1 This is the accepted manuscript of the following article: 2 https://www.sciencedirect.com/science/article/abs/pii/S0034425722001547 3 Site-specific scaling of remote sensing-based estimates of woody cover and aboveground 4 biomass for mapping long-term tropical dry forest degradation status 5 Tobias Fremouta,b,*, Jorge Cobián de Vinateac, Evert Thomasb, Wilson Huamánd, Mike Salazar- 6 Villegasc,e, Daniela Limache de la Fuentec, Paulo N. Bernardinoa,f, Rachel Atkinsonb, Elmar 7 Csaplovicsc, Bart Muysa 8 a Division of Forest, Nature and Landscape, Department of Earth and Environmental Sciences, KU 9 Leuven, Celestijnenlaan 200E, B-3001 Leuven, Belgium 10 b Alliance Bioversity International - CIAT, Av. La Molina 1895, La Molina, Lima, Peru 11 c Institute of Remote Sensing and Photogrammetry, TU Dresden, Helmholtzstraße 10, D-01069 12 Dresden, Germany 13 d Facultad de Ciencias Biológicas, Universidad Nacional de San Antonio Abaad de Cuzco, Av. de La 14 Cultura 773, Cusco, Peru 15 e Alliance Bioversity International – CIAT, Recta Cali-Palmira Km 17, Cali, Valle del Cauca, Colombia 16 f Laboratory of Geo-Information Science and Remote Sensing, Wageningen University, Wageningen, 17 The Netherlands 18 19 * corresponding author: tobias.fremout@gmail.com 20 21 E-mail addresses: Tobias Fremout: tobias.fremout@gmail.com; Jorge Cobián jorge.cdv@gmail.com; 22 Evert Thomas: e.thomas@cgiar.org; Wilson Huamán: wilson243807@gmail.com; Mike Salazar- 23 Villegas: mhvillegas@gmail.com; Daniela Limache de la Fuente: d.limache.delafuente@gmail.com; 24 Paulo Negri Bernardino: negribernardino.paulo@kuleuven.be; Rachel Atkinson: r.atkinson@cgiar.org; 25 Elmar Csaplovics: elmar.csaplovics@tu-dresden.de; Bart Muys: bart.muys@kuleuven.be; 26 27 1 28 Abstract 29 Remote sensing-based approaches are important for evaluating ecosystem degradation and the efficient 30 planning of ecosystem restoration efforts. However, the large majority of remote sensing-based 31 degradation assessments are trend-based, implying that they can only detect degradation that occurred 32 after medium or high-resolution satellite imagery became available. This makes them less suitable to 33 map long-term degradation in ecosystems that have been under high human pressure since before. The 34 main goal of this study was to develop a robust operational approach to map forest degradation status 35 in heterogeneous landscapes with a long-standing degradation history to inform the planning of 36 restoration interventions. We hereby use the tropical dry forests of Lambayeque, Peru, as a case study. 37 Instead of using a trend-based assessment, we evaluated forest degradation status by comparing current 38 woody cover (WC) and aboveground biomass (AGB) estimates obtained from remote sensing imagery 39 with benchmark values consisting of the 95th percentile WC and AGB values inside environmentally 40 homogenous land capability classes. Using boosted regression tree models and a combination of optical 41 (Sentinel-2) and synthetic aperture radar (Sentinel-1) data of different seasons, we mapped WC and 42 AGB, using training data obtained through very high-resolution imagery and field measurements. 43 Further, we aimed at assessing (i) whether the inclusion of Sentinel-1 data improves mapping accuracy 44 in comparison to using only Sentinel-2 data, and (ii) whether the use of multi-seasonal data improves 45 accuracy in comparison to single-season data. Models combining multi-seasonal Sentinel-1 and 46 Sentinel-2 data resulted in the most accurate WC predictions (mean absolute error (MAE): 16%; MAE 47 normalized by dividing by the inter-quartile range of training data: 26%) and AGB predictions (MAE: 48 28.6 t/ha; normalized MAE: 65%), but differences in predictive accuracy with single season models or 49 models using only Sentinel-2 data were small. The most accurate models estimated an average WC of 50 41% and an average AGB of 23.4 t/ha. Average WC and AGB reduction due to degradation was 35% 51 and 36%, respectively, indicating that these forests are highly degraded. The site-specific scaling of 52 WC and AGB allows to efficiently estimate forest degradation status irrespective of the time when this 53 degradation occurred, and to express degradation status against site-specific benchmarks. On the 54 condition that there are still some areas that are sufficiently undegraded to be used as a benchmark, the 2 55 approach can be used to prioritize forest restoration actions and inform targets for restoration in 56 heterogeneous landscapes suffering the impacts of undocumented long-term degradation. 57 1 Introduction 58 Drylands, usually defined as areas with an aridity index (i.e. annual precipitation over annual potential 59 evapotranspiration) lower than 0.65 (UNEP, 1992), cover around 45% of the Earth’s land surface. They 60 are home to more than 2.5 billion people and more than half of the world’s livestock (Gaur and Squires, 61 2017; Prăvălie, 2016). As a consequence of rising evapotranspiration levels due to climate change, 62 drylands are projected to cover at least half of the Earth’s surface by the end of the century (Huang et 63 al., 2016). Although dryland forests support lower biomass levels than their humid counterparts, they 64 play a crucial role in the delivery of ecosystem services to local communities, including soil 65 stabilization, watershed protection, and the provision of fuelwood and year-round forage (Safriel and 66 Adeel, 2005). However, drylands are under increasing pressure of agricultural expansion, and are 67 particularly vulnerable to land degradation (Reynolds et al., 2007; Walther et al., 2011). 68 Mapping the degradation status of dryland forests is important for the planning of conservation and 69 restoration efforts. Forest degradation can be broadly described as the ‘reduction of the capacity of the 70 forest to provide goods and services’ (FAO, 2010). Many interpretations of the concept of forest 71 degradation exist, which lead Thompson et al. (2012) to propose an operational framework for 72 monitoring forest degradation using five criteria: productivity, biodiversity, unusual disturbances (e.g. 73 invasive species), protective functions (e.g. erosion control), and carbon storage. As all of these can be 74 assessed with remote sensing to some extent (Thompson et al., 2012), remote sensing-based approaches 75 are indispensable for assessing forest degradation over large areas. However, dryland forests have 76 several characteristics that pose challenges to remote sensing, including open canopies leading to mixed 77 pixels consisting of woody vegetation, herbaceous vegetation, and bare soil, and a high share of 78 deciduous woody species, which can be difficult to distinguish from the soil background during the dry 79 season (Eisfelder et al., 2012). Additionally, the wet season is usually short, hindering the construction 80 of cloudless wet season composite images over large areas, while large inter-annual rainfall variability 3 81 makes the analysis of trends in vegetation characteristics a complex undertaking (Burrell et al., 2017). 82 Along with the threatened status of many dryland forests, these challenges have motivated numerous 83 remote sensing efforts in dryland forests over the last decade (e.g., Yang et al. 2012, Halperin et al. 84 2016, Baumann et al. 2018). While many challenges remain, some promising approaches have emerged, 85 such as the use of multi-seasonal data (e.g., Brandt et al. 2016, Higginbottom et al. 2018, Van Passel et 86 al. 2020), the combination of optical and synthetic aperture radar (SAR) data (e.g., Baumann et al. 2018, 87 Hirschmugl et al. 2018, Reiche et al. 2018), as well as more advanced techniques such as spectral 88 mixture analysis (e.g., Yang et al. 2012, Cao et al. 2015, Mayes et al. 2015) and time-series 89 segmentation (e.g., Burrell et al. 2017, Schneibel et al. 2017, Bernardino et al. 2020). 90 Remote sensing-based forest degradation assessments usually rely on identifying negative trends in 91 forest attributes (e.g., forest cover, productivity, biomass) through time (Hirschmugl et al., 2017). While 92 crucial for early detection of forest degradation and better understanding of forest degradation processes 93 and its drivers, trend-based methods can only detect degradation that occurred after medium or high- 94 resolution satellite imagery became available a few decades ago, making them less suitable to assess 95 the degradation status of areas that have already undergone widespread degradation before (Prince, 96 2004; Wessels et al., 2008). In such a context, it is crucial to identify a non-degraded reference state to 97 compare remote sensing-based indicators with, using a space-for-time substitution (Verón et al., 2006; 98 Wessels et al., 2008). 99 Proposed solutions include using protected areas as reference areas (Garbulsky and Paruelo, 2004), 100 using areas in which forest structural attributes remain stable through time as reference areas (Romero- 101 Sanchez and Ponce-Hernandez, 2017), modelling potential Normalized Difference Vegetation Index 102 (NDVI) values (Boer and Puigdefábregas, 2003), or the local net primary productivity scaling (LNS) 103 method proposed by Prince (2004). The LNS method involves grouping pixels in land capability classes 104 with homogeneous site conditions, estimating net primary productivity as the sum of NDVI values over 105 the growing season, and estimating the relative degradation status of each pixel by relating its net 106 primary productivity to a reference value of the land capability class in which it is located (Prince et al. 107 (2009) use the 90th percentile of the net primary productivity within each class). The LNS approach has 4 108 been applied in several contexts, often in landscapes dominated by rangelands and/or croplands 109 (Jackson and Prince, 2016a; Li et al., 2020; Noojipady et al., 2015; Prince et al., 2009; Wessels et al., 110 2008) and has been extended to the local scaling of the fractional cover of photosynthetic and non- 111 photosynthetic vegetation (Jackson and Prince, 2016b), but it has not been used yet to evaluate forest 112 degradation. 113 As loss of primary productivity is only one aspect of dryland degradation (Herrmann et al., 2020) and 114 altered forest structure is an important aspect of forest degradation (Gao et al., 2020; Thompson et al., 115 2012), here we adapt the LNS approach to the local scaling of forest woody cover (WC) and 116 aboveground biomass (AGB). Under the forest degradation framework proposed by Thompson et al. 117 (2012), AGB is directly related to the carbon storage criterion, while WC is also closely related to 118 protective functions (with higher WC providing better protection to erosion). Considering that land 119 capability classes are not necessarily spatially contiguous, we prefer to use the term ‘site-specific 120 scaling’ rather than ‘local scaling’, also for consistency with ‘site quality’ concept widely used in 121 forestry (Bontemps and Bouriaud, 2014; Carmean, 1975). As this site-specific scaling approach uses a 122 space-for-time substitution (i.e. using variation in degradation status in space as a proxy to understand 123 the temporal process of degradation), it relies on the assumption that there are still some areas left that 124 are sufficiently undegraded which can be used as a benchmark to compare current WC and AGB levels 125 with. 126 The tropical dry forests (TDFs) of northwestern Peru, characterized by a high number of endemic 127 woody species (Linares-Palomino, 2006), have supported large human populations since Pre- 128 Columbian times (Hocquenghem, 2001). While these TDFs initially recovered soon after European 129 colonization in the 16th century as a consequence of widespread mortality among native communities, 130 they came under large pressure again in the 19th century as a consequence of large-scale cotton 131 production, livestock grazing and charcoal production, and they have been under high human pressure 132 ever since (Hocquenghem 2001, Fremout et al. 2020). The situation in northwestern Peru is not unique 133 but similar to TDF ecosystems elsewhere, which have historically supported higher human population 134 densities than rainforests thanks to more fertile soils, vegetation that is easily cleared by fire, climatic 5 135 conditions conducive to annual crop production and livestock husbandry, and lower prevalence of pests 136 and diseases (Janzen, 1988; Murphy and Lugo, 1986). While around 60% of the original TDF cover in 137 northwestern Peru remains (MINAM, 2018), the remaining forests are severely impacted by 138 overexploitation, overgrazing, and forest fires (Fremout et al. 2020). Many restoration and conservation 139 projects have emerged to counter the degradation of these forests (Cerrón et al., 2019), but spatial 140 planning to prioritize areas for restoration and conservation is lacking. Remote sensing research in the 141 region is almost non-existent (but see Baena et al., 2017; Muenchow et al., 2020; Padrón and Navarro- 142 Cerrillo, 2007), resulting in a lack of reliable spatial data to inform spatial planning efforts. Due to the 143 often sparse vegetation in the lowland TDFs of northwestern Peru, global remote sensing products 144 usually only detect the TDFs in the highlands, resulting in many of the TDFs being omitted from global 145 tree cover maps (e.g., Hansen et al. 2013) or wrongly classified as grasslands and shrublands in global 146 land cover maps (e.g., Chen et al. 2015, ESA 2017, Buchhorn et al. 2020). 