ingeman-Nielsen, T., Isaksen, K., Ishikawa, M., Johansson, M., Johannsson, K., Joo, A., Kaverin, D., Kholodov, A., Konstantinov, P., Kröger, T., Lambiel, C., Lanckman, J.-P., Luo, D., Malkova, G., Meiklejohn, I., Moskalenko, N., Oliva, M., Phillips, M., Ramos, M., Sannel, A.B.K., Sergeev, D., Seybold, C., Skryabin, P., Vasiliev, A., Wu, Q., Yoshikawa, K., Zheleznyak, M., Lantuit, H., 2019. Permafrost is warming at a global scale. Nature Communications, 10, 264. Braun, C., Mustafa, O., Nordt, A., Pfeiffer, S., Peter, H.-U., 2012. Environmental monitoring and management proposals for the Fildes Region, King George Island, Antarctica. Polar Research, 31, 18206. Calviño-Cancela, M., Martín-Herrero, J., 2016. Spectral discrimination of vegetation classes in ice-free areas of Antarctica. Remote Sensing, 8, 856. Cannone, N., 2004. Minimum area assessment and different sampling approaches for the study of vegetation communities in Antarctica. Antarctic Science, 16, 157–164. Cannone, N., Ellis Evans, J.C., Strachan, R., Guglielmin, M., 2006. Interactions between climate, vegetation and the active layer in soils at two Maritime Antarctic sites. Antarctic Science, 18, 323–333. Carrivick, J.L, Smith, M.W., Quincey, D.J., 2016. Structure from Motion in the Geosciences. Wiley-Blackwell, Chichester (208 pp.).

26

Casanovas, P., Black, M., Fretwell, P., Convey P., 2015. Mapping lichen distribution on the Antarctic Peninsula using remote sensing, lichen spectra and photographic documentation by citizen scientists. Polar Research, 34, 25633. Cassano, J.J., 2014. Observations of atmospheric boundary layer temperature profiles with a small unmanned aerial vehicle. Antarctic Science, 26, 205-213. Congalton, R.G. 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment, 37, 35–46. Convey, P., Chown, S.L., Clarke, A., Barnes, D.K.A., Bokhorst, S., Cummings, V., Ducklow, H.W., Frati, F., Green, T.G.A., Gordon, S., Griffiths, H.J., Howard-Williams, C., Huiskes, A.H. L., Laybourn-Parry, J., Berry Lyons, W., McMinn, A., Morley, S.A., Peck, L.S., Quesada, A., Robinson, S.A., Schiaparelli, S., Wall, D.H., 2014. The spatial structure of Antarctic biodiversity. Ecological Monographs, 84, 203–244. Dabski, M., Zmarz, A., Pabjanek, P., Korczak-Abshire, M., Karsznia, I., Chwedorzewska, K.J., 2018. UAV-based detection and spatial analyses of periglacial landforms on Demay Point (King George Island, South Shetland Islands, Antarctica). Geomorphology, 290, 29-38. de Pablo, M.A., Ramos, M., Molina, A., 2017. Snow cover evolution, on 2009-2014, at the Limnopolar Lake CALM-S site on Byers Peninsula, Livingston Island, Antarctica. CATENA, 149, 538-547. Duffy, G.A., Lee, J.R., 2019. Ice-free area expansion compounds the non-native species threat to Antarctic terrestrial biodiversity. Biological Conservation, 232, 253-257.

