Wild Boar Collision Data and Satellite Computer Vision Refine Habitat Suitability Mapping across France

Wild Boar Collision Data and Satellite Computer Vision Refine Habitat Suitability Mapping across France | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Wild Boar Collision Data and Satellite Computer Vision Refine Habitat Suitability Mapping across France Adrian Ferdinand Meyer, Théo Reibel, Kevin Morelle, Mathias Kneubühler, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8798859/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Wild boar (Sus scrofa) populations have expanded rapidly across Europe, leading to escalating human-wildlife conflicts (HWCs), notably wildlife-vehicle collisions (WVCs), increased agricultural damage and disease transmission. In continental France these issues are compounded by the species’ ecological adaptability exhibiting increasing overlaps with urbanization and transportation networks. In this study, we leverage time-series environmental data, computer vision and Species Distribution Modeling (SDM) to predict wild boar habitat suitability and investigate its spatiotemporal drivers. We use presence-only data from WVC reports on the national railway and road networks, along with publicly available GBIF observation collections, to enhance the predictive power of our SDMs, while addressing inherent sampling biases of these datasets with tailored corrections. A key innovation of this study is the integration of large scale, very-high-resolution land cover predictors explicitly focused on wild boar resource preference. By fine-tuning a multitemporal Vision Transformer foundational AI model on multispectral satellite remote sensing imagery we capture subtle seasonal phenological differences. Our results highlight clear spatial, seasonal and annual variations in wild boar habitat suitability. The multitemporal SDM pipeline offers improved ecological realism and resilience to climate extremes, yielding meaningful predictions when extrapolating to novel environmental scenarios. The methodological and ecological insights gained through this study provide actionable knowledge for French transportation planning, agriculture and wildlife management. Identifying regions with high seasonal habitat suitability can inform targeted and preventive interventions. More broadly, our results demonstrate that advanced, data-driven methods are becoming indispensable for proactively and sustainably addressing HWCs in an increasingly anthropogenic world. Wild Boar (Sus scrofa) Species Distribution Models Multitemporal Modeling Wildlife-Vehicle Collisions Remote Sensing Vision Transformer Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Introduction Across Europe, expanding ungulate populations and the proliferation of infrastructure and agriculture have significantly intensified human-wildlife conflicts (HWCs). Especially wildlife-vehicle collisions (WVCs), crop damage, and disease transmission in wild boar (Sus scrofa) exemplify these conflicts (Langbein et al. 2011; Massei et al. 2015). Due to its rapid population growth and ecological adaptability, wild boar increasingly exploits heterogeneous, human-dominated landscapes, including agricultural mosaics, suburban areas, and transport networks (Morelle et al. 2014; Stillfried et al. 2017; Barasona et al. 2021). These conflicts highlight transport safety concerns, but elevated interactions between wild boar and humans also underline broader implications for veterinary epidemiology (e.g. African Swine Fever, ASF), agriculture, and wildlife management (Sauter-Louis et al. 2021). Modeling the spatiotemporal extent of wild boar distribution represents an expedient monitoring approach. Species distribution models (SDMs) provide a practical way to translate heterogeneous data into spatial predictions of relative suitability or occurrence, and they are not only widely used in wildlife ecology and management (Elith et al. 2011), but can also be specifically applied to the wild boar context at national scale (Bosch et al. 2014). Effectively managing wild boar populations and conflicts requires dynamic models capable of capturing spatiotemporal shifts in habitat use. Wild boar movements and habitat preferences vary seasonally with resource availability and anthropogenic pressures such as hunting (Thurfjell et al. 2013; Morelle and Lejeune 2015). These seasonal variations complicate static modeling approaches that rely on outdated or coarse-resolution environmental predictors, risking misrepresentation of actual habitat use and conflict hotspots (Culshaw et al. 2021). Moreover, most existing wild boar SDMs are based on static land cover (LC) maps, such as Corine Land Cover, that fail to reflect dynamic landscape changes driven by crop rotations, fruitification events, climatic extremes, and human disturbances (Pettorelli et al. 2005). Recent advances in SDMs and remote sensing (RS) offer pathways toward a more robust, multitemporal understanding of habitat suitability (Wang et al. 2025). Dynamic SDMs, which explicitly incorporate temporally resolved satellite-derived environmental predictors, significantly enhance ecological realism (He et al. 2015). The lack of seasonally explicit SDMs for wild boar remains a gap in Europe, limiting precise, targeted management interventions in agriculture, transportation, disease management, and conservation (Enetwild Consortium et al. 2024). France provides an ideal case study due to its diverse landscapes, growing wild boar populations involved in frequent HWCs, and extensive national-scale monitoring data, yet it currently lacks a consistent, seasonal SDM framework. Additionally, wild boar have been observed to increase their home range during winter over multiple French climate zones (Calenge et al. 2002). Especially the rapidly rising wild boar incidents on the French railway network (Société nationale des chemins de fer français, SNCF) motivate a national-level investigation (Santos et al. 2017). In this study, we address this critical gap by developing an innovative multitemporal SDM pipeline that integrates modern RS approaches, advanced machine learning (ML), and carefully curated occurrence data for continental France. Specifically, we leverage novel LC classification derived from a state-of-the-art Vision Transformer (ViT) deep learning architecture fine-tuned on high-resolution multispectral satellite imagery (Jakubik et al. 2023). By training this model on multitemporal RS stacks, we produce highly detailed annual LC predictors at 30 m resolution that accurately capture subtle phenological and anthropogenic land-use dynamics crucial to wild boar ecology such as crop rotation or forest mast. Multistage satellite time-series data-fusion ML frameworks combining state-of-the-art computer vision with environmental proxy features based on expert annotation have proven particularly effective for such LC paradigms (Meyer et al. 2025). Our methodological framework builds on structured WVC reports from SNCF and roadkill reports by the Réseau Routier National (RRN) databases alongside opportunistic observations by the Global Biodiversity Information Facility (GBIF). Each presence-only dataset provides distinct ecological insights but carries inherent spatial biases. We explicitly apply spatial bias-correction strategies to prevent SDMs from inadvertently learning observational patterns instead of true ecological relationships (Baker et al. 2024). Importantly, rather than treating these dataset differences as nuisances, we leverage them strategically: WVC datasets uniquely contribute systematic, spatially structured observations reflecting animal movements relative to linear infrastructure, while GBIF enriches ecological breadth and increases geographical coverage. Therefore, we address the following research objectives: (i) Do the seasonally explicit habitat suitability maps produced by our SDM approach provide ecologically meaningful proxies suitable for stakeholder use? (ii) Does multitemporal SDM calibration improve temporal transferability compared with a monotemporal “average year” baseline? (iii) Does integrating rail and road WVC data enhance the ecological realism of predicted management-relevant interface areas over classical “GBIF-only” models? This work provides a rigorous, data-driven foundation for targeted interventions in transportation planning, agriculture, veterinary epidemiology, and wildlife management. By highlighting the value of integrating multitemporal RS data and structured collision records, our framework demonstrates the feasibility of modern ML modeling pipelines. Moreover, as the developed LC classification approach leverages universally available satellite data, our methodology offers broad global transferability beyond the French context and can inform similar predictive modeling efforts for other conflict-prone ungulates. Methods Process Pipeline Overview Our methodological framework (Fig. 1) involves several