147 The main goal of this study was to develop a robust operational approach to map forest degradation 148 status in heterogeneous landscapes with a long-standing degradation history, to inform the large-scale 149 planning of forest restoration actions. We did so by adopting and extending the LNS method (Prince, 150 2004; Prince et al., 2009) to the site-specific scaling of forest woody cover (WC) and aboveground 151 biomass (AGB), using the TDFs of Lambayeque in northwestern Peru as a case study. We combined 152 multispectral optical (Sentinel-2) and synthetic aperture radar (SAR; Sentinel-1) data to map the WC 153 and AGB of these forests, using very high-resolution images and field data as training data. To obtain 154 estimates of forest degradation status, these WC and AGB estimates were compared with benchmark 155 WC and AGB values obtained from environmentally homogeneous land capability classes. Further, we 156 aimed at assessing (i) whether the inclusion of Sentinel-1 data improves mapping accuracy in 157 comparison to using only Sentinel-2 data, and (ii) whether the use of multi-seasonal data improves 158 mapping accuracy in comparison to single-season data. 6 159 2 Methodology 160 2.1 Study region 161 The study area comprises the TDFs of Lambayeque, one of Peru´s northernmost departments along the 162 Pacific coast (Figure 1). Part of the Tumbes-Piura dry forest ecoregion (Olson et al., 2001), the potential 163 natural vegetation in most of Lambayeque consists of different types of TDFs, varying mostly with 164 elevation. In the hyper-arid (aridity index < 0.05) lowlands bordering the Sechura desert (locally called 165 ‘bosques secos de llanura’), tree cover is generally sparse (tree cover < 33%), except in riparian areas. 166 With an average annual precipitation around 100 mm or lower, these forests strongly rely on 167 groundwater and fog coming in from the Pacific Ocean. They are often dominated by deep-rooted semi- 168 deciduous Prosopis pallida (Fabaceae) trees and evergreen sclerophyllous trees and shrubs such as 169 Colicodendron scabridum (Capparaceae) and Beautempsia avicenniifolia (Capparaceae). Closer to the 170 Andean cordillera, the lowland TDFs transition into the arid (aridity index between 0.05 and 0.20) hilly 171 TDFs (‘bosques secos de colina’), dominated by deciduous species such as Loxopterygium huasango 172 (Anacardiaceae) and Bursera graveolens (Burseraceae). At higher elevation, these forests transition 173 into semi-arid (aridity index between 0.20 and 0.50) montane TDFs (‘bosques secos de montaña’), with 174 annual precipitation levels of ca. 300-500 mm. The montane TDFs have an upper altitudinal limit of ca. 175 1600 masl (Linares-Palomino, 2006) and are often dominated by deciduous tree species with succulent 176 stems such as Eriotheca ruizii (Malvaceae) and Erythrina velutina (Fabaceae). 177 As most of the TDFs of Lambayeque are too dry for rainfed agriculture, cropping land is largely 178 restricted to river valleys. Pressure for arable land is high, as evidenced by the encroachment of two 179 protected areas (Santuario Histórico Bosque de Pómac and Reserva Ecológica Chaparrí) by smallholder 180 farmers over the past two decades. While agricultural areas in river valleys are still expanding, most 181 deforestation in recent years has taken place in the plains in the northwestern part of the department as 182 a consequence of a large-scale irrigation project that tunnels water from the Andes to these lowland 183 TDFs (‘Proyecto Olmos’). Remaining forest areas in Lambayeque are often highly degraded; important 184 drivers include charcoal production and overgrazing by free-roaming livestock. 7 185 2.2 Modelling forest extent and structure 186 2.2.1 Delineation of the study region and stratification 187 We delineated our study region as all non-converted tropical dry forest areas in Lambayeque, with an 188 elevation below 1600 masl (Linares-Palomino, 2006) and excluding desert areas as indicated in the 189 national ecosystems map of Peru (Mapa Nacional de Ecosistemas; MINAM, 2018). Converted areas 190 were excluded from the study region through land use modelling (forest vs. converted areas) using 191 multi-seasonal remote sensing data (see Supplementary Material 1 for methodological details and 192 results), resulting in a study region of ca. 8,000 km2 (Figure 1). Based on aridity (i.e. annual precipitation 193 divided by potential evapotranspiration), we stratified our study region in hyper-arid, arid and semi-arid 194 areas (Figure 1), roughly corresponding with the lowland, hilly and montane forest types described 195 above. The aridity map was constructed using WorldClim temperature and precipitation data (Fick and 196 Hijmans, 2017) and the Hargreaves formula for estimating potential evapotranspiration (Hargreaves 197 and Allen, 2003). 198 Figure 1: The location of Lambayeque in Peru (panel a), study region with aridity zones (panel b) and Normalized Difference Vegetation Index (NDVI; panel c).The study region consists of non-converted tropical dry forests areas in Lambayeque (see Supplementary Material 1 for details on how the forest extent map was obtained). . The NDVI map (panel c) depicts the maximum of the NDVI composites over 3 seasons (wet, transition, and dry season) derived from Sentinel-2 data (see section 2.2.3). 199 8 200 2.2.2 Reference data 201 Woody cover 202 Training data to calibrate the woody cover (WC) models were acquired through semi-automated visual 203 interpretation of very high-resolution Google Earth imagery, using the most recent imagery available 204 at the time of analysis (April 2020). The training data consisted of 300 Google Earth image chips (i.e. 205 small subsets of Google Earth images) in each of the three aridity strata (i.e. a total of 900 image chips) 206 obtained through random stratified sampling. More specifically, they consisted of a random subset of 207 the image chips that were classified as ‘forest’ through visual assessment and used for training the forest 208 extent model (see Supplementary Material 1 for further details). We downloaded the Google Earth 209 image chips using the ‘RgoogleMaps’ package for R (Loecher and Ropkins, 2015) as image chips with 210 an extent of 95x95m (i.e. one of the integer zoom levels available through the package). While WC 211 estimates are often made by assessing a number of sampling points within very high-resolution images 212 (e.g., Carreiras et al. 2006, Coulston et al. 2012, Anchang et al. 2020), we followed the approach 213 proposed by Nagelkirk & Dahlin (2020) instead: Since woody vegetation in our study region is usually 214 darker than surrounding herbaceous vegetation and soil, we used pixel brightness (i.e., the sum of red- 215 green-blue (RGB) values) to threshold the very high-resolution images. This approach, illustrated in 216 (Figure 2), was implemented using the ‘imager’ package for R (Barthelme, 2020). Using a semi- 217 automated approach, we estimated the total WC in each of the selected images by selecting the most 218 appropriate brightness threshold by visual inspection of 10 automatically generated brightness threshold 219 levels between 0.3 and 0.6. When the thresholding did not allow to confidently assess the woody cover 220 of an image, no WC estimates were made. This was the case for images containing clouds, hillshade, 221 or both evergreen and deciduous trees (in which case these trees sometimes needed a different 222 brightness threshold), as well as for some images taken during the wet season (not enough contrast 223 between herbaceous cover and trees) . Consequently, a total of 720 images could be evaluated. 9 224 225 Figure 2: Illustration of the woody cover assessment method using thresholding of Google Earth imagery: (a) 226 190x190m Google Earth image chip, (b) 95x95m Google Earth image chip, (c) woody cover (WC) obtained by 227 thresholding the brightness (i.e., the sum of the red-green-blue (RGB) values) of the 95x95m image chip. 228 Aboveground biomass 229 While WC is undoubtedly an important forest degradation indicator, woody vegetation in TDFs may 230 range from shrubs to large trees, so areas with similar WC may differ considerably in terms of carbon 231 stocks and delivery of ecosystem services. We therefore complemented the woody cover data with field 232 measurements of aboveground biomass (AGB), carried out over a six-week period in April and May 233 2019. Using a random stratified sampling scheme taking into account (1) the aridity classes defined 234 above, (2) the distribution of three common tree species (because a secondary objective of the fieldwork 235 was to study the natural regeneration of tree species), (3) results from previous remote sensing 236 modelling similar to those presented here (Cobián de Vinatea, 2019), and (4) accessibility, we selected 237 93 sampling locations (Figure 1) (see Supplementary Material 2 for details on the sampling design). At 238 each of these locations, we set up two plots of 20x10 m, the first plot as close as possible to the selected 239 coordinates, and the second within a close distance (ca. 100 m) but aiming to capture local variability 240 in forest structure. In each of the transects, we identified all adult woody individuals (i.e., with diameter 241 at breast height (DBH) > 5 cm for trees and diameter at ground level (DGL) > 5 cm for shrubs) and 242 measured their DBH for trees and DGL for shrubs. The AGB of the measured trees was estimated using 243 the pantropical allometric equation proposed by Chave et al. (2014), while shrub biomass was estimated 244 using the equation proposed by Návar (2014), which was validated in TDFs of Mexico and Puerto Rico. 245 . Formulas 1 and 2 were used to estimate the AGB of trees and shrubs, respectively: 10 246 𝐴𝐺𝐵𝑡𝑟𝑒𝑒 = exp[−1.803 − 0.976𝐸 + 0.976 𝑙𝑛(𝜌𝑤 ) + 2.673 𝑙𝑛(𝐷𝐵𝐻) − 0.0299[𝑙𝑛(𝐷)]2] [1] 247 𝐴𝐺𝐵𝑠ℎ𝑟𝑢𝑏 = 0.059 × 𝐷𝐺𝐿2.43 [2] 248 With E representing a measure for environmental stress (depending on water stress and temperature 249 seasonality; see Chave et al., 2014) and 𝜌w representing the species-specific wood density, which was 250 sourced from secondary data such as the Global Wood Density Database (Chave et al., 2009). 251 AGB data were collected in 186 plots. Five of the plots did not have woody plant individuals and were 252 therefore not used for training the AGB models (see section 0). Similarly, one plot had a tree with a 253 diameter higher than 80 cm and its crown surface was larger than the plot surface (200 m2), so it was 254 not used for training the AGB models either. 255 2.2.3 Predictor variables 256 To model the WC and AGB of the forests in our study region, we prepared a total of 123 predictor 257 variables (Table 1), consisting of multispectral optical data (Sentinel-2) and synthetic aperture radar 258 (SAR) data (Sentinel-1), the details of which are given in the following. 