27

Fretwell, P.T., Convey, P., Fleming, A.H., Peat, H.J., Hughes, K.A., 2011. Detecting and mapping vegetation distribution on the Antarctic Peninsula from remote sensing data. Polar Biology, 34, 273–281. Funaki, M., Higashino S.-I., Sakanaka, S., Iwata N., Nakamura, N., Hirasawa, N., Obara, N., Kuwabara, M., 2014. Small unmanned aerial vehicles for aeromagnetic surveys and their flights in the South Shetland Islands, Antarctica. Polar Science, 8, 342-356. Goebel, M.E., Perryman, W.L., Hinke, J.T., Krause, D.J., Hann, N.A., Gardner, S., LeRoi, D.J., 2015. A small unmanned aerial system for estimating abundance and size of Antarctic predators. Polar Biology, 38, 619-630. Green, T.G.A., Brabyn, L., Beard, C., Sancho, L.G., 2012. Extremely low lichen growth rates in Taylor Valley, Dry Valleys, continental Antarctica. Polar Biology, 35, 535–541. Green, T.G.A., Sancho, L.G., Pintado, A., Schroeter, B., 2011. Functional and spatial pressures on terrestrial vegetation in Antarctica forced by global warming. Polar Biology, 34, 1643–1656. Guglielmin, M., Dalle Fratte, M., Cannone, N., 2014. Permafrost warming and vegetation changes in continental Antarctica. Environmental Research Letters, 9, 045001. Hu, S.-S., 1998. Moss community types and species diversity of Southern Fields Peninsula (King George Island, South Shetland Islands) Antarctica. Journal of Hattori Botany Laboratory, 84, 187-198. Jawak, S.D., Udhayaraj, A.D., Luis, A.J. Geospatial mapping of vegetation in the Antarctic environment using very high-resolution WorldView-2 imagery. Proc. SPIE 9877, Land Surface and Cryosphere Remote Sensing III, 2016, 98772N. 28

Kejna, M., Arazny, A., Sobota, I., 2013. Climatic change on King George Island in the years 1948–2011. Polish Polar Research, 34, 213–235. Kim, S.-H., Hong, C.-L., 2012. Antarctic land-cover classification using IKONOS and Hyperion data at Terra Nova Bay. International Journal of Remote Sensing, 33, 7151–7164. Korczak-Abshire, M., Zmarz, A., Rodzewicz, M., Kycko, M., Karsznia, I., Chwedorzewska, K.J., 2019. Study of fauna population changes on Penguin Island and Turret Point Oasis (King George Island, Antarctica) using an unmanned aerial vehicle. Polar Biology, 42: 217-224. Krause, D.J., Hinke, J.T., Perryman, W.L., Goebel, M.E., LeRoi, D.J., 2017. An accurate and adaptable photogrammetric approach for estimating the mass and body condition of pinnipeds using an unmanned aerial system. PLoS ONE, 12, e0187465. Lee, J.R., Raymond, B., Bracegirdle, T.J., Chadès, I., Fuller, R.A., Shaw, J.D., Terauds, A., 2017. Climate change drives expansion of Antarctic ice-free habitat. Nature, 547, 49-54. Li, T., Zhang, B., Cheng, X., Westoby, M.J., Li, Z., Ma, C., Hui, F., Shokr, M., Liu, Y., Chen, Z., Zhai, M., Li, X., 2019. Resolving fine-scale surface features on polar sea ice: A first assessment of UAS photogrammetry without ground control. Remote Sensing, 11, 784. Li, Y., Kromer, B., Schukraft, G., Bubenzer, O,, Huang, M.-R., Wang, Z.-M., Bian, L.-G., Li., C.S., 2014. Growth rate of Usnea aurantiacoatra (Jacq.) Bory on Fildes Peninsula, Antarctica and its climatic background. PLoS ONE, 9(6), e100735. Lindsay, D.C., 1971. Vegetation of the South Shetlands. BAS Bulletin, 25, 59-83.