integrated analytical stages (detailed in section 2.2-2.7) to rigorously model seasonal habitat suitability for wild boar over the years 2017 to 2023 at a 1 km 2 resolution pixel grid covering continental France as our study area. Corsica is excluded due to its insular topography boasting strong variations in wild boar behavior such as heavy hybridization with domestic pigs (Iacolina et al. 2018) and documented endemic subspecies formation (Schleimer et al. 2022) beyond the scope of this study. Briefly, we (i) assemble three complementary presence-only streams (rail and road collision records plus GBIF observations, Fig. 1 top-left), (ii) reduce clustering and redundant sampling in environmental space to mitigate pseudo-replication (Fig. 1 dark-red pathway), (iii) model source-specific observation effort and use it to weight background sampling (Fig. 1 violet pathway), (iv) construct a predictor stack that combines annually updated LC maps from the satellite-based ViT with seasonal remote-sensing indices, climate, and anthropogenic layers (Fig. 1 central grey and yellow pathways), (v) fit and compare monotemporal versus multitemporal SDMs for two seasons (Fig. 1 bottom), and (vi) evaluate temporal transferability using leave-one-year-out (LOYO) validation and presence-only diagnostics (Fig. 1 bottom-right). The pipeline produces annual 30 m LC maps explicitly tailored to wild boar ecology capturing critical resources and shelter habitats as an intermediate product (Fig. 1 yellow pathway). Subsequently we derive predictive seasonal habitat-suitability index layers at 1 km² resolution that remain robust and interpretable across interannual variability, providing an essential resource for operational planning, targeted management and early-warning applications. Presence Data and Thinning We compile our comprehensive presence-only dataset of wild boar occurrences across France for the years 2017–2023 from three primary sources combined to maximize coverage of wild boar occurrences in different landscapes (Fig. 2): (i) Railway collision reports provided by SNCF Réseau (the French rail network infrastructure operator) documenting wild boar struck by trains extracted from the “Brehat” database (Arthaud et al. 2024), (ii) Roadkill carcass records from the toll-free national road network (RRN) reporting vehicle collisions with wild boar provided by the French Ministry for Ecological Transition (DGITM/DMR/SAM4, Sept 2024) and (iii) combined wild boar observations aggregated through GBIF, which include opportunistic sightings, various published scientific observation collections and citizen-science derived reports by the general public for example through mobile apps (Beck et al. 2012; Enetwild Consortium et al. 2018). We only include points with verified coordinate positions (error smaller than 0.5 km) and time stamps (temporal resolution at least one month) and filter out all duplicates in the same 1 km 2 cell. Each record is assigned to one of two “seasons” based on the month of observation, allowing us to model habitat selection separately for “summer” (Mar–Aug) and “winter” (Sep–Feb). This split is chosen to largely cover the French hunting season within the winter dataset: While the exact dates and special regulations can vary by department according to the federal biodiversity office OFB the hunting period usually opens in early September and ends by the end of February (Calenge et al. 2002; Vajas et al. 2023); stalking hunt can be allowed with a special permit in summer (Maillard et al. 2010). Presence-only datasets are typically spatially clustered, which can bias model fitting toward densely sampled environments. We therefore applied environmental-space thinning to reduce redundant presences while preserving the breadth of occupied environmental conditions (Aiello-Lammens et al. 2015). We opted for subsampling occurrences using greedy farthest-point sampling (Chamon and Ribeiro 2017) in a decorrelated Mahalanobis-distance representation of the predictor space (de Oliveira et al. 2014) for each given season/year/dataset combination. In practice, we keep ~40% of the original records (Fig. 2) to balance reducing the impact of pseudoreplication (Fourcade et al. 2014) while retaining an ecologically meaningful signal which preserves the fundamental ecological niche of the wild boar as our target species (Jiménez and Soberón 2022). Algorithmic details are provided in Supplementary Information (SI) S1. Bias Correction of Background Points Background point sampling and selection is a key step in presence-only based SDMs, because it can significantly distort suitability predictions unless sampling effort is explicitly modeled (Lobo et al. 2010; Barbet-Massin et al. 2012). Uneven observer effort and accessibility can lead models to interpret sampling patterns as habitat preference requiring effective spatial correction strategies (Baker et al. 2024). To mitigate such effects, we construct spatially explicit sampling bias surfaces (Fig. 3) for each occurrence source and each year/season to model observation effort as a proxy for realistic sampling intensity (Lobo et al. 2010). This increasingly endorsed practice in presence-only SDMs aligns background point generation with actual detectability (Beck et al. 2014). Syfert et al. (2013) found that incorporating a sampling bias grid significantly enhanced goodness-of-fit metrics, even though it did not completely eliminate bias effects. Meta-analyses and simulations also support modest but meaningful gains from bias correction on independent test data (Baker et al. 2024). Implementation approach including algorithmic details and background information is described in SI S2. GBIF occurrence data are available in large quantities for wild boar in Europe (Croft et al. 2025), but are known to suffer from strong spatial biases such as over-representation near roads, urban centers and accessible sites (Beck et al. 2014; Meyer et al. 2016). These biases may correlate with environmental predictors and lead to misleading niche estimations (Baker et al. 2022; Dubos et al. 2022). To mitigate such effects, recent work recommends modeling sampling effort explicitly (Sicacha-Parada et al. 2019) and implementing density surfaces or accessibility proxies to weight background sampling (Barber et al. 2022). We modeled observer effort using kernel-density estimation of reporting intensity combined with an accessibility proxy (Human Footprint Index) to account for the tendency of observations to cluster near roads, settlements, and other accessible areas (Sanderson et al. 2002; Hughes et al. 2021) (Fig. 3B). For implementation details see SI S2. WVCs and traffic infrastructure data can serve as a highly valuable monitoring prior (Schwartz et al. 2020), with railway operations boasting a particularly high scheduled regularity (Jasińska et al. 2019). Because WVCs are exclusively reported in the immediate vicinity of infrastructure networks we assume that higher traffic volume generally leads to a higher likelihood of detecting collisions (Morelle et al. 2013), although we acknowledge that traffic density is only one of several interacting factors influencing WVCs (Benard 2023). We modeled detection effort as a functional proxy of network extent and traffic intensity, using annual average daily traffic for roads (“TMJA” dataset obtained from the French Ministry for Ecological Transition, Aug 2025) and annual average daily train frequency for rail sections (proprietary dataset provided by SNCF Réseau, Mar 2025) (Fig. 3C and D). For implementation details see SI S2. The three bias layers were normalized and combined into a composite bias surface per season (Fig. 3A), which then defined the probability of drawing background points. This composite weighting retains the footprint of each observation stream while reducing the risk of spurious correlations driven purely by accessibility or infrastructure placement. Mathematical definitions including blending parameters and interpolation of missing traffic counts, as well as a full list of bias maps are reported in SI S2. LC Predictors and ViT A key innovation of this study is the use of annually updated high-resolution land-cover (LC) predictors designed around wild boar resources and shelter (Fig. 4, left three columns). To represent the dynamic shift of LC more faithfully than standardized static products, we produced annual LC maps at 30 m resolution for 2017–2023 using a multispectral ViT semantic-segmentation foundation model “Prithvi-100M” (Jakubik et al. 2023) fine-tuned on harmonized Landsat–Sentinel-2 surface reflectance imagery (HLS). HLS is provided by NASA and combines the four satellite missions Landsat-8/9 and Sentinel-2A/B into a common atmospherically and radiometrically corrected virtual constellation (Claverie et al. 2018). Training labels were compiled from multiple high-quality national and international GIS resources to provide robust training labels. The 12 chosen