259 Multispectral optical imagery (Sentinel-2) 260 Multispectral Sentinel-2 images were obtained from the Copernicus Open Access Hub 261 (https://scihub.copernicus.eu/) as Level-2A products, i.e. geometrically and atmospherically corrected 262 bottom-of-atmosphere reflectance, at a spatial resolution of 20 m. Given the complex relief of 263 Lambayeque, we further pre-processed them by applying a topographic correction, using the C- 264 correction method (Teillet et al., 1982) implemented in the ‘RStoolbox’ package for R (Leutner et al., 265 2019). We combined images from different sensing dates in 2019 and 2020 (Table S1, Supplementary 266 Material 3) to create median reflectance composite images for each season covering the entire study 267 region. Seasons considered here were the wet season, dry season, and the transition season (i.e. shortly 268 after the wet season).. The sensing dates that were combined for each of the seasons are given in Table 269 S1 (Supplementary Material 3). The Sentinel-2 bands were used to derive a series of spectral indices 270 (Table 1). Taking advantage of the strong seasonality in the study region, we also calculated the 11 271 minimum, maximum, and range of NDVI values across the three seasons. Further, we used the ‘glcm’ 272 package for R (Zvoleff, 2019) to calculate two image texture metrics, the mean and the variance, from 273 the NDVI images for each of the three seasons, using a grey-level co-occurrence matrix (Haralick, 274 1979) with 16 grey levels, window sizes of 5x5 and 7x7 pixels, and all four possible directions (0°, 45°, 275 90°, and 135°) (Table 1). 276 Synthetic aperture radar imagery (Sentinel-1) 277 We downloaded Sentinel-1 synthetic aperture radar (SAR) imagery as dual-polarized (i.e., vertical- 278 vertical (VV) and vertical-horizontal (VH)) data in Interferometric Wide-Swath Mode at a spatial 279 resolution of 20 m from the Copernicus Open Access Hub. Single-date VV and VH backscatter imagery 280 was downloaded for each of the three seasons (Table 1). Analogous to the Sentinel-2 data, we used data 281 from each of the three seasons (wet, dry and transitions season; see Table S1 in Supplementary Material 282 3 for sensing dates). Contrary to the Sentinel-2 data, the retrieved Sentinel-1 SAR data were at Level-1 283 and therefore still required several pre-processing steps. The following pre-processing steps, in this 284 order, were applied using the Sentinel Application Platform toolbox (ESA, 2019): application of an 285 orbit file, radiometric calibration, speckle filtering, radiometric terrain flattening, and Range-Doppler 286 geometric terrain correction. We also calculated two image texture metrics (mean and variance) from 287 each of the VV and VH images in the same way as the NDVI texture metrics mentioned above (Table 288 1). 289 12 290 Table 1: Remote sensing predictor variables included for modelling woody cover (WC) and aboveground biomass 291 (AGB). The Sentinel-2 bands and indices are median composites (see Table S1 in Supplementary Material 3 for 292 sensing dates). Sensor Bands, indices or parameters Definition Sentinel-2 Bands (wet, B02 Blue, 490 nm transition, B03 Green, 560 nm dry season) B04 Red, 665 nm B05 Red edge, 704 nm B06 Red Edge, 749 nm B07 Red Edge, 783 nm B8A Near Infrared (NIR), 865 nm B11 Short-wave Infrared (SWIR), 1610 nm B12 Short-wave Infrared (SWIR), 2190 nm Indices NDVI (Tucker, 1979) (B8A-B04)/(B8A+B04) (wet, transition, GDVI (Wu, 2014) (B8A2-B042)/(B8A2+B042) dry season) GNDVI (Gitelson et al., 1996) (B8A-B03)/(B8A+B03) EVI (Huete et al., 2002) 2.5*(B8A-B04)/((B8A+6.0*B04- 7.5*B02)+1.0) NDSVI (Qi et al., 2000) (B11-B04)/(B11+B04) SAVI (Huete, 1988) ((B8A-B04)/(B8A+B04+0.5))*(1+0.5) MSAVI (Qi et al., 1994) (2*B8A+1-(((2*B8A+1)2-8*(B8A-B04)))1/2)/2 SATVI (Marsett et al., 2006) ((B11-B04)/(B11+B04+1))*(1+1)-B12/2 MTCI (Dash and Curran, 2004) (B06-B05)/(B05-B04) CIr (Gitelson et al., 2003) (B07/B05)-1 IRECI (Frampton et al., 2013) (B07-B04)/(B05/B06) NDWI (Gao, 1996) (B8A-B11)/(B8A+B11) MNDWI (Xu, 2006) (B03-B8A)/(B03+B8A) NDBI (Zha et al., 2003) (B12-B8A)/(B12+B8A) NDTI (Van Deventer et al., (B11-B12)/(B11+B12) 1997) Seasonal Min. NDVI Min., max. and range of VV and VH across the Max. NDVI three seasons (wet, transition, dry season) Range NDVI Texture NDVI 5x5 mean Grey-level co-occurrence matrix (GLCM) (wet, NDVI 5x5 variance mean and variance of NDVI backscatter, using transition, NDVI 7x7 mean 5x5 and 7x7 pixel windows dry season) NDVI 7x7 variance Sentinel-1 Polarization VV Vertical transmit-vertical channel (wet, VH Vertical transmit-horizontal channel transition, dry season) Seasonal Min, max., range of VV Min., max. and range of VV and VH across the Min., max., range of VH three seasons (wet, transition, dry season) Texture VV 5x5 mean Grey-level co-occurrence matrix (GLCM) (wet, VV 5x5 variance mean and variance of VV and VH backscatter, transition, VH 5x5 mean using 5x5 and 7x7 pixel windows dry season) VH 5x5 variance VV 7x7 mean VV 7x7 variance VH 7x7 mean VH 7x7 variance 293 13 294 2.2.4 Model calibration and evaluation 295 Next, we used the predictor variables described above to model the WC and AGB in our study region. 296 To evaluate the usefulness of including Sentinel-1 data and multi-season data, we constructed models 297 with and without Sentinel-1 data using data from each of the three seasons and for all seasons together, 298 resulting in eight models for both WC and AGB. 299 Modelling was carried out using boosted regression trees (BRTs), which consist of an ensemble of 300 decision trees that are built in a sequential manner. BRTs are able to handle a large number of correlated 301 predictor variables and to model non-linear relationships and complex interactions between predictor 302 variables (Breiman, 2001; Hastie et al., 2001). BRTs were fitted using the ‘gbm’ package for R 303 (Greenwell et al., 2019) while tuning the hyperparameters through cross-validation using the ‘caret’ 304 package, varying (1) the learning rate, (2) the number of trees, (3) tree depth, and (4) the minimum 305 number of observations in the terminal nodes. The learning rate was varied between 0.001 and 0.15, the 306 number of trees between 500 and 1500, tree depth between 2 and 9, and the minimum number of 307 observations in the terminal nodes between 5 and 30. 308 To avoid overly optimistic cross-validation results due to spatial autocorrelation (Roberts et al., 2017; 309 Schratz et al., 2019) and to tune the hyperparameters in such a way to optimize model transferability 310 (Muscarella et al., 2014; Roberts et al., 2017), we used spatial blocks of 30x30 km to split the model 311 calibration data in training and testing data, using the ‘blockCV’ package for R (Valavi et al., 2019). 312 For each of the models, the data were split in ten folds, which was repeated five times with different 313 spatial configurations of the spatial blocks (i.e. a 5-repeated 10-fold cross validation). During cross- 314 validation, the forest extent models were evaluated using the overall accuracy metric (i.e. correctly 315 classified predictions divided by the total number of predictions), which was considered appropriate 316 given the balanced training dataset. The WC and AGB models were evaluated using the mean absolute 317 error (MAE), i.e. the mean difference between predicted and observed values, which was preferred over 318 the root mean squared error (RMSE) for its lower sensitivity against outliers and its straightforward 319 interpretation. To make the MAE values of the WC and AGB models comparable, we also calculated 320 normalized MAE values by dividing the MAE by the interquartile range of the WC and AGB values 14 321 used for model calibration. The final predictions were always made using the complete set of training 322 data and the tuned hyperparameter settings. 323 2.2.5 Site-specific scaling of woody cover and aboveground biomass 324 The TDFs of Lambayeque are a highly heterogeneous ecosystem, ranging from the hyper-arid TDFs 325 bordering the Sechura desert in the West to the semi-arid montane TDFs bordering sub-humid cloud 326 forests in the East (see section 2.1). Under such heterogeneous and harsh environmental conditions, it 327 is clear that these forests cannot all support the same WC and AGB levels, even when non-degraded. 328 Without benchmark WC and AGB values representative for the local environmental conditions, regions 329 with low potential WC and AGB could therefore be confused with degraded areas. To estimate the 330 spatially explicit potential WC and AGB across our study region, we followed a similar approach as in 331 the LNS method (Prince, 2004; Prince et al., 2009) by defining environmentally homogeneous land 332 capability classes that are expected to be able to support similar WC and AGB levels, using the 333 clustering approach described below. Methodological details about the site-specific scaling used to 334 obtain the WC and AGB reduction maps are given below and summarized in Figure 3. 335 As potential evapotranspiration strongly exceeds annual precipitation throughout the study region, we 336 considered the amount of water available for plant growth to be the main factor limiting WC and AGB, 337 and therefore defined the land capability classes by clustering variables serving as proxies for water 338 availability. Aridity, elevation, longitude, and latitude were included to obtain land capability classes 339 with homogeneous climatic conditions. To account for differences in soil drainage and run-off as 340 influenced by soil texture and topography, we used soil sand content (Hengl et al., 2017) and the SAGA 341 Wetness Index (Conrad et al., 2015), the latter of which was derived from the ALOS-PALSAR digital 342 elevation model, downloaded from the Alaska Satellite Facility (asf.alaska.edu). Based on these 343 variables, we subdivided the modelled forest extent in a set of clusters using the Clustering Large 344 Applications (CLARA) method, an extension of the partitioning-around-medoids (PAM) method for 345 large datasets (Kaufman and Rousseeuw, 1990), implemented in the ‘cluster’ package for R (Maechler 346 et al., 2019). The optimal number of clusters was determined with the ‘optCluster’ package for R using 347 four cluster stability measures: average proportion of non-overlap, average distance, average distance 15 348 between means, and figure of merit (Sekula et al., 2020), all of which are indicators of how much 349 clustering results change when one of the clustering variables is removed. The number of clusters was 350 varied between 5 and 75 in intervals of 5, and 55 clusters was determined as the optimal number of 351 clusters. All variables were normalized prior to clustering to ensure equal weighting. 352 The next step was to obtain potential WC and AGB values for each of the land capability classes, for 353 which we used the 95th percentile of the predicted WC and AGB values within each class. The 354 degradation status of each pixel was then estimated by calculating the percentage WC and AGB 355 reduction relative to these potential WC and AGB values. All pixels above this threshold were 356 considered non-degraded. The use of the 95th percentile should reduce the effect of outlying WC and 357 AGB estimates on estimating these benchmark values (Prince et al., 2009), for example as a 358 consequence of wrongly including some irrigated agricultural areas in the forest extent map. However, 359 it relies on the assumption that at least 5% of the pixels within each land capability class is sufficiently 360 non-degraded to be used as estimates of potential WC and AGB. To assess the sensitivity of our results 361 to the choice of this threshold, we also repeated the analysis using the 90th percentile. 16 362 363 Figure 3: Flowchart showing a summary of the methodology followed to obtain woody cover (WC) and 364 aboveground biomass (AGB) reduction maps. CLARA stands for the Clustering Large Applications method. Note 365 that this flowchart provides a simplified overview without (pre-)processing of predictor variables. 