29

López-Martínez, J., Serrano, E., Schmid, T., Mink, S., Linés, C., 2012. Periglacial processes and landforms in the South Shetland Islands (northern Antarctic Peninsula region). Geomorphology, 155–156, 62–79. Lovelock, C.E., Robinson, S.A., 2002. Surface reflectance properties of Antarctic moss and their relationship to plant species, pigment composition and photosynthetic function. Plant, Cell and Environment, 25, 1239–1250. Lucieer, A., Turner, D., King, D.H., Robinson, S.A., 2014. Using an Unmanned Aerial Vehicle (UAV) to capture micro-topography of Antarctic moss beds. International Journal of Applied Earth Observation and Geoinformation, 27, 53-62. Malenovský, Z., Lucieer, A., King, D.H., Turnbull, J.D., Robinson, S.A., 2017. Unmanned aircraft system advances health mapping of fragile polar vegetation. Methods in Ecology and Evolution, 8, 1842-1857. Malenovský, Z., Turnbull, J.D., Lucieer, A., Robinson S.A., 2015. Antarctic moss stress assessment based on chlorophyll content and leaf density retrieved from imaging spectroscopy data. New Phytologist, 208, 608-624. Michel, R.F.M., Schaefer, C.E.G.R., Martínez, J.L., Simas, F.N.B., Haus, N.W., Serrano, E., Bockheim, J.G., 2014. Soils and landforms from Fildes Peninsula and Ardley Island, Maritime Antarctica. Geomorphology, 225: 76–86. Mora, C., Jiménez, J.J., Pina, P., Catalão, J., Vieira, G., 2017. Evaluation of single-band snowpatch mapping using high-resolution microwave remote sensing: an application in the maritime Antarctic. The Cryosphere, 11, 139-155.

30

Mustafa, O., Barbosa, A., Krause, D.J., Peter, H.-U., Vieira, G., Rümmler, M.-C., 2018. State of knowledge: Antarctic wildlife response to unmanned aerial systems. Polar Biology, 41, 2387. Olech, M., 2002. Plant Communities on King George Island. In: Beyer L., Bölter M. (eds) Geoecology of Antarctic Ice-Free Coastal Landscapes, vol 154, 215-231, Springer, Berlin, Heidelberg. Oliva, M., Navarro, F., Hrbáček, F., Hernández, A., Nývlt, D., Pereira, P., Ruiz-Fernández, J., Trigo, R., 2017. Recent regional climate cooling on the Antarctic Peninsula and associated impacts on the cryosphere. Science of the Total Environment, 580, 210– 223. Oliva, M., Pereira, P., Antoniades, D., 2018. The environmental consequences of permafrost degradation in a changing climate. Science of The Total Environment, 616–617, 435437. Øvstedal, D.O., Lewis-Smith R.I., Lichens of Antarctica and South Georgia. A guide to their identification and ecology. Cambridge University Press: Cambridge, 2001, 411 pp. Peat H.J., Clarke A., Convey P., 2007. Diversity and biogeography of the Antarctic flora. Journal of Biogeography, 34: 132-146. Pereira, P., Brevik, E., Trevisani, S., 2018. Mapping the environment. Science of The Total Environment, 610–611, 17-23. Peter, H.-U., Buesser, C., Mustafa, O., Pfeiffer S. Risk assessment for the Fildes Peninsula and Ardley Island, and development of management plans for their designation as

31

specially protected or specially managed areas. Research Report. Umweltbundesamt, Germany, 2008 (508 pp.). Pfeifer, C., Barbosa, A., Mustafa, O., Peter, H.-U., Rümmler, M.C., Brenning, A., 2019. Using fixed-wing UAV for detecting and mapping the distribution and abundance of penguins on the South Shetlands Islands, Antarctica. Drones, 3, 39. Pina, P., Vieira, G., Bandeira, L., Mora C., 2016. Accurate determination of surface reference data in digital photographs in ice-free surfaces of Maritime Antarctica. Science of the Total Environment, 573, 290-302. Pina P., Pereira F., Marques J.S., Heleno S., 2019. Detection of stone circles in periglacial regions of Antarctica in UAV datasets. In Morales A., Fierrez J., Salvador Sánchez J., Ribeiro B. (eds.), Pattern Recognition and Image Analysis, Lecture Notes in Computer Science,

11867:

279-288,

Springer

Nature

Switzerland

AG.