classes (see SI S3a) directly correspond to wild boar habitat and foraging opportunities, given their opportunistic omnivorous diet primarily consisting of vegetation and agricultural crops (Brandt et al. 2006; Ballari and Barrios-García 2014). Wild boar frequently exploit energy-rich cultivated crops when available, shifting to natural resources such as acorns, chestnuts, roots, and tubers in forests or outside the cropping season. Brandt et al. (2006) highlight the significant dietary contributions of tree mast (oak and beech nuts) and cultivated crops to the French wild boar diet. Annotations were sourced for crop types from the Registre Parcellaire Graphique (RPG) (Cantelaube and Carles 2014), forest types from the French national forest inventory (IGN BD Forêt v2) augmented with annual forest-loss updates (Hansen et al. 2013), and additional land-use categories from Theia LC products (Inglada et al. 2017; Puissant et al. 2019), complemented by Corine Land Cover and OpenStreetMap for built-up and water classes. We trained with an auxiliary ‘Other/NoData’ class for segmentation stability, which was excluded from the subsequent workflow. Details on class composition can be found in SI S3a. For SDM integration, the 30 m LC maps were aggregated to the 1 km² modeling grid as (i) fractional cover per class and (ii) distance-to-class layers capturing edge and proximity effects, which are known to be important for wild boar use of forest-agriculture interfaces (Thurfjell et al. 2009). Model training details, hyperparameters, hardware requirements and inference configuration are provided in SI S3b. Environmental Predictor Stack Beyond LC composition, we included dynamic and static predictors capturing vegetation condition, climate, terrain, fragmentation, and anthropogenic pressures (Fig. 4). RS time series can substantially improve SDMs by representing within- and between-year variation in habitat conditions (Pettorelli et al. 2005; He et al. 2015). We derived spectral indices from HLS mosaics at key phenological stages (April, July, October) and assigned them seasonally (spring/summer indices to summer models; autumn indices to winter models). To reduce redundancy and multicollinearity, predictors were pruned using variance inflation factor (VIF) filtering following established practice in wild boar SDMs (Bosch et al. 2014). Climate predictors combined dynamic seasonal anomalies from ERA5-Land reanalysis (temperature, precipitation, soil moisture, solar radiation; plus snow and minima for winter) (Muñoz-Sabater et al. 2021) with higher-resolution long-term bioclimatic normals from WorldClim to represent microclimatic gradients relevant to wild boar persistence and movement (Enetwild Consortium et al. 2019; Vetter et al. 2020). Terrain variables (elevation, slope, aspect) were derived from the “SRTM DEM” dataset (obtained from USGS, see SI S4) to account for topographic structure and known wild-boar specific constraints (Acevedo et al. 2006). We further included landscape fragmentation resampled from the “effective mesh density” dataset (EEA/FOEN 2011) and the “Copernicus small woody features” dataset (Faucqueur et al. 2019) as proxies for permeability, shelter and corridor structure, which can shape wild boar movement and conflict risk (Welander 2000; Ficetola et al. 2014). Wetlands were included from the “INPN Zones humides” dataset (Rapinel et al. 2023) as seasonal resources and refuges (Barasona et al. 2021). Finally, we integrated coarse regional hunting-bag statistics (departmental counts by OFB, Aug 2025) as a spatial proxy for hunting pressure and associated behavioral shifts, activated only in winter models (Calenge et al. 2002; Keuling et al. 2008; Thurfjell et al. 2013). All predictors were harmonized to the 1 km² grid. Data provenance, processing details and correlation matrices as basis for collinearity feature pruning are reported in SI S4. SDM Modeling Strategy We modeled habitat suitability separately for summer and winter using two complementary SDM engines: MaxEnt and gradient-boosted decision trees (GBM). MaxEnt is a widely used presence-only approach with strong theoretical grounding and a long record of robust performance when sampling bias is addressed and model complexity is controlled (Phillips et al. 2006; Elith et al. 2011). GBMs can capture nonlinearities and interactions and provide a flexible benchmark against a regularized, more interpretable model class. We implemented MaxEnt (Phillips et al., 2006) through the elapid Python 3 interface (Anderson 2023), while the kernel is based on glmnet ensuring compatibility and comparability of models and results to the popular open-source R release maxnet (Phillips et al. 2017). We implemented GBM in Python 3 using the LightGBM library (Ke et al. 2017). Multitemporal calibration of SDMs has often been suggested as a pathway for overcoming existing SDM limitations (Reside et al. 2010; Martínez-Minaya et al. 2018; Eduardo et al. 2022). To test whether exposing the learner to interannual environmental variability improves robustness, we compared (i) a multitemporal calibration in which training data retain year-specific predictors, and (ii) a monotemporal baseline in which predictors are averaged across training years into an “average year” representation comparable to the monotemporal approach of Bosch et al. (2014) or Guisan & Thuiller (2005). We evaluated temporal transferability using a leave-one-year-out (LOYO) design: models were trained on all but one year and then projected to the held-out year using its corresponding predictor stack. For each LOYO fold, we sampled 15,000 background points per year from the composite bias surface and fitted models to presences and weighted background points. Hyperparameters were tuned using automated search (Optuna) within the training data, with spatial blocking used during tuning to reduce optimistic bias from spatial autocorrelation (Roberts et al. 2017; Valavi et al. 2019). Full search spaces, convergence settings, and calibration details are reported in SI S5. Model outputs are transformed to a continuous habitat-suitability index (0–1) using the cloglog transformation commonly used for MaxEnt-style presence-only predictions, facilitating interpretation as relative suitability as an approximate probability of presence (Phillips and Dudík 2008). Evaluation and Uncertainty Quantification For each year/season combination run we calculate the Continuous Boyce Index (CBI, Spearman variant) (Boyce et al. 2002; Hirzel et al. 2006) as our main presence-only diagnostic quantifying how much the predicted suitability values deviate from a random expectation, given the distribution of presence points across prediction bins. We complementarily calculate model discrimination capability expressed as the Area under the Receiver Operating Characteristic Curve (AUC). Because the wild boar is a widespread generalist with few clearly unsuitable areas, rank-based presence-only metrics are often more informative than discrimination metrics relying on pseudo-absences (see chapter Discussion: SDM Metrics). We report AUC for comparability with earlier SDM literature and for computing permutation-based feature importance. We additionally calculate variable response curves by systematically varying one predictor at a time while keeping other predictors constant at their mean values (SI S6). We quantify environmental novelty and projection extrapolation capability of our models with two complementary indices: (i) For novelty type 1 (NT1) – values outside the univariate training ranges – we compute the Multivariate Environmental Similarity Surface (MESS) (Elith et al. 2010), which evaluates each pixel against the full distribution of training values per predictor and reports the minimum similarity across variables. (ii) For novelty type 2 (NT2) – novel combinations of predictors even when each falls within its marginal range – we compute ExDet-NT2 (Mesgaran et al. 2014) as a Mahalanobis distance-based measure relative to the training mean covariance structure (scaled by the maximum training distance). Thresholds are specified in Fig. 10. Results Vision Transformer Fine-tuning the ViT on multispectral satellite imagery successfully predicted high-resolution (30 m) annual LC maps for France, distinguishing 12 ecologically relevant habitat classes with strong metrics (Table 1): Mean Intersection-over-Union (IoU) = 47.7 %, mean pixelwise class accuracy (mAcc) = 71.8 %. Achieved values slightly surpassed the multitemporal crop segmentation achieved in the original reference (Jakubik et al. 2023). Detailed results, confusion matrices and an evaluation against 1k manually annotated points are available in SI S3c. The resulting LC maps demonstrate a clear and detailed delineation of regional landscape patterns and class-specific distributions (Fig. 5). On a national scale, broad spatial gradients are captured accurately (Fig. 5 top right). The zoomed-in section (Fig. 5 