366 17 367 3 Results 368 A summary of the collected woody cover (WC) and aboveground biomass (AGB) data is given in Table 369 2. Mean WC estimates ranged from 15% in the hyper-arid zone to 71% in the semi-arid zone, while 370 median WC estimates ranged from 10% in the hyper-arid zone to 87% in the semi-arid zone. Mean 371 AGB values were lower in the arid and semi-arid zone (42.6 and 42.0 t/ha) than in the hyper-arid zone 372 (56.9 t/ha), the latter being a consequence of the inclusion of a few well-conserved riverine forests with 373 exceptionally high AGB levels. Median AGB values were lowest in the hyper-arid zone (21.6 t/ha) and 374 highest in the semi-arid zone (29.2 t/ha). Livestock dung and cut stems were encountered in 56% and 375 47% of the plots, respectively, illustrating the large human pressure on these forests. 376 Table 2: Summary of reference data used to model woody cover (WC) and aboveground biomass (AGB) in 377 different aridity zones of the tropical dry forests of Lambayeque. n refers to the number of images used for WC 378 estimation and the number of inventoried plots for AGB. SD refers to standard deviation. Woody cover (%) Aboveground biomass (t/ha) Aridity zone n Mean WC (± SD) Median WC n Mean AGB (± SD) Median AGB Hyper-arid 257 15 (± 15) 10 35 56.9 (± 84.0) 21.6 Arid 241 43 (± 21) 43 62 42.6 (± 55.6) 24.5 Semi-arid 222 71 (± 35) 87 83 42.0 (± 39.6) 29.2 Total 720 42 (± 34) 34 180 45.1 (± 56.1) 26.3 379 Figure 4 presents the results of the cross-validation of WC and AGB models. The mean absolute error 380 (MAE) of the WC models ranged from 16.0% (multi-season model using both Sentinel-1 and Sentinel- 381 2 data) to 18.0% (wet season model using only Sentinel-2 data). Including Sentinel-1 data in addition 382 to Sentinel-2 data consistently improved model accuracy, but only with a mean difference in MAE of 383 0.7%. Slightly larger differences were found between the multi-season and single season models, with 384 a mean difference in accuracy of 1.1%. 385 The AGB models followed the same trends as the WC models (Figure 5), but differences in predictive 386 accuracy were even smaller and not larger than the variability observed across repetitions of the 10-fold 387 cross validation with different configurations of spatial blocks (indicated by error bars in Figure 4), 388 except for the considerably lower performance of the transition season model. The most accurate model 389 was the multi-seasonal model combining both Sentinel-1 and Sentinel-2 data, with a cross-validated 390 MAE of 28.6 t/ha. 18 391 392 Figure 4: Cross-validated accuracy of (a) woody cover (WC) and (b) aboveground biomass (AGB) predictions 393 using boosted regression trees and data from different sources (Sentinel-2 only and Sentinel-2 + Sentinel-1) and 394 different seasons (multi-season, dry season, transition season, and wet season). The MAE (mean absolute error) 395 values shown are those obtained from the models with tuned hyperparameter settings (see section 0). The error 396 bars indicate the range between accuracy metric values obtained between the five repetitions of the 10-fold cross- 397 validation, each with a different configuration of spatial blocks (see section 0). 398 After normalizing the MAE values by dividing them by the interquartile range of the WC and AGB 399 values used for model calibration, the most accurate WC and AGB models presented a normalized MAE 400 of 26% and 65%, respectively, indicating that WC was mapped with much greater accuracy than AGB. 401 The poor accuracy of the AGB predictions was confirmed by the R2 value of the final models, which 402 was 0.81 for WC and 0.27 for AGB. 403 The maps with predicted and potential WC and AGB and the corresponding WC and AGB reduction 404 estimates are shown in Figure 5, while an illustration of the WC reduction mapping at 3 locations is 405 given in Figure 6. As expected, both predicted and potential WC and AGB showed an increase from 406 the south-west to the north-east (Figure 5), corresponding to the aridity gradient spanning our study 19 407 region. Average predicted WC ranged from 18% in the hyper-arid zone to 67% in the semi-arid zone 408 (Table 3). Overall, 11% of our study region was predicted to have WC values lower than 10%, which 409 would not be considered forest under the classical definition (FAO, 2002). When compared with the 410 benchmark values, determined as the 95th percentile values in each of the land capability classes, 411 average WC reduction was highest in the hyper-arid zone (47%) and lowest in the semi-arid zone (20%). 412 Average predicted AGB ranged from 17.7 t/ha in the hyper-arid zone to 36.2 t/ha in the semi-arid zone. 413 Average AGB reduction was highest in the arid zone (43%) and lowest in the hyper-arid zone (29%). 414 Overall, average AGB reduction (36%) was slightly higher than average WC reduction (35%) (Table 415 3). The results of the same analysis using the 90th percentile for the benchmark values are available in 416 Table S2 (Supplementary Material 3). 417 Table 3: Predicted mean current woody cover (WC) and aboveground biomass (AGB), average potential WC and 418 AGB, and average WC and AGB reduction for different aridity zones. The values shown here correspond to the 419 maps generated by the most accurate WC and AGB models (see Figure 4). They are based on the 95th percentile 420 WC and AGB values within each of the land capability classes as benchmark values; the results of the same 421 analysis using the 90th percentiles are given in Table S2 (Supplementary Material 3). Woody cover (WC) Aboveground biomass (AGB) Aridity Current Potential Reduction Current Potential Reduction zone (%) (± SD) (%) (± SD) (%) (± SD) (t/ha) (± SD) (t/ha) (± SD) (%) (± SD) Hyper-arid 18 (± 14) 37 (± 20) 47 (± 22) 17.7 (± 7.3) 27.4 (± 13.2) 29 (± 21) Arid 47 (± 18) 69 (± 12) 33 (± 21) 21.4 (± 8.3) 38.4 (± 10.0) 43 (± 19) Semi-arid 67 (± 13) 85 (± 3) 20 (± 14) 36.2 (± 15.2)) 53.9 (± 9.8) 34 (± 22) Total 41 (± 24) 61 (± 24) 35 (± 22) 23.4 (± 12.3) 37.9 (± 14.9) 36 (± 21) 422 20 423 424 Figure 5: Mapping results of woody cover (WC) and aboveground biomass (AGB) (panels a and b), potential 425 WC and AGB (panels c and d), and WC and AGB reduction (panels f and g). The maps shown here are those 426 obtained from the most accurate forest extent, WC, and AGB models (see Figure 4). The potential WC and AGB 427 maps are based on the 95th percentile WC and AGB values within each of the land capability classes. 21 428 429 Figure 6: Illustration of woody cover reduction mapping at three locations with varying degrees of degradation. 430 Panel (a) shows the very high-resolution Google Earth image chips of these locations, panel (b) shows the 431 estimated woody cover reduction. Note that woody cover estimates of individuals pixels reflect the woody cover 432 of a 95x95 m window rather than of the individual pixels. 433 22 434 4 Discussion 435 4.1 Mapping woody cover and aboveground biomass 436 The inclusion of SAR data (Sentinel-1) in addition to optical data (Sentinel-2) consistently improved 437 the predictive accuracy of the woody cover (WC) models and aboveground biomass (AGB) models. 438 Although differences in accuracy were small, this confirms the potential of combining optical data with 439 SAR data in TDF and savannah ecosystems for WC (Anchang et al., 2020; Baumann et al., 2018; 440 Higginbottom et al., 2018; Zhang et al., 2019) and AGB mapping (Forkuor et al., 2020; Pötzschner et 441 al., 2022). Similarly, both WC and AGB were most accurately modelled using multi-seasonal data, in 442 line with findings in TDF and savannah ecosystems elsewhere (Forkuor et al., 2020; Higginbottom et 443 al., 2018; Karlson et al., 2015; Van Passel et al., 2020). Considering the single season models, WC was 444 most accurately modelled using dry season data, corroborating findings in TDF and savannah 445 ecosystems elsewhere (Higginbottom et al., 2018; Urbazaev et al., 2015; Van Passel et al., 2020). This 446 is presumably due to higher contrasts between woody vegetation and herbaceous vegetation during the 447 dry season. Among the single season models, the wet season model resulted in the most accurate AGB 448 predictions, contrary to the findings of Forkuor et al. (2020), who obtained more accurate predictions 449 using dry season data. However, the small differences between seasons and poor accuracy of the AGB 450 models warn against overinterpreting these results. 451 Contrary to studies that have recommended the use of transition season data for mapping WC or AGB 452 in savannah or TDF ecosystems (Gasparri et al., 2007; Nagelkirk and Dahlin, 2020), the transition 453 season was the season with the poorest performance. However, this may be related to the difficulty we 454 had to generate a cloud-free composite for this season, which resulted in a mosaic of images spanning 455 almost two months and therefore also spanning a large phenological window, which may have 456 confounded our predictions. Our field observations support this hypothesis: Towards the end of the field 457 campaign, which was carried out in the transition season, some species such as Bursera graveolens and 458 Loxopterygium huasango had already lost much of their leaves, and such phenological variability might 459 add noise to the data used for the AGB estimations in this season. 23 460 The AGB models had a considerably lower accuracy than the WC models. While part of this difference 461 can be explained by the lower number of training points for the AGB model (180 vs. 720) and the less 462 direct physical relationship between AGB and reflectance, it is likely also related to the small plot size 463 in which AGB was measured (10x20m), as such small plot sizes are likely to exacerbate sampling errors 464 due to higher plot perimeter-to-area ratio (Hernández-Stefanoni et al., 2018). Apart from the usual 465 challenges associated to remote sensing assessments in TDFs, such as mixed pixels (containing both 466 vegetation and soil), other challenges we encountered to obtain accurate AGB estimations include a 467 large variability in tree species compositions and phenology associated with the large aridity gradients 468 spanning the study region, and a dieback of Prosopis spp. (Fabaceae) trees in many parts of 469 Lambayeque, caused at least partly by a defoliating insect (Whaley et al., 2020). While already dead 470 trees were not included in the measurements, affected Prosopis trees that only lost part of their leaves 471 were measured, as the dieback is a slow and reversible process (Whaley et al., 2020). As such affected 472 trees were present in 18% of the sampled plots and Prosopis trees account for most of the AGB in many 473 of the lowland TDFs (Padrón and Navarro-Cerrillo, 2007), this has likely confounded our predictions. 474 Our results indicated that average current AGB levels in the TDFs of Lambayeque range between 17.7 475 t/ha in the hyper-arid zone to 36.2 t/ha in the semi-arid zone. While these low values are partly the 476 consequence of widespread forest degradation, our estimates of potential AGB levels of individual land 477 capability classes (ranging from 16.0 to 61.8 t/ha) are also lower than potential AGB levels in TDF 478 ecosystems elsewhere, which is related to the extremely low precipitation levels of the TDFs of 479 Lambayeque. A global review found that AGB in mature TDFs ranged from 39 t/ha to 334 t/ha 480 (Becknell et al., 2012). However annual precipitation in the sampling sites considered by the authors 481 ranged from around 600 to 2000 mm, whereas the average annual precipitation in our study region is 482 only 120 mm, never exceeding 600 mm even at the highest elevations (Fick and Hijmans, 2017), 483 although these forests also receive some additional moisture through fog coming from the Pacific 484 Ocean. 