https://doi.org/10.1007/978-3-030-31332-6_25 Pudełko, R., Angiel, P.J., Potocki, M., Jędrejek, A., Kozak, M., 2018. Fluctuation of glacial retreat rates in the eastern part of Warszawa Icefield, King George Island, Antarctica, 1979–2018. Remote Sensing, 10, 892. Ramos, M., Vieira, G., de Pablo, M. A., Molina, A., Abramov, A., Goyanes, G., 2017. Recent shallowing of the thaw depth at Crater Lake, Deception Island, Antarctica (2006– 2014). CATENA, 149, 519-528. Robinson, S.A., King, D.H., Bramley-Alves, J., Waterman, M.J., Ashcroft, M.B., Wasley, J., Turnbull, J.D., Miller, R.E., Ryan-Colton, E., Benny, T., Mullany, K., Clarke, L.J., Barry,

32

L.A., Hua, Q., 2018. Rapid change in East Antarctic terrestrial vegetation in response to regional drying. Nature Climate Change, 8, 879–884. Ruiz-Fernández, J., Oliva, M., García-Hernández, C., 2017. Topographic and geomorphologic controls on the distribution of vegetation formations in Elephant Point (Livingston Island, Maritime Antarctica). Science of the Total Environment, 587-588, 340-349. Ruiz-Fernández, J., Oliva, M., Nývlt, D., Cannone, N., García-Hernández, C., Guglielmin, M., Hrbáček, F., Roman, M., Fernández, S., López-Martínez, J., Antoniades, D., 2019. Patterns of spatio-temporal paraglacial response in the Antarctic Peninsula region and associated ecological implications. Earth-Science Reviews, 192, 379-402. Sancho, L.G., Green, T.G.A., Pintado, A., 2007. Slowest to fastest: Extreme range in lichen growth rates supports their use as an indicator of climate change in Antarctica. Flora, 202, 667–673. Sancho, L.G., Pintado, A., 2004. Evidence of high annual growth rate for lichens in the maritime Antarctic. Polar Biology, 27, 312–319. Sancho, L.G., Pintado, A., Green, T.G.A., 2019. Antarctic studies show lichens to be excellent biomonitors of climate change. Diversity, 11, 42. Sancho, L.G., Pintado, A., Navarro, F., Ramos, M., De Pablo, M.A., Blanquer, J.M., Raggio, J., Valladares, F., Green, T.G.A., 2017. Recent warming and cooling in the Antarctic Peninsula Region has rapid and large effects on lichen vegetation. Scientific Reports, 7, 5689. Schmid, T., Guillaso, S., López-Martínez, J., Nieto, A., Mink, S., Koch, M., 2018. Advances in mapping ice-free surfaces within the Northern Antarctic Peninsula region using 33

polarimetric RADARSAT-2 data. In IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, Valencia, Spain, 22-27 July 2018, 216-219. Schmid, T., López-Martínez, J., Guillaso, S., Serrano, E., D'Hondt, O., Koch, M., Nieto, A., O'Neill, T., Mink, S., Durán, J.J., Maestro, A., 2017. Geomorphological mapping of icefree areas using polarimetric RADARSAT-2 data on Fildes Peninsula and Ardley Island, Antarctica. Geomorphology, 293, 448-459. Shin, J.I., Kim, H.C., Lim, S.I., Hong, S.G., 2014. Vegetation abundance on the Barton Peninsula, Antarctica: estimation from high-resolution satellite images. Polar Biology, 37, 1579–1588. Siegert, M., Atkinson, A., Banwell, A., Brandon, M., Convey, P., Davies, B., Downie, R., Edwards, T., Hubbard, B., Marshall, G., Rogelj, J., Rumble, J., Stroeve, J., Vaughan, D., 2019. The Antarctic Peninsula under a 1.5◦C global warming scenario. Frontiers in Environmental Science, 7:102. Simonov, I.M., 1977. Physical-geographic description of the Fildes Peninsula (South Shetland Islands). Polar Geography, 1, 223–242. Soille, P., 2004. Morphological Image Analysis. Principles and Applications. 2nd edition, Springer: Berlin-Heidelberg, pp Turner, D., Lucieer, A., Malenovský, Z., King, D., Robinson, S.A., 2018. Assessment of Antarctic moss health from multi-sensor UAS imagery with Random Forest Modelling. International Journal of Applied Earth Observation and Geoinformation, 68, 168–179.