bottom right) illustrates precise classification of fine-scale features, such as river corridors and crop-field mosaics. Overall, the annual LC maps effectively capture the interannual and seasonal variability in vegetation dynamics and crop rotations, supporting nuanced habitat assessments at scale. Producing such maps requires significant computational resources (~50 GPU-hours per annual inference run), indicating a clear trade-off between spatial detail and operational efficiency. Nonetheless, the achieved classification quality and ecological relevance justify the computational cost, providing robust inputs for subsequent wild boar habitat suitability modeling. Per-class performance varied (Table 1), with notably high IoU for water (83.4 %), wheat (61.6 %), rapeseed (57.9 %), urban areas (53.1 %) and coniferous forest (47.3 %). Conversely, mixed and beech forests showed lower IoU scores (26.9 % and 21.3 %, respectively), indicating intra-domain variability (evident when assessing the confusion matrix, see SI S3c). Accuracy was consistently high (>78 %) for water, rapeseed, corn/maize, and coniferous forests, reflecting strong model differentiation potential. Table 1 - Classification metrics after fine-tuning the ViT for 120 Epochs. Class IoU [%] Accuracy [%] Wheat 61.59 78.32 Corn / Maize 55.05 84.11 Barley 38.57 72.85 Rapeseed 57.91 85.24 General Agriculture 32.42 38.25 Grassland 52.84 66.71 Oak Grove 34.08 70.21 Coniferous Forest 47.28 84.61 Mixed Forest 26.87 29.95 Beech Grove 21.25 80.45 Urbanization 53.14 79.77 Water 83.37 91.04 Total 47.69 (mIoU) 71.79 (mAcc) 60.25 (aAcc) Multitemporal Model Performance Across all LOYO folds, MaxEnt models exhibited consistently higher CBI values than GBM models under every configuration (Table 2). MaxEnt’s mean CBI (Spearman) was very high for both summer and winter, typically ~0.88–0.98, whereas GBM’s mean CBI ranged lower (~0.56–0.83). MaxEnt also showed much greater stability across years, with low standard deviation (often 0.02–0.10) indicating consistent performance on each held-out year. In contrast, GBM’s CBI was highly variable (SD up to ~0.5), and in some test years the GBM model’s CBI dropped to near zero or even negative. For example, the winter monotemporal GBM had one fold with a negative CBI (≈ –0.35), indicating worse-than-random predictions for that year’s data. Notably, MaxEnt never produced a negative CBI in any fold, and even its worst-case test fold remained moderately high (CBI ≈ 0.67). These results demonstrate that the MaxEnt engine achieved stronger generalization capacity than GBM across temporal folds. Table 2 – Summary of SDM results on test data from LOYO temporal CV. High values for CBI > 0.8 printed bold. Data Input Season Temporal Setup Model Engine Mean CBI ± SD Max CBI WVCs+GBIF Summer Monotemporal MaxEnt 0.943 ± 0.042 0.991 WVCs+GBIF Summer Monotemporal GBM 0.756 ± 0.418 1.000 WVCs+GBIF Summer Multitemporal MaxEnt 0.885 ± 0.119 0.984 WVCs+GBIF Summer Multitemporal GBM 0.602 ± 0.338 0.977 WVCs+GBIF Winter Monotemporal MaxEnt 0.967 ± 0.028 0.989 WVCs+GBIF Winter Monotemporal GBM 0.653 ± 0.548 0.985 WVCs+GBIF Winter Multitemporal MaxEnt 0.947 ± 0.093 0.993 WVCs+GBIF Winter Multitemporal GBM 0.560 ± 0.314 0.904 GBIF_only Summer Monotemporal MaxEnt 0.941 ± 0.045 0.991 GBIF_only Summer Monotemporal GBM 0.694 ± 0.385 0.986 GBIF_only Summer Multitemporal MaxEnt 0.914 ± 0.071 0.985 GBIF_only Summer Multitemporal GBM 0.603 ± 0.393 0.967 GBIF_only Winter Monotemporal MaxEnt 0.977 ± 0.016 0.991 GBIF_only Winter Monotemporal GBM 0.795 ± 0.131 0.961 GBIF_only Winter Multitemporal MaxEnt 0.976 ± 0.021 0.991 GBIF_only Winter Multitemporal GBM 0.831 ± 0.197 0.981 Despite these stark differences in CBI, the Area Under the Curve (AUC) on withheld-year data was acceptably high for both modeling approaches (full fold-wise AUC results are provided in SI S7) MaxEnt models achieved test AUCs typically in the 0.60–0.72 range for both summer and winter scenarios, while GBM models’ test AUC ranged roughly 0.62–0.75, occasionally reaching up to ~0.80 in winter multitemporal runs. In other words, both engines appeared to perform well by the AUC metric on validation data. However, the CBI reveals a crucial difference: GBM’s high AUC did not translate into ecologically reliable predictions. GBM models often attained very high training AUC (≈ 0.93), indicating a near-perfect fit to the training presences, yet their validation CBI was poor. This pattern reflects overfitting in the GBM models: they fit the training data extremely well but also failed to rank habitats correctly for unseen test data years. MaxEnt, by contrast, maintained only moderate training AUC (≈ 0.70), suggesting stronger regularization, but it retained high CBI on validation folds. Thus, MaxEnt models generalized better, yielding higher validation CBI despite similar (or slightly lower) AUC than GBM – a sign that MaxEnt predictions were more ecologically consistent. In terms of seasonal and temporal settings, monotemporal vs. multitemporal models performed similarly overall. Any differences due to temporal setup were small compared to the large engine effect. MaxEnt’s CBI remained high in both monotemporal and multitemporal modes (e.g. winter MaxEnt CBI ≈0.95–0.98 in both setups), and GBM’s performance, while lower, followed the same pattern (its best CBI occurred in monotemporal summer runs, but even there mean CBI was only ~0.75). Data input (WVCs+GBIF vs. GBIF_only) had minimal impact on these rank-based metrics. Models trained on the full presence dataset versus only GBIF records yielded very similar CBI outcomes (Table 2). This suggests that adding the extra WVC occurrence points (from sources beyond GBIF) did not drastically alter the models’ overall skill in predicting habitat suitability rankings. For MaxEnt in particular, CBI stayed consistently high regardless of data source. Feature Importance Because of the consistently higher CBI of MaxEnt over GBM models, permutation feature importance was only calculated for MaxEnt. Across seasons, the ten most important predictors consistently grouped into three clear ecological categories (Fig. 6 and 7): (i) proximity to mature broadleaf forest (particularly beech and oak), (ii) LC structure (fractional grassland and general agriculture; crop fractions and distances), and (iii) weather/energy constraints from ERA-5 (winter solar radiation; summer air temperature and shallow soil moisture/temperature). Winter Proximity to mature beech forests consistently emerged as the strongest predictor (approx. 15% across models). Initially, monotemporal models also strongly emphasized fractional grassland (approx. 11%) and general agriculture (approx. 11%). However, their importance decreased substantially under multitemporal calibration (grassland: Δ ≈ -4%; general agriculture: Δ ≈ -3%), indicating that their predictive value varies considerably across years. Small woody features may serve as a more relevant refuge in winter (up to 10%) than in summer (approx. 4%). In contrast, dynamic climatic predictors gained importance under multitemporal modeling. ERA-5 surface solar radiation, reflecting dynamic seasonal weather conditions, increased dramatically from around 4% in monotemporal models to 16% in multitemporal models (Δ ≈ +12%), becoming a critical predictor of interannual variation in winter habitat suitability. Moreover, incorporating structured WVC data notably reduced importance of features potentially influenced by observer bias: small woody features decreased from around 10% (GBIF_only monotemporal) to about 5% (WVC+GBIF multitemporal; Δ ≈ -5%), and dense urbanization (approx. 3%) dropped out entirely. Overall, multitemporal calibration combined with WVC data increased the cumulative predictive power of the top features from 70% to 73% (Δ ≈ +3%), underscoring improved focus on ecologically robust variables (Fig. 6). Summer During summer, multitemporal calibration emphasized dynamic climatic constraints even more distinctly. ERA-5 mean summer temperature rose markedly from around 6% importance in monotemporal models to about 15% in multitemporal models (Δ ≈ +9%), highlighting the influence of dynamic climatic variability on habitat selection. Fractional grassland decreased in importance under multitemporal models (from around 12% monotemporal to around 6% multitemporal; Δ ≈ -6%), suggesting high year-to-year variability in grassland suitability. Furthermore, a notable ecological shift occurred with multitemporal modeling: the importance of proximity to maize (approx. 8% monotemporal) was replaced by proximity to oak stands (approx. 8% multitemporal; Δ ≈ +8%), suggesting that once interannual variability is integrated, proximity to mast-producing broadleaf stands may supersede crop adjacency as a predictor. Proximity to beech similarly from about 8% monotemporal to around 4% multitemporal (Δ ≈ -4%), indicating further seasonal differences in resource reliance. After mast year events oak stands might offer extended availability of acorns into the following spring and summer compared to beech nuts which spoil more rapidly (Brandt et al. 2006). Wheat forms a stable summer resource (approx. 