24 485 4.2 Forest degradation status 486 While the TDFs of Lambayeque still maintain around 74% of their original extent (assuming all areas 487 with an elevation below 1600 masl and outside the Sechura desert were once covered by TDF; see 488 Supplementary Material 1), our results show that the remaining forests are strongly affected by forest 489 degradation, with an average WC reduction of 35% and an average AGB reduction of 36% (Table 3, 490 Figure 5). WC reduction was most pronounced in the hyper-arid zone in the lowlands of Lambayeque, 491 which may be related to higher densities of free-roaming goats, which are less common at higher 492 elevation. Other potential factors are easier accessibility, as most of the zone has a relatively flat relief, 493 and higher population density, as agriculture is mostly concentrated in the valleys of the hyper-arid 494 zone.. AGB reduction was most pronounced in the arid zone, roughly corresponding to the hilly TDFs. 495 This may be related to the fact that these TDFs are often dominated by Loxopterygium huasango and 496 Bursera graveolens, both of which have undergone extensive selective logging and have both been 497 classified as ‘critically endangered’ by the Peruvian government (El Peruano, 2006). 498 4.3 Potential applications for forest restoration planning 499 Compared to conventional trend-based forest degradation assessments, the site-specific scaling of WC 500 and AGB has the advantage that it allows estimating forest degradation status in relation to benchmark 501 WC and AGB values irrespective of the time when this degradation occurred. While not an alternative 502 to measuring changes in forest degradation status, its main advantage lies in the potential to detect 503 degradation signals that go deeper in time than the Landsat record (or other medium to high resolution 504 imagery), hence allowing to map forest degradation status in regions that have a long history of human 505 pressure. In addition, the site-specific scaling approach allows characterizing degradation as a 506 continuous variable (i.e. percentage reduction of WC and AGB as compared to benchmark WC and 507 AGB levels). Such quantitative appraisal of degradation is more easily understood by policy makers 508 and restoration planners, while trend-based assessments often result in binary results (Cohen et al., 509 2017). To facilitate the use of our results, the maps are made available in a user-friendly online tool 510 (https://bioversityinternational.shinyapps.io/restoration_priorities_Lambayeque), along with other 511 spatially explicit indicators that may be used to prioritize areas for restoration (e.g., distance to river, 25 512 distance to protected area, slope). This tool can then be used as one of multiple inputs to guide the 513 restoration priority setting process, in which also socio-economic aspects should be considered (e.g., 514 cost-benefit ratios, land tenure), ideally involving a wide range of relevant stakeholders (IUCN et al., 515 2014). 516 Using the WC and AGB reduction maps as estimates for forest degradation, assisted natural 517 regeneration may be prioritized in areas with intermediate degrees of degradation, as higher seed 518 availability and less harsh conditions in these areas are expected to favour seedling establishment 519 (Chazdon, 2017; Vieira and Scariot, 2006), while active tree planting may be prioritized in the most 520 heavily degraded areas, where probabilities of natural regeneration are lower. Furthermore, as forest 521 restoration usually aims at recovering forest attributes to a reference state, the benchmark WC and AGB 522 values can be used as quantitative targets for forest restoration initiatives. While conventional trend- 523 based assessments can only track relative degradation trends, the site-specific scaling approach can be 524 used to track absolute degradation trends through time (Jackson and Prince, 2016a; Li et al., 2020). This 525 allows to monitor the recovery towards these targets and provide insights on forest resilience, which is 526 not captured by quantifying WC and AGB reduction at a specific point in time. Another possible 527 application of the WC and AGB reduction maps is the selection of potential seed sources for tree 528 planting activities: these should be selected in non-degraded areas to minimize the risks of potential 529 erosion of genetic diversity that may have resulted from long-term degradation (Thomas et al., 2014). 530 4.4 Woody cover and aboveground biomass benchmark values 531 The reliability of the results of our site-specific scaling approach depends on the degree to which the 532 environmental conditions within specific land capability classes support similar WC and AGB levels. 533 Analogous to the local net productivity scaling (LNS) approach proposed by Prince (2004), degradation 534 will be overestimated if these land capability classes include WC or AGB hotspots due to more 535 favourable site conditions (Prince et al., 2009). The use of the 95th percentile as benchmark value should 536 reduce the interference from such WC and AGB outliers, but does not entirely preclude this risk. 537 Conversely, the use of the 95th percentile assumes that at least 5% of the grid cells within an land 538 capability class is sufficiently non-degraded to be used as a benchmark value. If this not the case, or 26 539 when the most undegraded forests occur in the most marginal sites, degradation may be underestimated, 540 which can be seen as an example of the shifting baseline syndrome (Pauly, 1995). Given the pervasive 541 human influence in the TDFs of northwestern Peru (Fremout et al. 2020), we chose the 95th percentile 542 rather than 90th percentile used by Prince et al. (2009), but the presence of non-degraded areas in each 543 of the land capability classes remains a strong assumption. 544 To evaluate to what extent these reference areas with WC and AGB values above the 95th percentile 545 thresholds correspond with non-degraded areas, long-term exclosures could be established to evaluate 546 what happens in these areas without any human disturbance, although this would also change the impact 547 of native fauna on the vegetation. While it remains difficult to assess which areas are truly undegraded, 548 especially when ecosystems may have shifted towards an alternative stable state (Ghazoul et al., 2015), 549 such exclosures would at least allow assessing the stability of the current state and therefore be useful 550 to inform restoration targets. As an alternative, and faster, approach to evaluate to what extent the forests 551 used as a reference correspond to natural old-growth forests, the use of tree-ring based metrics (such as 552 the range of canopy tree age) seems promising, an approach already successfully used to quantify the 553 naturalness of old growth forests in Europe (Di Filippo et al., 2017). 554 Using a space-for-time substitution, the local scaling approach allows identifying degradation that 555 occurred before medium-to-high resolution satellite data became available. Regardless, the benchmark 556 WC and AGB values we used are likely to reflect a relatively recent baseline situation, as the impact of 557 centuries of human occupation and livestock grazing on forest structure is largely unknown. The 558 benchmark values may be even less representative for the baseline of past millennia, where aspects such 559 as paleoclimatic changes (Smith and Mayle, 2021), changes in fire regimes (Power et al., 2008), and 560 human-driven megafauna extinctions (Malhi et al., 2016; Sandom et al., 2014) become relevant. While 561 their impacts on vegetation are still uncertain and subject to debate, there is evidence that some regions 562 were characterized by more open vegetation structure before these megafauna extinctions (Malhi et al., 563 2016; Sandom et al., 2014). 564 Degradation estimates were considerably smaller using the 90th percentile (around 14-22% smaller; see 565 Table S2, Supplementary Material 3), indicating that the choice of this threshold is not unimportant, 27 566 especially when the focus is on absolute estimates of degradation; here we were more interested in 567 relative differences to prioritize areas for restoration. While the choice of this threshold will always be 568 somewhat arbitrary (Prince et al., 2009), it can be informed by expert knowledge on the extent of 569 degradation. There is a similar trade-off between the number of land capability classes selected and the 570 likelihood that each class contains non-degraded vegetation areas: Higher numbers result in lower 571 within-class variability in site conditions, which decreases the risk of the occurrence of WC or AGB 572 hotspots, but increases the risk that there are not enough non-degraded areas within the land capability 573 class to be used as a benchmark. 574 4.5 Other aspects of forest degradation 575 Our approach provides a forest structure perspective to assessing forest degradation, but it is important 576 to note that there are also other dimensions to forest degradation (Putz and Redford, 2010; Thompson 577 et al., 2012). For example, some land managers may be more interested in the decline of forest 578 productivity, which can be assessed with the original LNS approach (Prince, 2004). Similarly, 579 community ecologists may be more interested in changes in forest community composition, for example 580 observed greening trends (usually interpreted as vegetation recovering from degradation) in Senegal 581 have been found to be associated with an impoverishment of the vegetation (Herrmann and Tappan, 582 2013). In our study region, widespread livestock grazing and selective logging (livestock dung and cut 583 stems were encountered in 56% and 47% of the plots, respectively) are likely to also have a considerable 584 impact on forest composition (Jara-Guerrero et al., 2021). However, such changes are much more 585 difficult to quantify using remote sensing data (but see for example Baena et al., 2017). Other 586 dimensions of forest degradation such as defaunation (i.e. the loss of local fauna) are almost impossible 587 to evaluate using remote sensing, but they should not be overlooked when assessing forest degradation 588 status. For example, defaunation can lead to long-term vegetation degradation, as essential ecosystem 589 functions in so-called ‘empty forests’ (Redford, 1992) such as seed dispersal, pollination and predator- 590 prey relationships may collapse (Culot et al., 2017; Dirzo et al., 2014; Peres et al., 2016). Also in the 591 TDFs of Lambayeque, several important seed dispersers (e.g., white-tailed deer (Odocoileus 592 virginianus), spectacled bear (Tremarctos ornatus), white-winged guan (Penelope albipennis) are 28 593 highly threatened.. While we expect such other dimensions of forest degradation to be correlated with 594 our degradation estimates, further research is needed to confirm this hypothesis. 595 4.6 Transferability 596 As WC and AGB can be mapped in forest ecosystems around the world using a variety of sensors, our 597 general approach of site-specific scaling of WC and AGB to estimate forest degradation should be easily 598 applicable elsewhere too, as long as sufficient non-degraded areas exist along a range of environmental 599 conditions, and as long as potential WC and AGB levels can be described by land capability classes. 600 The choice of variables to construct these land capability classes should be tailored to the ecosystem. 601 For example in rainforests where water availability is not a limiting factor, soil characteristics are likely 602 to play a more important role in shaping potential AGB patterns (Santiago-García et al., 2019; van der 603 Sande et al., 2018). 604 The brightness thresholding approach we used to map WC, as proposed by Nagelkirk and Dahlin (2020) 605 in African savannas, proved to be transferable to the TDFs of our study region and to be a cost-effective 606 alternative to field measurements of WC. Yet, transferability to other forest ecosystems depends on the 607 contrast between the woody vegetation and the herb layer, making it less suitable for more humid forest 608 ecosystems where the contrast between both is limited during a larger part of the year. This is further 609 illustrated by the difficulties we had to threshold some of the images in the montane TDFs in our study 610 region. 