34

Turner, D., Lucieer, A., Malenovský, Z., King, D.H., Robinson, S.A., 2014. Spatial coregistration of ultra-high resolution visible, multispectral and thermal images acquired with a micro-UAV over Antarctic moss beds. Remote Sensing, 6, 4003–4024. Turner, J., Colwell, S.R., Marshall, G.J., Lachlan-Cope, T.A., Carleton, A.M., Jones, P.D., Lagun, V., Reid, P.A., Iagovkina, S., 2005. Antarctic climate change during the last 50 years. International Journal on Climatology, 25, 2279–2294. Turner, J., Lu, H., White, I., King, J.C., Phillips, T., Hosking, J.S., Bracegirdle, T.J., Marshall, G.J., Mulvaney, R., Deb, P., 2016. Absence of 21st century warming on Antarctic Peninsula consistent with natural variability. Nature, 535, 411-415. Vapnik, V.N. The Nature of statistical learning theory. Springer-Verlag: Berlin, 1995, pp. Vieira, G., Bockheim, J., Guglielmin, M., Balks, M., Abramov, A.A., Boelhouwers, J., Cannone, N., Ganzert, L., Gilichinsky, D.A., Gotyachkin, S., López-Martínez, J., Meiklejohn, I., Raffi, R., Ramos, M., Schaefer, C., Serrano, E., Simas, F., Sletten, R., Wagner, D., 2010. Thermal state of permafrost and active-layer monitoring in the Antarctic: advances during the International Polar Year 2007–2009. Permafrost and Periglacial Processes, 21, 182–197. Vieira, G., Mora, C., Pina, P., Schaefer, C.E.R., 2014. A proxy for snow cover and winter ground surface cooling: mapping Usnea sp. communities using high resolution remote sensing imagery (Maritime Antarctica). Geomorphology, 225, 69–75. Westoby, M.J., Dunning, S.A., Woodward, J., Hein, A.S., Marrero, S.M., Winter, K., Sugden, D.E., 2015. Sedimentological characterization of Antarctic moraines using UAVs and Structure-from-Motion photogrammetry. Journal of Glaciology, 61, 1088-1102. 35

Zmarz, A., Rodzewicz, M., Dabski, M., Karsznia, I., Korczak-Abshire, M., Chwedorzewska, K.J., 2017. Application of UAV BVLOS remote sensing data for multi-faceted analysis of Antarctic ecosystem. Remote Sensing of Environment, 217, 375-388.

Figure 1. Location of Fildes Peninsula in (a) Antarctica, (b) in King George Island and (c) with the respective elevation map. The white rectangle in Meseta Norte indicates the UAV survey site. Figure 2. UAV-based datasets: (a) Mosaic built from 3063 individual images and (b) respective DEM. The circumscribed rectangle indicates the study area of 534,296 m2 and the division in two equal areas (UAV-west and UAV-east) to train and validate the classifier. Figure 3. Vegetation examples during 2017 summer in WV2 satellite in a NIR-G-B composite (left) and UAV in RGB (right) with predominance of Usnea sp. (top) and mosses (bottom) with image contrast enhanced to highlight vegetation. Figure 4. Classification of study site in the years: (a) 2006, (b) 2013 and (c) 2017. Figure 5. Changes detected in the periods: (a) 2006-2017, (b) 2006-2013 and (c) 2013-2017.

36

Figure 6. Examples of areas of Usnea sp. in 2006 that were kept unchanged in 2013 (transparent yellow) and those that have changed to bare ground in 2013 (red dots, centred at each pixel of 2 m). Background image is the shaded-relief DEM built from the UAV surveys and the scale bar is the same for all images.