6-8%). Small woody features and natural heath and altitude provide modest, steady contributions (approx. 3-5% each). Overall, cumulative predictive importance slightly decreased from monotemporal (approx. 69%) to multitemporal (approx. 61%, Δ ≈ ‑8%), reflecting a more balanced distribution of predictor importance, clearly distinguishing stable landscape resources from dynamically varying climatic variables. Discussion Ecological Inference Visual inspection of the seasonal predictions (Fig. 8) reveals concordant hotspots in both summer and winter: Provence-Côte d’Azur, the Pyrenees, the forests of Île-de-France, and the Alsace/Vosges chain. Large riparian corridors (e.g. Loire and Rhône) consistently exhibit elevated suitability. Persistently poor habitats include the high-altitude Alps and the open arable plains of Champagne, which lack shelter and diverse forage. Season-specific shifts were evident: The winter maps (Fig. 8 top) uniquely showed increased suitability in the milder Atlantic regions of Bretagne and Normandie, suggesting these areas might serve as important winter refuges. Conversely, the extensive coniferous forests of Aquitaine exhibited depressed suitability in winter compared to summer (Fig. 8 bottom), indicating that some large forest tracts are less utilizable during winter months (possibly due to reduced forage or ongoing disturbances such as hunting). Elevated suitability in peri-urban belts persisted even after bias correction, mirroring empirical reports of risk-tolerant behavior and stable urban or suburban boar groups in European cities (Stillfried et al. 2017; Marin et al. 2024). Dense centers remained less suitable than surrounding suburban greenbelt mosaics. Specifically, our models suggest elevated suitability hotspots in the wider suburban areas of the South-West (Marseille, Toulon, Nice, Montpellier, Narbonne, Perpignan) as well as Paris, Bordeaux, Toulouse and Nantes during both summer and winter. In the Upper Rhine valley (Strasbourg, Mulhouse) peri-urban suitability is specifically strong in the winter months. We note that generally most of continental France has been consistently mapped as an at least moderately suitable habitat for wild boar. A stronger concentration of hotspots during winter can be attributed to coastal zones, heterogenous lowlands with a moderate to warm climate, as well as more urbanized areas. Summer signal spreads out more evenly favoring some additional forest-agriculture interfaces reducing the least suitable areas too very singular stretches. These interpretations are consistent with known wild boar ecology: preferred use of forest-agriculture ecotones, mast-producing broadleaf stands, riparian corridors, and an increased reliance on peri-urban resources (Schley and Roper 2003; Brandt et al. 2006; Thurfjell et al. 2009; Stillfried et al. 2017; Marin et al. 2024). SDM Metrics Across all LOYO folds, MaxEnt produced higher and markedly more stable CBI than GBM, and its suitability maps were spatially coherent and ecologically interpretable (Fig. 8). This pattern is expected for presence-only data when models are strongly regularized in the feature space and evaluated with a rank-based metric (Phillips and Dudík 2008; Elith et al. 2011). By contrast, boosted trees can achieve very high in-sample discrimination yet exhibit poor temporal transfer if capacity is not tightly constrained relative to sample size and predictor cardinality. In our case, GBM frequently reached training AUCs > 0.9 but yielded mediocre or unstable CBIs on withheld years, whereas MaxEnt retained moderate training AUCs (~0.7) and high validation CBI values. This divergence illustrates why the CBI is the primary criterion for presence-only SDMs: CBI asks whether observed presences accumulate in high-suitability bins more than expected by chance and does not rely on a comparison against pseudo-absences (Hirzel et al. 2006; Di Cola et al. 2017). We therefore adhere to recommendations preferring the CBI for model comparison and ecological interpretation because the CBI does not require systematically collected absence data and is comparatively insensitive to prevalence and background sampling definition (Lobo et al. 2008; Jiménez-Valverde 2012; Jiménez and Soberón 2020). The AUC, although still widely reported, remains sensitive to the geographic extent of the background, species prevalence, and pseudo-absence design, and it can inflate apparent skill especially for widespread generalists or penalize ecologically sensible maps (Lobo et al. 2008; Jiménez-Valverde 2012; Warren et al. 2020). Additionally, generalist species like wild boar, by definition, exhibit broader ecological niches, occupying a wide range of environmental conditions, and typically have few clearly unsuitable areas in France. AUC relies on sharply contrasting presence-background separation and tends to be less sensitive in such a context (Warren et al. 2020). CBI instead quantifies how well the model’s predicted suitability corresponds to actual species use patterns across continuous gradients of suitability, providing a more nuanced and ecologically meaningful assessment of predictive accuracy (Paudel et al. 2015; Milano et al. 2024). The added benefit of WVC datasets Rail and road collision records supplied systematic, high-resolution presences along linear infrastructure – an observation process that complements the observer-biased reach of GBIF (Seiler 2004; Langbein et al. 2011; Jasińska et al. 2019; Schwartz et al. 2020). The winter 2019 difference map (Fig. 9) illustrates the practical effect of the integration of WVC datasets: Generally, consensus between WVCs+GBIF and GBIF_only models is high with ΔHSI between -0.2 and 0.2 for over 96.5% of the study area (mean 0.007 ± 0.094 SD). When WVCs were included, predicted suitability increases within the conifer-dominated forests of Aquitaine and in mountainous regions like the Pyrenees, the Massif Central, Provence-Côte d’Azur, Languedoc and the pre-Alps (Chartreuse, Savoies, Jura). Inflated GBIF signals in the West (Bretagne, Pays-de-la-Loire, Limousin-Indre) were down-regulated, most notably in a heavily sampled zone northwest of Nantes. This is also reflected by the disappearance of the “dense urbanization” proxy from Top10 feature importance, once WVCs are included. WVCs emphasize habitats where wild boar actually move across landscapes, capturing dispersal and risk corridors (e.g., beech and oak proximity, fragmentation, grassland-agriculture mosaics). Consequently, forest-edge and movement-related predictors (distance to beech, ERA-5 energy variables) gain relative weight, while purely observational correlates (small woody features, urban bias) diminish. These corrections are ecologically coherent: in winter, boar movements and resource tracking bring animals into contact with transport corridors and montane forest refugia, which collision data detects but GBIF_only underrepresents. WVCs also yielded a sharper more concentrated feature signal in both seasons, indicating a model that explains variance through a smaller, more coherent and interpretable predictor subset. Conversely, GBIF potentially overemphasizes easily accessible zones (Beck et al. 2014; Hughes et al. 2021). While both WVCs+GBIF as well as GBIF_only models yield high scalar metrics, the net effect is not a dramatic performance change but a material improvement in the spatial realism of predicted conflict corridors and seasonal refugia (Morelle et al. 2013, 2014). Multitemporal Modeling Advances Extrapolation Capability Because of consistently more robust performance of MaxEnt models over GBM we restricted the environmental novelty analysis to MaxEnt, calculating flagged fractional pixel count of the total AoI: NT1 if MESS 1, indicating novel predictor combinations outside training ranges (Fig. 10 top). Overall, multitemporal models consistently reduced novelty (NT1) compared to monotemporal approaches, especially under typical climatic conditions. During the winters 2017–2022 monotemporal models generally showed moderate NT1 novelty, ranging from 1–6 %, but spiked sharply to 10% in the anomalously warm winter of 2023. In contrast, multitemporal winter models remained consistently low in novelty (<0.2 %), rising only modestly to 2% in 2023. Summer scenarios exhibited more pronounced variability. Monotemporal NT1 novelty remained relatively low (~3%) from 2018 to 2020, but surged dramatically to 32% during the anomalous summer of 2021. It subsequently decreased to 5% in 2022, stabilizing around 9% in 2023. Multitemporal summer models maintained substantially lower novelty overall (baseline 0.02–0.04%), but still showed brief increases during anomalies (5% in 2021, 1% in 2022) before returning to baseline levels (0.07 %) in 