611 As the inclusion of multi-seasonal data and inclusion of Sentinel-1 data only led to small improvements 612 in accuracy, using single-season Sentinel-2 data could make our approach easier to implement. 613 However, this does not seem to be a general trend, considering that Forkuor et al. (2020) obtained 614 stronger improvements in accuracy when evaluating the usefulness of multi-seasonal data and the 615 addition of Sentinel-2 data when mapping AGB in West-African dry forests. In regions where it is 616 difficult to obtain cloudless composites, using only Sentinel-1 or other SAR data could even be the 617 better option. 29 618 Overall, the site-specific scaling of WC and AGB allows estimating forest degradation status 619 irrespective of the time when this degradation occurred, on the condition that there are still some 620 sufficiently undegraded areas that can be used as a benchmark. The approach can be used to prioritize 621 forest restoration actions in environmentally heterogeneous regions suffering the impacts of long-term 622 degradation, in TDFs and beyond. 623 Author responsibilities: TM, BM, and ET conceived and designed the research; TF, JC, and WH 624 collected the data; TF and JC analyzed the data; TF drafted the manuscript; TF, JC, MSV, DLF, PNB, 625 WH, RA, EC, BM, ET edited the manuscript. 626 Data availability 627 The data to train the WC and AGB models are available at [insert data repository], as well as the R 628 script of the boosted regression tree models and the CLARA clustering. 629 Acknowledgements 630 This work was supported by the Flemish Interuniversity Council (VLIR-UOS) (grant number 631 NDOC2016PR002), the German Federal Ministry of Economic Cooperation and Development (BMZ), 632 commissioned and administered through the Deutsche Gesellschaft für Internationale Zusammenarbeit 633 (GIZ) Fund for International Agricultural Research (FIA) (grant number 8121944), and the CGIAR 634 Fund Donors (https://www.cgiar.org/funders/). We thank Walter Cespedes for support with the field 635 measurements and the Gerencia Regional de Recursos Naturales of the local government of 636 Lambayeque for supporting the transport to the sampling sites, with a special thanks to Regis Reyes for 637 organizing the logistic support. We thank Koenraad Van Meerbeek, Marielos Peña-Claros, Stefaan 638 Dondeyne and Raf Aerts for their useful feedback on the manuscript. 639 Literature cited 640 Anchang, J.Y., Prihodko, L., Ji, W., Kumar, S.S., Ross, C.W., Yu, Q., Lind, B., Sarr, M.A., Diouf, A.A., 641 Hanan, N.P., 2020. Toward Operational Mapping of Woody Canopy Cover in Tropical Savannas 642 Using Google Earth Engine. Front. Environ. Sci. 8, 4. https://doi.org/10.3389/fenvs.2020.00004 643 Baena, S., Moat, J., Whaley, O., Boyd, D.S., 2017. Identifying species from the air: UAVs and the very 644 high resolution challenge for plant conservation. PLoS One 12, e0188714. 30 645 https://doi.org/10.1371/journal.pone.0188714 646 Barthelme, S., 2020. imager: Image Processing Library Based on “CImg”. R package version 0.42.1. 647 Baumann, M., Levers, C., Macchi, L., Bluhm, H., Waske, B., Gasparri, N.I., Kuemmerle, T., 2018. 648 Mapping continuous fields of tree and shrub cover across the Gran Chaco using Landsat 8 and 649 Sentinel-1 data. Remote Sens. Environ. 216, 201–211. https://doi.org/10.1016/j.rse.2018.06.044 650 Becknell, J.M., Kissing Kucek, L., Powers, J.S., 2012. Aboveground biomass in mature and secondary 651 seasonally dry tropical forests: A literature review and global synthesis. For. Ecol. Manage. 276, 652 88–95. https://doi.org/10.1016/j.foreco.2012.03.033 653 Bernardino, P.N., De Keersmaecker, W., Fensholt, R., Verbesselt, J., Somers, B., Horion, S., 2020. 654 Global-scale characterization of turning points in arid and semi-arid ecosystems functioning. 655 Glob. Ecol. Biogeogr. 29, 1230–1245. https://doi.org/10.1111/geb.13099 656 Boer, M.M., Puigdefábregas, J., 2003. Predicting potential vegetation index values as a reference for 657 the assessment and monitoring of dryland condition. Int. J. Remote Sens. 24, 1135–1141. 658 https://doi.org/10.1080/0143116021000026788 659 Bontemps, J.D., Bouriaud, O., 2014. Predictive approaches to forest site productivity: Recent trends, 660 challenges and future perspectives. Forestry 87, 109–128. https://doi.org/10.1093/forestry/cpt034 661 Brandt, M., Hiernaux, P., Tagesson, T., Verger, A., Rasmussen, K., Diouf, A.A., Mbow, C., Mougin, 662 E., Fensholt, R., 2016. Woody plant cover estimation in drylands from Earth Observation based 663 seasonal metrics. Remote Sens. Environ. 172, 28–38. https://doi.org/10.1016/j.rse.2015.10.036 664 Breiman, L., 2001. Random forests. Mach. Learn. https://doi.org/10.1023/A:1010933404324 665 Buchhorn, M., Lesiv, M., Tsendbazar, N.E., Herold, M., Bertels, L., Smets, B., 2020. Copernicus global 666 land cover layers-collection 2. Remote Sens. 12, 1044. https://doi.org/10.3390/rs12061044 667 Burrell, A.L., Evans, J.P., Liu, Y., 2017. Detecting dryland degradation using Time Series Segmentation 668 and Residual Trend analysis (TSS-RESTREND). Remote Sens. Environ. 197, 43–57. 669 https://doi.org/10.1016/j.rse.2017.05.018 670 Cao, S., Yu, Q., Sanchez-Azofeifa, A., Feng, J., Rivard, B., Gu, Z., 2015. Mapping tropical dry forest 671 succession using multiple criteria spectral mixture analysis. ISPRS J. Photogramm. Remote Sens. 672 109, 17–29. https://doi.org/10.1016/j.isprsjprs.2015.08.009 673 Carmean, W.H., 1975. Forest site quality evaluation in The United States. Adv. Agron. 27, 209–269. 674 https://doi.org/10.1016/S0065-2113(08)70011-7 675 Carreiras, J.M.B., Pereira, J.M.C., Pereira, J.S., 2006. Estimation of tree canopy cover in evergreen oak 676 woodlands using remote sensing. For. Ecol. Manage. 223, 45–53. 677 https://doi.org/10.1016/j.foreco.2005.10.056 678 Cerrón, J., Fremout, T., Atkinson, R., Thomas, E., Cornelius, J., 2019. Experiencias de restauración y 679 fuentes semilleras en el bosque seco tropical del norte del Perú. Bioversity International, World 680 Agroforestry. Lima, Peru. https://doi.org/10.13140/RG.2.2.13126.63040 681 Chave, J., Coomes, D., Jansen, S., Lewis, S.L., Swenson, N.G., Zanne, A.E., 2009. Towards a 682 worldwide wood economics spectrum. Ecol. Lett. 12, 351–366. https://doi.org/10.1111/j.1461- 683 0248.2009.01285.x 684 Chave, J., Réjou-Méchain, M., Búrquez, A., Chidumayo, E., Colgan, M.S., Delitti, W.B.C., Duque, A., 685 Eid, T., Fearnside, P.M., Goodman, R.C., Henry, M., Martínez-Yrízar, A., Mugasha, W.A., 686 Muller-Landau, H.C., Mencuccini, M., Nelson, B.W., Ngomanda, A., Nogueira, E.M., Ortiz- 687 Malavassi, E., Pélissier, R., Ploton, P., Ryan, C.M., Saldarriaga, J.G., Vieilledent, G., 2014. 688 Improved allometric models to estimate the aboveground biomass of tropical trees. Glob. Chang. 689 Biol. 20, 3177–3190. https://doi.org/10.1111/gcb.12629 690 Chazdon, R.L., 2017. Landscape Restoration, Natural Regeneration, and the Forests of the Future. Ann. 691 Missouri Bot. Gard. 102, 251–257. https://doi.org/10.3417/2016035 31 692 Chen, Jun, Chen, Jin, Liao, A., Cao, X., Chen, L., Chen, X., He, C., Han, G., Peng, S., Lu, M., Zhang, 693 W., Tong, X., Mills, J., 2015. Global land cover mapping at 30 m resolution: A POK-based 694 operational approach. ISPRS J. Photogramm. Remote Sens. 103, 7–27. 695 https://doi.org/10.1016/j.isprsjprs.2014.09.002 696 Cobián de Vinatea, J., 2019. Mapping the forest structure of the seasonally dry tropical forests of 697 Lambayeque, Peru. MSc thesis. TU Dresden, Dresden, Germany. 698 Cohen, W.B., Healey, S.P., Yang, Z., Stehman, S. V., Brewer, C.K., Brooks, E.B., Gorelick, N., Huang, 699 C., Hughes, M.J., Kennedy, R.E., Loveland, T.R., Moisen, G.G., Schroeder, T.A., Vogelmann, 700 J.E., Woodcock, C.E., Yang, L., Zhu, Z., 2017. How Similar Are Forest Disturbance Maps 701 Derived from Different Landsat Time Series Algorithms? Forests 8, 98. 702 https://doi.org/10.3390/f8040098 703 Conrad, O., Bechtel, B., Bock, M., Dietrich, H., Fischer, E., Gerlitz, L., Wehberg, J., Wichmann, V., 704 Böhner, J., 2015. System for Automated Geoscientific Analyses (SAGA) v. 2.1.4. Geosci. Model 705 Dev. Discuss. 8, 2271–2312. https://doi.org/10.5194/gmd-8-1991-2015 706 Coulston, J.W., Moisen, G.G., Wilson, B.T., Finco, M. V., Cohen, W.B., Brewer, C.K., 2012. Modeling 707 Percent Tree Canopy Cover. Photogramm. Eng. Remote Sens. 78, 715–727. 708 https://doi.org/10.14358/PERS.78.7.715 709 Culot, L., Bello, C., Batista, J.L.F., Do Couto, H.T.Z., Galetti, M., 2017. Synergistic effects of seed 710 disperser and predator loss on recruitment success and long-term consequences for carbon stocks 711 in tropical rainforests. Sci. Rep. 7, 7662. https://doi.org/10.1038/s41598-017-08222-4 712 Dash, J., Curran, P.J., 2004. The MERIS terrestrial chlorophyll index. Int. J. Remote Sens. 25, 5403– 713 5413. https://doi.org/10.1080/0143116042000274015 714 Di Filippo, A., Biondi, F., Piovesan, G., Ziaco, E., 2017. Tree ring-based metrics for assessing old- 715 growth forest naturalness. J. Appl. Ecol. 54, 737–749. https://doi.org/10.1111/1365-2664.12793 716 Dirzo, R., Young, H.S., Galetti, M., Ceballos, G., Isaac, N.J.B., Collen, B., 2014. Defaunation in the 717 Anthropocene. Science (80-. ). 345, 401–406. https://doi.org/10.1126/science.1251817 718 Eisfelder, C., Kuenzer, C., Dech, S., 2012. Derivation of biomass information for semi-arid areas using 719 remote-sensing data. Int. J. Remote Sens. 33, 2937–2984. 720 https://doi.org/10.1080/01431161.2011.620034 721 El Peruano, 2006. Aprueban Categorización de Especies Amenazadas de Flora Silvestre. Decreto 722 Supremo N° 043-2006-AG. 723 ESA, 2017. CCI Land Cover Product User Guide version 2. 724 FAO, 2010. Global Forest Resources Assessment 2010, FAO Forestry Paper 163. 725 FAO, 2002. Second expert meeting on harmonizing forest-related definitions for use by various 726 stakeholders. Rome, Italy. 727 Fick, S.E., Hijmans, R.J., 2017. WorldClim 2: new 1-km spatial resolution climate surfaces for global 728 land areas. Int. J. Climatol. 37, 4302–4315. https://doi.org/10.1002/joc.5086 729 Forkuor, G., Benewinde Zoungrana, J.B., Dimobe, K., Ouattara, B., Vadrevu, K.P., Tondoh, J.E., 2020. 730 Above-ground biomass mapping in West African dryland forest using Sentinel-1 and 2 datasets - 731 A case study. Remote Sens. Environ. 236, 111496. https://doi.org/10.1016/j.rse.2019.111496 732 Frampton, W.J., Dash, J., Watmough, G., Milton, E.J., 2013. Evaluating the capabilities of Sentinel-2 733 for quantitative estimation of biophysical variables in vegetation. ISPRS J. Photogramm. Remote 734 Sens. 82, 83–92. https://doi.org/10.1016/j.isprsjprs.2013.04.007 735 Fremout, T., Thomas, E., Gaisberger, H., Van Meerbeek, K., Muenchow, J., Briers, S., Gutierrez- 736 Miranda, C.E., Marcelo-Peña, J.L., Kindt, R., Atkinson, R., Cabrera, O., Espinosa, C.I., Aguirre- 737 Mendoza, Z., Muys, B., 2020. Mapping tree species vulnerability to multiple threats as a guide to 738 restoration and conservation of tropical dry forests. Glob. Chang. Biol. 26, 3552–3568. 32 739 https://doi.org/10.1111/gcb.15028 740 Gao, B.C., 1996. NDWI - A normalized difference water index for remote sensing of vegetation liquid 741 water from space. Remote Sens. Environ. 58, 257–266. https://doi.org/10.1016/S0034- 742 4257(96)00067-3 743 Gao, Y., Skutsch, M., Paneque-Gálvez, J., Ghilardi, A., 2020. Remote sensing of forest degradation: a 744 review. Environ. Res. Lett. 15. https://doi.org/10.1088/1748-9326/abaad7 745 Garbulsky, M.F., Paruelo, J.M., 2004. Remote sensing of protected areas to derive baseline vegetation 746 functioning characteristics. J. Veg. Sci. 15, 711–720. https://doi.org/10.1111/j.1654- 747 1103.2004.tb02313.x 748 Gasparri, N.I., Parmuchi, M.G., Bono, J., Karszenbaum, H., Montenegro, C.L., 2007. Utilidad de 749 imágenes Landsat 7 ETM+ de diferentes fechas para la estimación de biomasa aérea en bosques 750 subtropicales secos de Argentina, in: XII Congreso de La Asociación Española de Teledetección. 751 p. 9. 752 Gaur, M.K., Squires, V.R., 2017. Climate variability impacts on land use and livelihoods in drylands, 753 in: Gaur, M.K., Squires, V.R. (Eds.), Climate Variability Impacts on Land Use and Livelihoods 754 in Drylands. Springer International Publishing, New York, USA, pp. 3–20. 