Table 1. Remotely sensed datasets available: QB-QuickBird satellite, WV2-WorldView2 satellite and UAV-DJI Phantom3 Unmanned Aerial Vehicle.

21/Feb/2006

Local time GMT-3 (hh:mm) 13:37

Off-nadir angle (°) 0

Sun azimuth (°) 45.3

Sun elevation (°) 31.6

Cloud cover (%) 0

WV2

13/Apr/2013

14:07

27

28.3

16.3

0

3

WV2

19/Mar/2017

13:31

0

41.5

22.0

14

4

UAV

20-30/Jan/2017

variable

0

variable

variable

0

Data Set

Platform

Acquisition date

1

QB

2

Table 2. Characteristics of training/testing samples per surface class.

Classes Usnea sp. Mosses Bare

UAV-west Objects Pixels 292 (27.6%) 1953 (31.3%) 138 (13.0%) 845 (13.6%) 628 (59.4%) 3717 (55.1%) 37

UAV-east Objects Pixels 451 (31.1%) 3017 (37.9%) 205 (14.2%) 1229 (15.4%) 793 (54.7%) 3717 (46.7%)

Total

1058 (100%)

6234 (100%)

1449 (100%)

7963 (100%)

Table 3. Classifications performances in WV2 imagery in UAV site (2017). UAV-west oa (%) kappa SVM γ =0.03, C=100 γ =0.1, C=100 γ =0.03, C=1000 γ =0.1, C=1000 kNN k=1 k=3 k=7 k=9

UAV-east oa (%) kappa

96.9 96.3 96.6 96.3

0.950 0.945 0.940 0.940

94.2 95.1 95.6 93.5

0.900 0.915 0.940 0.886

87.1 90.9 90.3 89.5

0.772 0.840 0.830 0.817

88.1 87.9 85.6 86.5

0.766 0.766 0.724 0.740

Classified

Table 4. Confusion matrix and derived errors (%), commission (com.) and omission (omi.) per surface class for the best classification model (SVM classifier with parameters γ =0.03 and C=1000).

Class Usnea sp. Mosses Bare

UAV-west Reference Usnea sp. Mosses Bare 98.7 1.2 3.7 0.0 95.0 0.0 1.3 3.8 96.3

errors com. omi. 2.3 1.3 0.0 5.0 4.8 3.7

UAV-east Reference Usnea sp. Mosses Bare 98.6 0.1 2.8 0.0 99.9 6.0 1.4 0.0 91.2

Table 5. Classified areas by surface type. 2006

2013 38

2017

errors com. omi. 3.4 1.4 7.0 0.1 1.3 8.8

Usnea sp. Mosses Usnea sp. + mosses Bare

m2 131,378 39,214 170,592 214,820

% 34.1 10.2 44.3 55.7

m2 118,056 42,810 160,866 224,547

% 30.6 11.1 41.7 58.3

m2 117,836 35,380 153,216 232,196

Table 6. Surface variation per class (%) relatively to its area in the earlier date. Usnea sp. Mosses Bare

2006-2017 -10.3 -9.8 8.1

2006-2013 -10.1 9.2 4.5

2013-2017 -0.2 -17.4 3.4

Table 7. Exchanges between surfaces (% of total area). Changes Usnea sp. unchanged Usnea sp. to mosses Usnea sp. to bare Mosses unchanged Mosses to Usnea sp. Mosses to bare Bare unchanged Bare to Usnea sp. Bare to mosses

2006-2017 25.0 0.2 8.8 6.1 0.4 3.7 48.2 5.1 2.4

39

2006-2013 24.6 0.5 9.0 7.1 0.3 2.7 46.6 5.7 3.4

2013-2017 24.1 0.2 6.3 6.7 0.6 3.8 50.6 5.8 1.9

% 30.6 9.2 39.8 60.2