2023. NT2 novelty, representing uncertain predictor combinations, remained consistently low (0.4–1 %) across all scenarios and training schemes. Distinct spatial patterns of extrapolation further support these findings (Fig. 10, bottom panel): In the anomalous summer of 2021, severe flooding in northeastern France triggered high NT1 novelty in monotemporal models, notably along the Moselle and Meuse valleys. In contrast, multitemporal models showed only isolated novelty patches. During winter 2023/24, the warmest global winter on record, monotemporal models again indicated widespread NT1 novelty across northern France due to extreme temperatures. Conversely, multitemporal models exhibited minimal extrapolation, underscoring their robustness in capturing broader climatic variability. NT2 novelty (novel predictor combinations) was consistently limited to small, localized clusters, primarily in mountainous areas and densely populated city centers, and remained nearly identical across all seasons and modeling approaches. Comparing predictive performance by CBI, differences between mono- and multitemporal models were modest on average (Table 2), yet their consequences were substantial for robustness. Novelty diagnostics showed that multitemporal calibration consistently curtailed true extrapolation both spatially and quantitatively. These findings support the argument that exposing the learner to multi-year variability widens the realized environmental space, improves temporal transfer, and increases resilience to non-stationarity (Franklin 2010; Martínez-Minaya et al. 2018; Eduardo et al. 2022). We therefore argue that multitemporal calibration of generalist species SDMs is becoming an increasingly necessary property under future rapid climate variability scenarios. Limitations and Outlook Despite robust methodological approaches, several important limitations warrant consideration. Our models currently estimate habitat suitability, but we do not directly translate our results into abundance or density metrics. To calibrate HSI into actual population sizes future modeling approaches can bridge this gap by integrating occupancy-detection frameworks (MacKenzie et al. 2017), distance sampling (Buckland et al. 2015), or spatial capture-recapture methods (Royle et al. 2013). For instance, camera trapping setups or GPS collar telemetry combined with capture-recapture approaches provide the necessary spatially explicit capture histories, enabling the conversion of habitat suitability into population density estimates at fine spatial scales. While we are conducting a subsequent camera trap study alongside infrastructure corridors, we argue that the rapid reproductive capacity and high ecological plasticity of wild boar typically result in swift occupancy of suitable niches. Consequently, our HSI can be considered directly informative and operationally useful for guiding proactive management decisions, even in the absence of direct density estimates. Presence-only inference remains sensitive to residual bias and spatial dependence even under weighting; CBI as an evaluation method as well as bias correction techniques mitigate but cannot completely eliminate these issues (Syfert et al. 2013; Dubos et al. 2022; Baker et al. 2024). The benefits of such corrections are context-dependent: They can reduce redundancy and pseudo-replication, but point thinning always discards information and does not necessarily guarantee better transferability or faithful response curves (Fourcade et al. 2014; Aiello-Lammens et al. 2015; Ten Caten and Dallas 2023). A higher spatial resolution might increase environmental coverage at the potential cost of loss of generalizability. The spatial resolution of the current predictor stack, aggregated to a 1 km² grid, may also overlook finer-grained ecological processes or landscape elements critical to wild boar resource use, such as small habitat patches, agricultural edges, or localized human disturbances. Hence, leveraging finer-resolution predictors or hierarchical multi-scale modeling approaches (Johnson et al., 2004; DeCesare et al., 2012) could further enhance ecological realism. For example, models may not fully account for rapidly changing anthropogenic factors such as intensified hunting pressures, agricultural shifts, rapid urban expansion, or sudden infrastructure developments, all of which can dynamically alter habitat suitability beyond the captured predictor variability. Our extrapolation novelty analysis shows that integrating multitemporal environmental scenarios can in principle capture a wider range of such potential scenarios and therefore can offer a suitable pathway for navigating future environmental or anthropogenic extremes. Additionally, integrating Artificial Neural Networks (ANNs), particularly Convolutional Neural Networks (CNNs) or ViTs, as an SDM engine represents a natural and beneficial progression in our framework. Neural networks excel in capturing complex nonlinear relationships, interactions, and spatiotemporal patterns that conventional SDMs might overlook. Early fusion approaches (such as integrating LC maps and satellite-derived phenological indicators at high resolution directly into a fully trainable ANN SDM) could maximize the information content and predictive accuracy. These networks might efficiently leverage multi-scale spatial information, analyzing both the immediate and contextual landscape configurations around occurrence points, offering substantial ecological interpretability improvements. However, robust ANN implementations potentially require even larger and richer datasets. As WVCs involving wild boar are steadily rising and reporting protocols by SNCF and other agencies are improving, we anticipate that future data will enhance spatiotemporal resolution and comprehensiveness, strengthening the predictive capability and applicability of such models. Applying our seasonal suitability layers makes it possible to time and place interventions with greater precision: (i) Transport agencies can prioritize winter-risk segments for fencing, signage and crossing structures where models reveal recurrent WVC corridors (Seiler 2004; Langbein et al. 2011). Agricultural services can anticipate summer hotspots around energy-rich crops and deploy deterrents or target hunting accordingly (Ficetola et al. 2014). (ii) Veterinary authorities can focus surveillance and biosecurity along predicted movement corridors and refugia relevant to ASF spread (Morelle et al. 2020; Sauter-Louis et al. 2021). (iii) Urban planners should explicitly consider the suburban greenbelts as functional habitat and manage attractants, connectivity, and public communication accordingly (Stillfried et al. 2017; Marin et al. 2024). Because the pipeline can digest new satellite data and occurrence streams as soon as they become available, it is generally fast and flexible. On the other hand it also needs to be rerun periodically to operationally track short-term changes and provide early warning under evolving climatic or LU conditions. Conclusion Our results confirm that large portions of France provide highly suitable habitat for wild boar, underscoring the species' exceptional adaptability and generalist behavior. Although this adaptability poses considerable management challenges – especially regarding WVCs, agricultural conflicts, and urban encroachment – the presented modeling workflow offers a robust framework to identify and proactively mitigate conflict hotspots. Technically, the analysis demonstrates that MaxEnt-based multitemporal SDMs substantially outperform monotemporal models in terms of predictive robustness. They offer reduced extrapolation under future extreme climatic conditions and high spatial coherence of predicted wild boar habitat suitability. Incorporating rail and road collision datasets alongside GBIF observations increases data quantity, mitigates spatial biases and enhances realism. Ecologically, the persistent suitability of peri-urban and suburban greenbelts, even after rigorous bias correction, highlights urban habitats as a highly important refuges and resource-rich environments for wild boar populations in France. Although current limitations – such as spatial precision of collision data, absence of direct abundance estimation, and computational intensity of the ViT pipeline – persist, our approach offers concrete pathways for spatially and temporally explicit predictions on wild boar occurrence distribution. Collectively, this study underscores the critical importance of multitemporal data collection and modeling, careful bias correction, and diversified data integration. Our research has broad implications for wildlife conservation, transportation safety, agricultural management, veterinary epidemiology and environmental conflict mitigation. By bridging SDMs, RS, and ML, this study provides a data-driven foundation for evidence-based decision-making. Our research is not only applicable in