755 https://doi.org/10.1007/978-3-319-56681-8 756 Ghazoul, J., Burivalova, Z., Garcia-Ulloa, J., King, L.A., 2015. Conceptualizing Forest Degradation. 757 Trends Ecol. Evol. 30, 622–632. https://doi.org/10.1016/j.tree.2015.08.001 758 Gitelson, A.A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote sensing of 759 global vegetation from EOS- MODIS. Remote Sens. Environ. 58, 289–298. 760 https://doi.org/10.1016/S0034-4257(96)00072-7 761 Gitelson, A.A., Verma, S.B., Vina, A., Rundquist, D.C., Keydan, G., Leavitt, B., Arkebauer, T.J., 762 Burba, G.G., Suyker, A.E., 2003. Novel technique for remote estimation of CO2 flux in maize. 763 Geophys. Res. Lett. 30, 1486. https://doi.org/10.1029/2002GL016543 764 Greenwell, B., Boehmke, B., Cunningham, J., GBM developers, 2019. Boosted Regression Models. R 765 package version 2.1.5. https://CRAN.R-project.org/package=gbm. 766 Halperin, J., LeMay, V., Coops, N., Verchot, L., Marshall, P., Lochhead, K., 2016. Canopy cover 767 estimation in miombo woodlands of Zambia: Comparison of Landsat 8 OLI versus RapidEye 768 imagery using parametric, nonparametric, and semiparametric methods. Remote Sens. Environ. 769 179, 170–182. https://doi.org/10.1016/j.rse.2016.03.028 770 Hansen, M.C.M., Potapov, P.V.P. V, Moore, R., Hancher, M., Turubanova, S.A., Tyukavina, A., Thau, 771 D., Stehman, S. V., Goetz, S.J., Loveland, T.R., Kommareddy, A., Egorov, A., Chini, L., Justice, 772 C.O., Townshend, J.R.G.G., 2013. High-resolution global maps of 21st-century forest cover 773 change. Science (80-. ). 342, 850–853. https://doi.org/10.1126/science.1244693 774 Haralick, R.M., 1979. Statistical and structural approaches to texture. Proc. IEEE 67, 786–804. 775 https://doi.org/10.1109/PROC.1979.11328 776 Hargreaves, G.H., Allen, R.G., 2003. History and evaluation of Hargreaves evapotranspiration 777 equation. J. Irrig. Drain. Eng. 129, 53–63. https://doi.org/10.1061/(ASCE)0733- 778 9437(2003)129:1(53) 779 Hastie, T., Tibshirani, R., Friedman, J.H., 2001. The Elements of Statistical Learning: Data Mining, 780 Inference, and Prediction. Springer, New York, USA. 781 Hengl, T., Mendes de Jesus, J., Heuvelink, G.B.M., Ruiperez Gonzalez, M., Kilibarda, M., Blagotić, 782 A., Shangguan, W., Wright, M.N., Geng, X., Bauer-Marschallinger, B., Guevara, M.A., Vargas, 783 R., MacMillan, R.A., Batjes, N.H., Leenaars, J.G.B., Ribeiro, E., Wheeler, I., Mantel, S., Kempen, 784 B., 2017. SoilGrids250m: Global gridded soil information based on machine learning. PLoS One 785 12, e0169748. https://doi.org/10.1371/journal.pone.0169748 33 786 Hernández-Stefanoni, J.L., Reyes-Palomeque, G., Castillo-Santiago, M. ángel, George-Chacón, S.P., 787 Huechacona-Ruiz, A.H., Tun-Dzul, F., Rondon-Rivera, D., Dupuy, J.M., 2018. Effects of sample 788 plot size and GPS location errors on aboveground biomass estimates from LiDAR in tropical dry 789 forests. Remote Sens. 10, 1586. https://doi.org/10.3390/rs10101586 790 Herrmann, S., Diouf, A.A., Sall, I., 2020. Beyond bioproductivity: Engaging local perspectives in land 791 degradation monitoring and assessment. J. Arid Environ. 173, 104002. 792 https://doi.org/10.1016/j.jaridenv.2019.104002 793 Herrmann, S.M., Tappan, G.G., 2013. Vegetation impoverishment despite greening: A case study from 794 central Senegal. J. Arid Environ. 90, 55–66. https://doi.org/10.1016/j.jaridenv.2012.10.020 795 Higginbottom, T.P., Symeonakis, E., Meyer, H., van der Linden, S., 2018. Mapping fractional woody 796 cover in semi-arid savannahs using multi-seasonal composites from Landsat data. ISPRS J. 797 Photogramm. Remote Sens. 139, 88–102. https://doi.org/10.1016/j.isprsjprs.2018.02.010 798 Hirschmugl, M., Gallaun, H., Dees, M., Datta, P., Deutscher, J., Koutsias, N., Schardt, M., 2017. 799 Methods for Mapping Forest Disturbance and Degradation from Optical Earth Observation Data: 800 a Review. Curr. For. Reports 3, 32–45. https://doi.org/10.1007/s40725-017-0047-2 801 Hirschmugl, M., Sobe, C., Deutscher, J., Schardt, M., 2018. Combined use of optical and synthetic 802 aperture radar data for REDD+ applications in Malawi. Land 7, 116. 803 https://doi.org/10.3390/land7040116 804 Hocquenghem, A.M., 2001. Una historia del bosque seco. Debate Agrar. 33, 39–60. 805 Huang, J., Yu, H., Guan, X., Wang, G., Guo, R., 2016. Accelerated dryland expansion under climate 806 change. Nat. Clim. Chang. 6, 166–171. https://doi.org/10.1038/nclimate2837 807 Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview of the 808 radiometric and biophysical performance of the MODIS vegetation indices. Remote Sens. 809 Environ. 83, 195–213. https://doi.org/10.1016/S0034-4257(02)00096-2 810 Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25, 295–309. 811 https://doi.org/10.1016/0034-4257(88)90106-X 812 IUCN, WRI, IUCN and WRI, 2014. A guide to the Restoration Opportunities Assessment Methodology 813 (ROAM): Assessing forest landscape restoration opportunities at the national or sub-national 814 level., (Road-test edition). Gland, Switzerland. 815 Jackson, H., Prince, S.D., 2016a. Degradation of net primary production in a semiarid rangeland. 816 Biogeosciences 13, 4721–4734. https://doi.org/10.5194/bg-13-4721-2016 817 Jackson, H., Prince, S.D., 2016b. Degradation of non-photosynthetic vegetation in a semi-arid 818 rangeland. Remote Sens. 8, 692. https://doi.org/10.3390/rs8080692 819 Janzen, D.H., 1988. Tropical Dry Forests: The Most Endangered Major Tropical Ecosystem, in: Wilson, 820 E.O., Peter, F.M. (Eds.), Biodiversity. National Academies Press, Washington D.C., pp. 130–137. 821 Jara-Guerrero, A., González-Sánchez, D., Escudero, A., Espinosa, C.I., 2021. Chronic Disturbance in a 822 Tropical Dry Forest: Disentangling Direct and Indirect Pathways Behind the Loss of Plant 823 Richness. Front. For. Glob. Chang. 4. https://doi.org/10.3389/ffgc.2021.723985 824 Karlson, M., Ostwald, M., Reese, H., Sanou, J., Tankoano, B., Mattsson, E., 2015. Mapping tree canopy 825 cover and aboveground biomass in Sudano-Sahelian woodlands using Landsat 8 and random 826 forest. Remote Sens. 7, 10017–10041. https://doi.org/10.3390/rs70810017 827 Kaufman, L., Rousseeuw, P.J., 1990. Finding Groups in Data: An Introduction to Cluster Analysis. 828 Wiley, New York. https://doi.org/10.1007/s13398-014-0173-7.2 829 Leutner, B., Horning, N., Schwalb-Wilmann, J., 2019. RStoolbox: Tools for Remote Sensing Data 830 Analysis. R package version 0.2.4. 831 Li, F., Meng, J., Zhu, L., You, N., 2020. Spatial pattern and temporal trend of land degradation in the 34 832 Heihe River Basin of China using local net primary production scaling. L. Degrad. Dev. 31, 518– 833 530. https://doi.org/10.1002/ldr.3468 834 Linares-Palomino, R., 2006. Phytogeography and floristics of seasonally dry tropical forests in Peru, 835 in: Pennington, R.T., Ratter, J.A. (Eds.), Neotropical Savannas and Seasonally Dry Forests. Plant 836 Diversity, Biogeography and Conservation. CRC Press, Boca Raton, USA, pp. 257–280. 837 https://doi.org/doi:10.1201/9781420004496.ch11 838 Loecher, M., Ropkins, K., 2015. RgoogleMaps and loa: Unleashing R graphics power on map tiles. J. 839 Stat. Softw. 63, 1–18. https://doi.org/10.18637/jss.v063.i04 840 Maechler, M., Rousseeuw, P., Struyf, A., Hubert, M., Hornik, K., 2019. cluster: Cluster Analysis Basics 841 and Extensions. R package version 2.1.0. 842 Malhi, Y., Doughty, C.E., Galetti, M., Smith, F.A., Svenning, J., 2016. Megafauna and ecosystem 843 function from the Pleistocene to the Anthropocene 113. https://doi.org/10.1073/pnas.1502540113 844 Marsett, R.C., Qi, J., Heilman, P., Biedenbender, S.H., Watson, M.C., Amer, S., Weltz, M., Goodrich, 845 D., Marsett, R., 2006. Remote sensing for grassland management in the arid Southwest. Rangel. 846 Ecol. Manag. 59, 530–540. https://doi.org/10.2111/05-201R.1 847 Mayes, M.T., Mustard, J.F., Melillo, J.M., 2015. Forest cover change in Miombo Woodlands: Modeling 848 land cover of African dry tropical forests with linear spectral mixture analysis. Remote Sens. 849 Environ. 165, 203–215. https://doi.org/10.1016/j.rse.2015.05.006 850 MINAM, 2018. Mapa Nacional de Ecosistemas: Memoria Descriptiva 37. 851 https://doi.org/10.1016/j.jpedsurg.2017.07.007 852 Muenchow, J., Dieker, P., Böttcher, T., Brock, J., Didenko, G., Fremout, T., Jakubka, D., Jentsch, A., 853 Nüst, D., Richter, M., Rodríguez, E.F., Rodríguez, R.A., Rollenbeck, R., Salazar Zarsosa, P., 854 Schratz, P., Brenning, A., 2020. Monitoring and predictive mapping of floristic biodiversity along 855 a climatic gradient in ENSO’s terrestrial core region, NW Peru. Ecography (Cop.). 43, 1–13. 856 https://doi.org/10.1111/ecog.05091 857 Murphy, P.G., Lugo, A.E., 1986. Ecology of Tropical Dry Forest. Annu. Rev. Ecol. Syst. 17, 67–88. 858 https://doi.org/https://doi.org/10.1146/annurev.es.17.110186.000435 859 Muscarella, R., Galante, P.J., Soley-Guardia, M., Boria, R.A., Kass, J.M., Uriarte, M., Anderson, R.P., 860 2014. ENMeval: An R package for conducting spatially independent evaluations and estimating 861 optimal model complexity for MaxEnt ecological niche models. Methods Ecol. Evol. 5, 1198– 862 1205. https://doi.org/10.1111/2041-210X.12261 863 Nagelkirk, R.L., Dahlin, K.M., 2020. Woody cover fractions in African Savannas from Landsat and 864 High-Resolution Imagery. Remote Sens. 12, 813. https://doi.org/10.3390/rs12050813 865 Návar, J., 2014. Pan tropical biomass equations for Mexico’s dry forests. Agron. Colomb. 32, 367–376. 866 https://doi.org/10.15446/agron.colomb.v32n3.45627 867 Noojipady, P., Prince, S.D., Rishmawi, K., 2015. Reductions in productivity due to land degradation in 868 the drylands of the southwestern United States. Ecosyst. Heal. Sustain. 1, 1–15. 869 https://doi.org/10.1890/EHS15-0020.1 870 Olson, D.M., Dinerstein, E., Wikramanayake, E.D., Burgess, N.D., Powell, G.V.N., Underwood, E.C., 871 D’amico, J.A., Itoua, I., Strand, H.E., Morrison, J.C., Loucks, C.J., Allnutt, T.F., Ricketts, T.H., 872 Kura, Y., Lamoreux, J.F., Wettengel, W.W., Hedao, P., Kassem, K.R., 2001. Terrestrial 873 Ecoregions of the World: A New Map of Life on Earth. Bioscience 51, 933. 874 https://doi.org/10.1641/0006-3568(2001)051[0933:TEOTWA]2.0.CO;2 875 Padrón, E., Navarro-Cerrillo, R.M., 2007. Aboveground biomass in Prosopis pallida (Humb. and Bonpl. 876 ex Willd.) H. B. K. ecosystems using Landsat 7 ETM+ images. Rev. Chil. Hist. Nat. 80, 43–53. 877 https://doi.org/10.4067/S0716-078X2007000100004 878 Pauly, D., 1995. Anecdotes and the shifting baseline syndrome of fisheries. Trends Ecol. Evol. 10, 430. 35 879 https://doi.org/10.1016/S0169-5347(00)89171-5 880 Peres, C.A., Emilio, T., Schietti, J., Desmoulière, S.J.M., Levi, T., 2016. Dispersal limitation induces 881 long-term biomass collapse in overhunted Amazonian forests. Proc. Natl. Acad. Sci. U. S. A. 113, 882 892–897. https://doi.org/10.1073/pnas.1516525113 883 Pötzschner, F., Baumann, M., Gasparri, N.I., Conti, G., Loto, D., Piquer-Rodríguez, M., Kuemmerle, 884 T., 2022. Ecoregion-wide, multi-sensor biomass mapping highlights a major underestimation of 885 dry forests carbon stocks. Remote Sens. Environ. 269. https://doi.org/10.1016/j.rse.2021.112849 886 Power, M.J., Marlon, J., Ortiz, N., Bartlein, P.J., Harrison, S.P., Mayle, F.E., Ballouche, A., Bradshaw, 887 R.H.W., Carcaillet, C., Cordova, C., Mooney, S., Moreno, P.I., Prentice, I.C., Thonicke, K., 888 Tinner, W., Whitlock, C., Zhang, Y., Zhao, Y., Ali, A.A., Anderson, R.S., Beer, R., Behling, H., 889 Briles, C., Brown, K.J., Brunelle, A., Bush, M., Camill, P., Chu, G.Q., Clark, J., Colombaroli, D., 890 Connor, S., Daniau, A.L., Daniels, M., Dodson, J., Doughty, E., Edwards, M.E., Finsinger, W., 891 Foster, D., Frechette, J., Gaillard, M.J., Gavin, D.G., Gobet, E., Haberle, S., Hallett, D.J., Higuera, 892 P., Hope, G., Horn, S., Inoue, J., Kaltenrieder, P., Kennedy, L., Kong, Z.C., Larsen, C., Long, C.J., 893 Lynch, J., Lynch, E.A., McGlone, M., Meeks, S., Mensing, S., Meyer, G., Minckley, T., Mohr, J., 894 Nelson, D.M., New, J., Newnham, R., Noti, R., Oswald, W., Pierce, J., Richard, P.J.H., Rowe, C., 895 Sanchez Goñi, M.F., Shuman, B.N., Takahara, H., Toney, J., Turney, C., Urrego-Sanchez, D.H., 896 Umbanhowar, C., Vandergoes, M., Vanniere, B., Vescovi, E., Walsh, M., Wang, X., Williams, 897 N., Wilmshurst, J., Zhang, J.H., 2008. Changes in fire regimes since the last glacial maximum: An 898 assessment based on a global synthesis and analysis of charcoal data. Clim. Dyn. 30, 887–907. 