France but largely builds on open satellite data which is universally available across diverse ecological and infrastructural landscapes worldwide. In the future the methodological insights from this research can be transferred further to other ungulate species that are frequently involved in HWCs and specifically WVCs, such as red deer ( Cervus elaphus ), roe deer ( Capreolus capreolus ), reindeer ( Rangifer tarandus ), elk ( Alces alces ), offering a transferable approach and generalizable pathway to data-driven wildlife management in human-modified landscapes. Declarations Author contributions A.M. and D.J. conceived the study. A.M. led the methodological development, performed the analyses, and drafted the manuscript. D.J. provided methodological input for modeling and statistics. T.R. supported data preprocessing and model implementation, especially for the Vision Transformer. M.K. contributed to the remote‑sensing components. K.M. contributed to ecological interpretation and management relevance. D.J. provided supervision and project oversight. All authors contributed critically to the drafts and approved the final manuscript. Data and Code availability GBIF occurrence records used in this study are publicly available from the Global Biodiversity Information Facility (GBIF). Railway collision records (SNCF Réseau) and road collision records (Réseau Routier National) were used under data-sharing agreements and are not publicly available; access may be granted by the respective data owners upon reasonable request. HLS satellite datasets are available globally and free of charge from NASA. Pretrained ViT model weights are available through huggingface. Processed predictor layers, trained model weights, and model outputs can be obtained from the corresponding author upon reasonable request. Access to the private code repository is restricted because of the intermediate data products necessary to run the scripts but may be granted from the corresponding author upon reasonable request. Funding This study was co-funded by SNCF Réseau (Paris, France) and the Institute Geomatics, FHNW (Muttenz, Switzerland). Competing interests The authors declare no competing interests. Ethical approval This study did not involve capture, handling, or experimental procedures on animals. References Acevedo P, Escudero MA, Muńoz R, Gortázar C (2006) Factors affecting wild boar abundance across an environmental gradient in Spain. Acta Theriol (Warsz) 51:327–336. https://doi.org/10.1007/BF03192685 Aiello-Lammens M, Boria R, Radosavljevic A, et al (2015) spThin: An R package for spatial thinning of species occurrence records for use in ecological niche models. Ecography 38:541–545. https://doi.org/10.1111/ecog.01132 Anderson CB (2023) elapid: Species distribution modeling tools for Python. J Open Source Softw 8:4930. https://doi.org/10.21105/joss.04930 Arthaud F, Lecoeur G, Pierre A (2024) Transformers à Grande Vitesse. J Rail Transp Plan Manag 29:100418. https://doi.org/10.1016/j.jrtpm.2023.100418 Baker DJ, Maclean IMD, Gaston KJ (2024) Effective strategies for correcting spatial sampling bias in species distribution models without independent test data. Divers Distrib 30:e13802. https://doi.org/10.1111/ddi.13802 Baker DJ, Maclean IMD, Goodall M, Gaston KJ (2022) Correlations between spatial sampling biases and environmental niches affect species distribution models. Glob Ecol Biogeogr 31:1038–1050. https://doi.org/10.1111/geb.13491 Ballari SA, Barrios-García MN (2014) A review of wild boar Sus scrofa diet and factors affecting food selection in native and introduced ranges. Mammal Rev 44:124–134. https://doi.org/10.1111/mam.12015 Barasona JA, Carpio A, Boadella M, et al (2021) Expansion of native wild boar populations is a new threat for semi-arid wetland areas. Ecol Indic 125:107563. https://doi.org/10.1016/j.ecolind.2021.107563 Barber RA, Ball SG, Morris RKA, Gilbert F (2022) Target-group backgrounds prove effective at correcting sampling bias in Maxent models. Divers Distrib 28:128–141. https://doi.org/10.1111/ddi.13442 Barbet-Massin M, Jiguet F, Albert CH, Thuiller W (2012) Selecting pseudo-absences for species distribution models: how, where and how many? Methods Ecol Evol 3:327–338. https://doi.org/10.1111/j.2041-210X.2011.00172.x Beck J, Ballesteros-Mejia L, Buchmann CM, et al (2012) What’s on the horizon for macroecology? Ecography 35:673–683. https://doi.org/10.1111/j.1600-0587.2012.07364.x Beck J, Böller M, Erhardt A, Schwanghart W (2014) Spatial bias in the GBIF database and its effect on modeling species’ geographic distributions. Ecol Inform 19:10–15. https://doi.org/10.1016/j.ecoinf.2013.11.002 Benard A (2023) Road ecology: toward a predictive model of wildlife-vehicle collisions. Univ Claude Bernard Lyon 1 Bosch J, Mardones F, Pérez A, et al (2014) A maximum entropy model for predicting wild boar distribution in Spain. Span J Agric Res 12:984–999. https://doi.org/10.5424/sjar/2014124-5717 Boyce MS, Vernier PR, Nielsen SE, Schmiegelow FKA (2002) Evaluating resource selection functions. Ecol Model 157:281–300. https://doi.org/10.1016/S0304-3800(02)00200-4 Brandt S, Baubet E, Vassant J, Servanty S (2006) Régime alimentaire du sanglier en milieu forestier de plaine agricole. Faune Sauvage 273:20–27 Buckland ST, Rexstad EA, Marques TA, Oedekoven CS (2015) Distance sampling: methods and applications. Springer Calenge C, Maillard D, Vassant J, Brandt S (2002) Summer and hunting season home ranges of wild boar (Sus scrofa) in two habitats in France. Game Wildl Sci 19 (4):281–301 Cantelaube P, Carles M (2014) Le registre parcellaire graphique : des données géographiques pour décrire la couverture du sol agricole. NOVAE - Ingénierie Savoir-Faire Innovants 58–64 Chamon LF, Ribeiro A (2017) Greedy sampling of graph signals. IEEE Trans Signal Process 66:34–47 Claverie M, Ju J, Masek JG, et al (2018) The Harmonized Landsat and Sentinel-2 surface reflectance data set. Remote Sens Environ 219:145–161. https://doi.org/10.1016/j.rse.2018.09.002 Croft S, Warren D, Blanco-Aguiar J, Smith G (2025) Predicting the distribution of common wild mammal species across Europe - are there sufficient occurrence data? Eur J Wildl Res 71:133. https://doi.org/10.1007/s10344-025-02014-2 Culshaw V, Mairal M, Sanmartín I (2021) Biogeography Meets Niche Modeling: Inferring the Role of Deep Time Climate Change When Data Is Limited. Front Ecol Evol 9:. https://doi.org/10.3389/fevo.2021.662092 de Oliveira G, Rangel TF, Lima-Ribeiro MS, et al (2014) Evaluating, partitioning, and mapping the spatial autocorrelation component in ecological niche modeling: a new approach based on environmentally equidistant records. Ecography 37:637–647. https://doi.org/10.1111/j.1600-0587.2013.00564.x Di Cola V, Broennimann O, Petitpierre B, et al (2017) ecospat: an R package to support spatial analyses and modeling of species niches and distributions. Ecography 774–787. https://doi.org/10.1111/ecog.02671 Dubos N, Préau C, Lenormand M, et al (2022) Assessing the effect of sample bias correction in species distribution models. Ecol Indic 145:109487. https://doi.org/10.1016/j.ecolind.2022.109487 Eduardo AA, Liparini A, Martinez PA, et al (2022) Assessing multitemporal calibration for species distribution models. Ecol Inform 71:101787. https://doi.org/10.1016/j.ecoinf.2022.101787 EEA/FOEN (2011) Landscape fragmentation in Europe. European Environment Agency (EEA), Swiss Federal Office for the Environment (FOEN), Copenhagen, Denmark Elith J, Kearney M, Phillips S (2010) The art of modelling range-shifting species. Methods Ecol Evol 1:330–342. https://doi.org/10.1111/j.2041-210X.2010.00036.x Elith J, Phillips SJ, Hastie T, et al (2011) A statistical explanation of MaxEnt for ecologists. Divers Distrib 17:43–57. https://doi.org/10.1111/j.1472-4642.2010.00725.x Enetwild Consortium, Croft S, Blanco-Aguiar JA, et al (2024) Modelling wild boar abundance at high resolution. EFSA Support Publ 21:8965E. https://doi.org/10.2903/sp.efsa.2024.EN-8965 Enetwild Consortium, Croft S, Smith G, et al (2018) Wild boar in focus: Review of existing models on spatial distribution and density of wild boar and proposal for next steps. EFSA Support Publ 15:1490E. https://doi.org/10.2903/sp.efsa.2018.EN-1490 Enetwild Consortium, Croft S, Smith G, et al (2019) Wild boar in focus: initial model outputs of wild boar distribution based on occurrence data and identification of priority areas for data collection. EFSA Support Publ 16:1533E Faucqueur L, Morin N, Masse A, et al (2019) A new Copernicus high resolution layer at pan-European scale: small woody features. In: Remote Sensing for Agriculture, Ecosystems, and Hydrology XXI. SPIE, pp 268–278 Ficetola GF, Bonardi A, Mairota P, et al (2014) Predicting wild boar damages to croplands in a mosaic of agricultural and natural areas. Curr Zool 60:170–179. https://doi.org/10.1093/czoolo/60.2.170 Fourcade Y, Engler JO, Rödder D, Secondi J (2014) Mapping Species Distributions with MAXENT Using a