899 https://doi.org/10.1007/s00382-007-0334-x 900 Prăvălie, R., 2016. Drylands extent and environmental issues. A global approach. Earth-Science Rev. 901 161, 259–278. https://doi.org/10.1016/j.earscirev.2016.08.003 902 Prince, S.D., 2004. Mapping Desertification in Southern Africa, in: Gutman, G., Janetos, A.C., Justice, 903 C.O., Moran, E.F., Mustard, J.F., Rindfuss, R.R., Skole, D., Turner, B.L., Cochrane, M.A. (Eds.), 904 Land Change Science: Observing, Monitoring and Understanding Trajectories of Change on the 905 Earth’s Surface. Springer Netherlands, Dordrecht, pp. 163–184. https://doi.org/10.1007/978-1- 906 4020-2562-4_10 907 Prince, S.D., Becker-Reshef, I., Rishmawi, K., 2009. Detection and mapping of long-term land 908 degradation using local net production scaling: Application to Zimbabwe. Remote Sens. Environ. 909 113, 1046–1057. https://doi.org/10.1016/j.rse.2009.01.016 910 Putz, F.E., Redford, K.H., 2010. The Importance of Defining “Forest”: Tropical Forest Degradation, 911 Deforestation, Long-term Phase Shifts, and Further Transitions. Biotropica 42, 10–20. 912 https://doi.org/10.1111/j.1744-7429.2009.00567.x 913 Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A modified soil adjusted 914 vegetation index. Remote Sens. Environ. 48, 119–126. https://doi.org/10.1016/0034- 915 4257(94)90134-1 916 Qi, J., Marsett, R.C., Moran, M.S., Goodrich, D.C., Heilman, P., Kerr, Y.H., Dedieu, G., Chehbouni, 917 A., Zhang, X.X., 2000. Spatial and temporal dynamics of vegetation in the San Pedro River basin 918 area. Agric. For. Meteorol. 105, 55–68. https://doi.org/10.1016/S0168-1923(00)00195-7 919 Redford, K., 1992. The empty forest. Bioscience 42, 412–422. 920 https://doi.org/10.1126/science.345.6195.396 921 Reiche, J., Hamunyela, E., Verbesselt, J., Hoekman, D., Herold, M., 2018. Improving near-real time 922 deforestation monitoring in tropical dry forests by combining dense Sentinel-1 time series with 923 Landsat and ALOS-2 PALSAR-2. Remote Sens. Environ. 204, 147–161. 924 https://doi.org/10.1016/j.rse.2017.10.034 925 Reynolds, J.F., Stafford Smith, D.M., Lambin, E.F., Turner, B.L., Mortimore, M., Batterbury, S.P.J., 926 Downing, T.E., Dowlatabadi, H., Fernández, R.J., Herrick, J.E., Huber-Sannwald, E., Jiang, H., 927 Leemans, R., Lynam, T., Maestre, F.T., Ayarza, M., Walker, B., 2007. Global desertification: 36 928 Building a science for dryland development. Science (80-. ). 316, 847–851. 929 https://doi.org/10.1126/science.1131634 930 Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera-Arroita, G., Hauenstein, S., Lahoz- 931 Monfort, J.J., Schröder, B., Thuiller, W., Warton, D.I., Wintle, B.A., Hartig, F., Dormann, C.F., 932 2017. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic 933 structure. Ecography (Cop.). 40, 913–929. https://doi.org/10.1111/ecog.02881 934 Romero-Sanchez, M.E., Ponce-Hernandez, R., 2017. Assessing and monitoring forest degradation in a 935 deciduous tropical forest in Mexico via remote sensing indicators. Forests. 936 https://doi.org/10.3390/f8090302 937 Safriel, U., Adeel, Z., 2005. Dryland Systems, in: Hassan, R., Scholes, R., Ash, N. (Eds.), Ecosystems 938 and Human Well-Being: Current State and Trends: Findings of the Condition and Trends Working 939 Group. Island Press, Washington D.C., USA, pp. 625–662. 940 Sandom, C.J., Ejrnaes, R., Hansen, M.D.D., Svenning, J.C., 2014. High herbivore density associated 941 with vegetation diversity in interglacial ecosystems. Proc. Natl. Acad. Sci. U. S. A. 111, 4162– 942 4167. https://doi.org/10.1073/pnas.1311014111 943 Santiago-García, R.J., Finegan, B., Bosque-Pérez, N.A., 2019. Soil is the main predictor of secondary 944 rain forest estimated aboveground biomass across a Neotropical landscape. Biotropica 51, 10–17. 945 https://doi.org/10.1111/btp.12621 946 Schneibel, A., Stellmes, M., Röder, A., Frantz, D., Kowalski, B., Haß, E., Hill, J., 2017. Assessment of 947 spatio-temporal changes of smallholder cultivation patterns in the Angolan Miombo belt using 948 segmentation of Landsat time series. Remote Sens. Environ. 195, 118–129. 949 https://doi.org/10.1016/j.rse.2017.04.012 950 Schratz, P., Muenchow, J., Iturritxa, E., Richter, J., Brenning, A., 2019. Hyperparameter tuning and 951 performance assessment of statistical and machine-learning algorithms using spatial data. Ecol. 952 Modell. 406, 109–220. https://doi.org/10.1016/j.ecolmodel.2019.06.002 953 Sekula, M., Datta, Somnath, Datta, Susmita, 2020. optCluster: Determine Optimal Clustering 954 Algorithm and Number of Clusters. R package version 1.3.0. 955 Smith, R.J., Mayle, F.E., 2021. Impact of mid- to late Holocene precipitation changes on vegetation 956 across lowland tropical South America : a paleo-data synthesis 134–155. 957 https://doi.org/10.1017/qua.2017.89 958 Teillet, P.M., Guindon, B., Goodenough, D.G., 1982. On the slope-aspect correction of multispectral 959 scanner data. Can. J. Remote Sens. 8, 84–106. https://doi.org/10.1080/07038992.1982.10855028 960 Thomas, E., Jalonen, R., Loo, J., Boshier, D., Gallo, L., Cavers, S., Bordács, S., Smith, P., Bozzano, 961 M., 2014. Genetic considerations in ecosystem restoration using native tree species. For. Ecol. 962 Manage. 333, 66–75. https://doi.org/10.1016/j.foreco.2014.07.015 963 Thompson, I.D., Guariguata, M.R., Okabe, K., Bahamondez, C., Nasi, R., Heymell, V., 2012. An 964 Operational Framework for Defining and Monitoring Forest Degradation 18. 965 Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring vegetation. 966 Remote Sens. Environ. 8, 127–150. https://doi.org/10.1016/0034-4257(79)90013-0 967 UNEP, 1992. World atlas of desertification, United Nations Environment Programme. Edward Arnold, 968 London. 969 Urbazaev, M., Thiel, C., Mathieu, R., Naidoo, L., Levick, S.R., Smit, I.P.J.J., Asner, G.P., Schmullius, 970 C., 2015. Assessment of the mapping of fractional woody cover in southern African savannas 971 using multi-temporal and polarimetric ALOS PALSAR L-band images. Remote Sens. Environ. 972 166, 138–153. https://doi.org/10.1016/j.rse.2015.06.013 973 Valavi, R., Elith, J., Lahoz-Monfort, J.J., Guillera-Arroita, G., 2019. blockCV: An r package for 974 generating spatially or environmentally separated folds for k-fold cross-validation of species 37 975 distribution models. Methods Ecol. Evol. 10, 225–232. https://doi.org/10.1111/2041-210X.13107 976 van der Sande, M.T., Arets, E.J.M.M., Peña-Claros, M., Hoosbeek, M.R., Cáceres-Siani, Y., van der 977 Hout, P., Poorter, L., 2018. Soil fertility and species traits, but not diversity, drive productivity 978 and biomass stocks in a Guyanese tropical rainforest. Funct. Ecol. 32, 461–474. 979 https://doi.org/10.1111/1365-2435.12968 980 Van Deventer, A.P., Ward, A.D., Gowda, P.M., Lyon, J.G., 1997. Using thematic mapper data to 981 identify contrasting soil plains and tillage practices. Photogramm. Eng. Remote Sensing 63, 87– 982 93. 983 Van Passel, J., De Keersmaecker, W., Somers, B., 2020. Monitoring woody cover dynamics in tropical 984 dry forest ecosystems using Sentinel-2 satellite imagery. Remote Sens. 12, 1276. 985 https://doi.org/10.3390/rs12081276 986 Verón, S.R., Paruelo, J.M., Oesterheld, M., 2006. Assessing desertification. J. Arid Environ. 66, 751– 987 763. https://doi.org/10.1016/j.jaridenv.2006.01.021 988 Vieira, D.L.M., Scariot, A., 2006. Principles of natural regeneration of Tropical Dry Forests for 989 regeneration. Restor. Ecol. 14, 11–20. https://doi.org/10.1111/j.1526-100X.2006.00100.x 990 Walther, C., Sietz, D., Lu, M.K.B., Lüdeke, M.K.B., Walther, C., 2011. Categorisation of typical 991 vulnerability patterns in global drylands. Glob. Environ. Chang. 21, 431–440. 992 https://doi.org/10.1016/j.gloenvcha.2010.11.005 993 Wessels, K.J., Prince, S.D., Reshef, I., 2008. Mapping land degradation by comparison of vegetation 994 production to spatially derived estimates of potential production. J. Arid Environ. 72, 1940–1949. 995 https://doi.org/10.1016/j.jaridenv.2008.05.011 996 Whaley, O.Q., Borda, C., Moat, J., Wilkinson, T., Bravo-Sánchez, A., Gagné, R.J., 2020. Ecología y 997 diagnóstico de Enallodiplosis discordis (Diptera:Cecidomyiidae): un nuevo defoliador feroz con 998 repercusiones directas en la pérdida del bosque seco de Prosopis y los medios de vida en Perú. 999 Rev. Peru. Biol. 27, 451–482. https://doi.org/http://dx.doi. org/10.15381/rpb.v27i4.19200 1000 Wu, W., 2014. The Generalized Difference Vegetation Index (GDVI) for dryland characterization. 1001 Remote Sens. 6, 1211–1233. https://doi.org/10.3390/rs6021211 1002 Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance open water 1003 features in remotely sensed imagery. Int. J. Remote Sens. 27, 3025–3033. 1004 https://doi.org/10.1080/01431160600589179 1005 Yang, J., Weisberg, P.J., Bristow, N.A., 2012. Landsat remote sensing approaches for monitoring long- 1006 term tree cover dynamics in semi-arid woodlands: Comparison of vegetation indices and spectral 1007 mixture analysis. Remote Sens. Environ. 119, 62–71. https://doi.org/10.1016/j.rse.2011.12.004 1008 Zha, Y., Gao, J., Ni, S., 2003. Use of normalized difference built-up index in automatically mapping 1009 urban areas from TM imagery. Int. J. Remote Sens. 24, 583–594. 1010 https://doi.org/10.1080/01431160304987 1011 Zhang, W., Brandt, M., Wang, Q., Prishchepov, A. V., Tucker, C.J., Li, Y., Lyu, H., Fensholt, R., 2019. 1012 From woody cover to woody canopies: How Sentinel-1 and Sentinel-2 data advance the mapping 1013 of woody plants in savannas. Remote Sens. Environ. 234, 111465. 1014 https://doi.org/10.1016/j.rse.2019.111465 1015 Zvoleff, A., 2019. glcm: Calculate Textures from Grey-Level Co-Occurrence Matrices (GLCMs). R 1016 package version 1.6.4. 1017 1018 38 1019 List of figure captions 1020 Figure 7: The location of Lambayeque in Peru (panel a), study region with aridity zones (panel b) and Normalized 1021 Difference Vegetation Index (NDVI; panel c).The study region consists of non-converted tropical dry forests areas 1022 in Lambayeque (see Supplementary Material 1 for details on how the forest extent map was obtained). . The NDVI 1023 map (panel c) depicts the maximum of the NDVI composites over 3 seasons (wet, transition, and dry season) 1024 derived from Sentinel-2 data (see section 2.2.3). 1025 Figure 8: Illustration of the woody cover assessment method using thresholding of Google Earth imagery: (a) 1026 190x190m Google Earth image chip, (b) 95x95m Google Earth image chip, (c) woody cover (WC) obtained by 1027 thresholding the brightness (i.e., the sum of the red-green-blue (RGB) values) of the 95x95m image chip. 1028 Figure 9: Flowchart showing a summary of the methodology followed to obtain woody cover (WC) and 1029 aboveground biomass (AGB) reduction maps. CLARA stands for the Clustering Large Applications method. Note 1030 that this flowchart provides a simplified overview without (pre-)processing of predictor variables. 1031 Figure 10: Cross-validated accuracy of (a) woody cover (WC) and (b) aboveground biomass (AGB) predictions 1032 using boosted regression trees and data from different sources (Sentinel-2 only and Sentinel-2 + Sentinel-1) and 1033 different seasons (multi-season, dry season, transition season, and wet season). The MAE (mean absolute error) 1034 values shown are those obtained from the models with tuned hyperparameter settings (see section 0). The error 1035 bars indicate the range between accuracy metric values obtained between the five repetitions of the 10-fold cross- 1036 validation, each with a different configuration of spatial blocks (see section 0). 1037 Figure 11: Illustration of woody cover reduction mapping at three locations with varying degrees of degradation. 1038 Panel (a) shows the very high-resolution Google Earth image chips of these locations, panel (b) shows the 1039 estimated woody cover reduction. Note that woody cover estimates of individuals pixels reflect the woody cover 1040 of a 95x95 m window rather than of the individual pixels. 1041 1042 39