Geographically Biased Sample of Presence Data: A Performance Assessment of Methods for Correcting Sampling Bias. PLOS ONE 9:e97122. https://doi.org/10.1371/journal.pone.0097122 Franklin J (2010) Moving beyond static species distribution models in support of conservation biogeography. Divers Distrib 16:321–330. https://doi.org/10.1111/j.1472-4642.2010.00641.x Guisan A, Thuiller W (2005) Predicting species distribution: offering more than simple habitat models. Ecol Lett 8:993–1009. https://doi.org/10.1111/j.1461-0248.2005.00792.x Hansen MC, Potapov PV, Moore R, et al (2013) High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 342:850–853. https://doi.org/10.1126/science.1244693 He KS, Bradley BA, Cord AF, et al (2015) Will remote sensing shape the next generation of species distribution models? Remote Sens Ecol Conserv 1:4–18. https://doi.org/10.1002/rse2.7 Hirzel AH, Le Lay G, Helfer V, et al (2006) Evaluating the ability of habitat suitability models to predict species presences. Ecol Model 199:142–152. https://doi.org/10.1016/j.ecolmodel.2006.05.017 Hughes AC, Orr MC, Ma K, et al (2021) Sampling biases shape our view of the natural world. Ecography 44:1259–1269. https://doi.org/10.1111/ecog.05926 Iacolina L, Pertoldi C, Amills M, et al (2018) Hotspots of recent hybridization between pigs and wild boars in Europe. Sci Rep 8:17372. https://doi.org/10.1038/s41598-018-35865-8 Inglada J, Vincent A, Arias M, et al (2017) Operational High Resolution Land Cover Map Production at the Country Scale Using Satellite Image Time Series. Remote Sens 9:95. https://doi.org/10.3390/rs9010095 Jakubik J, Roy S, Phillips CE, et al (2023) Foundation Models for Generalist Geospatial Artificial Intelligence Jasińska KD, Żmihorski M, Krauze-Gryz D, et al (2019) Linking habitat composition, local population densities and traffic characteristics to spatial patterns of ungulate-train collisions. J Appl Ecol 56:2630–2640. https://doi.org/10.1111/1365-2664.13495 Jiménez L, Soberón J (2022) Estimating the fundamental niche: Accounting for the uneven availability of existing climates in the calibration area. Ecol Model 464:109823. https://doi.org/10.1016/j.ecolmodel.2021.109823 Jiménez L, Soberón J (2020) Leaving the area under the receiving operating characteristic curve behind: An evaluation method for species distribution modelling applications based on presence-only data. Methods Ecol Evol 11:1571–1586. https://doi.org/10.1111/2041-210X.13479 Jiménez-Valverde A (2012) Insights into the area under the receiver operating characteristic curve (AUC) as a discrimination measure in species distribution modelling. Glob Ecol Biogeogr 21:498–507. https://doi.org/10.1111/j.1466-8238.2011.00683.x Ke G, Meng Q, Finley T, et al (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In: Advances in Neural Information Processing Systems. Curran Associates, Inc. Keuling O, Stier N, Roth M (2008) How does hunting influence activity and spatial usage in wild boar Sus scrofa L.? Eur J Wildl Res 54:729–737. https://doi.org/10.1007/s10344-008-0204-9 Langbein J, Putman R, Pokorny B (2011) Traffic collisions involving deer and other ungulates in Europe and available measures for mitigation. In: Apollonio M, Andersen R, Putman R (eds) Ungulate Management in Europe: Problems and Practices. Cambridge University Press, Cambridge, pp 215–259 Lobo JM, Jiménez-Valverde A, Hortal J (2010) The uncertain nature of absences and their importance in species distribution modelling. Ecography 33:103–114. https://doi.org/10.1111/j.1600-0587.2009.06039.x Lobo JM, Jiménez-Valverde A, Real R (2008) AUC: a misleading measure of the performance of predictive distribution models. Glob Ecol Biogeogr 17:145–151. https://doi.org/10.1111/j.1466-8238.2007.00358.x MacKenzie DI, Nichols JD, Royle JA, et al (2017) Occupancy estimation and modeling: inferring patterns and dynamics of species occurrence. Elsevier Maillard D, Gaillard JM, Hewison M a. J, et al (2010) Ungulates and their management in France. Cambridge University Press, p 441 Marin C, Couderchet L, Le Campion G, Werno J (2024) Wildlife and the city. Modelling wild boar use of urban nature: Empirical contribution, methodological proposal. Urban Ecosyst 27:1291–1312. https://doi.org/10.1007/s11252-024-01510-8 Martínez-Minaya J, Cameletti M, Conesa D, Pennino MG (2018) Species distribution modeling: a statistical review with focus in spatio-temporal issues. Stoch Environ Res Risk Assess 32:3227–3244. https://doi.org/10.1007/s00477-018-1548-7 Massei G, Kindberg J, Licoppe A, et al (2015) Wild boar populations up, numbers of hunters down? A review of trends and implications for Europe. Pest Manag Sci 71:492–500. https://doi.org/10.1002/ps.3965 Mesgaran MB, Cousens RD, Webber BL (2014) Here be dragons: a tool for quantifying novelty due to covariate range and correlation change when projecting species distribution models. Divers Distrib 20:1147–1159. https://doi.org/10.1111/ddi.12209 Meyer A, Milani G, Lack N, et al (2025) Machine learning-based multimodal data fusion for the Swiss land use statistics. Eur J Remote Sens 58:2508420. https://doi.org/10.1080/22797254.2025.2508420 Meyer C, Weigelt P, Kreft H (2016) Multidimensional biases, gaps and uncertainties in global plant occurrence information. Ecol Lett 19:992–1006. https://doi.org/10.1111/ele.12624 Milano F, Casazza G, Galimberti A, et al (2024) Combining distribution modelling and phylogeography to understand present, past and future of an endangered spider. BMC Ecol Evol 24:106. https://doi.org/10.1186/s12862-024-02295-2 Morelle K, Bubnicki J, Churski M, et al (2020) Disease-Induced Mortality Outweighs Hunting in Causing Wild Boar Population Crash After African Swine Fever Outbreak. Front Vet Sci 7:. https://doi.org/10.3389/fvets.2020.00378 Morelle K, Lehaire F, Lejeune P (2014) Is wild boar heading towards movement ecology? A review of trends and gaps. Wildl Biol 20:196–205. https://doi.org/10.2981/wlb.00017 Morelle K, Lehaire F, Lejeune P (2013) Spatio-temporal patterns of wildlife-vehicle collisions in a region with a high-density road network. Nat Conserv 5:53–73. https://doi.org/10.3897/natureconservation.5.4634 Morelle K, Lejeune P (2015) Seasonal variations of wild boar Sus scrofa distribution in agricultural landscapes: a species distribution modelling approach. Eur J Wildl Res 61:45–56. https://doi.org/10.1007/s10344-014-0872-6 Muñoz-Sabater J, Dutra E, Agustí-Panareda A, et al (2021) ERA5-Land: a state-of-the-art global reanalysis dataset for land applications. Earth Syst Sci Data 13:4349–4383. https://doi.org/10.5194/essd-13-4349-2021 Paudel PK, Hais M, Kindlmann P (2015) Habitat suitability models of mountain ungulates: identifying potential areas for conservation. Zool Stud 54:e37. https://doi.org/10.1186/s40555-015-0116-9 Pettorelli N, Vik JO, Mysterud A, et al (2005) Using the satellite-derived NDVI to assess ecological responses to environmental change. Trends Ecol Evol 20:503–510. https://doi.org/10.1016/j.tree.2005.05.011 Phillips SJ, Anderson RP, Dudík M, et al (2017) Opening the black box: An open‐source release of Maxent. Ecography 40:887–893. https://doi.org/10.1111/ecog.03049 Phillips SJ, Anderson RP, Schapire RE (2006) Maximum entropy modeling of species geographic distributions. Ecol Model 190:231–259. https://doi.org/10.1016/j.ecolmodel.2005.03.026 Phillips SJ, Dudík M (2008) Modeling of species distributions with Maxent: new extensions and a comprehensive evaluation. Ecography 31:161–175. https://doi.org/10.1111/j.0906-7590.2008.5203.x Puissant A, Sellé A, Baghdadi N, et al (2019) The ‘urban’ component of the French Land Data and Services Centre (THEIA). In: 2019 Joint Urban Remote Sensing Event (JURSE). pp 1–4 Rapinel S, Panhelleux L, Lemercier B, et al (2023) Projet de cartographie nationale des milieux humides – Rapport méthodologique de la prélocalisation des milieux et des zones humides, de la cartographie des habitats et de l’évaluation des fonctions. Patrinat (OFB-MNHN-CNRS-IRD) Reside AE, VanDerWal JJ, Kutt AS, Perkins GC (2010) Weather, Not Climate, Defines Distributions of Vagile Bird Species. PLOS ONE 5:e13569. https://doi.org/10.1371/journal.pone.0013569 Roberts DR, Bahn V, Ciuti S, et al (2017) Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40:913–929. https://doi.org/10.1111/ecog.02881 Royle JA, Chandler RB, Sollmann R, Gardner B (2013) Spatial capture-recapture. Academic press Sanderson EW, Jaiteh M, Levy MA, et al (2002) The Human Footprint and the Last of the Wild: The human footprint is a global map of human influence on the land surface, which suggests that human beings are stewards of nature, whether we like it or not. BioScience 52:891–904. https://doi.org/10.1641/0006-3568(2002)052%5B0891:THFATL%5D2.0.CO;2 Santos SM, Carvalho F, Mira A (2017) Current knowledge on wildlife mortality in railways. Railw Ecol 11–22 Sauter-Louis C, Conraths FJ, Probst C, et al (2021) African Swine Fever in Wild Boar in Europe—A Review. Viruses 13:1717. https://doi.org/10.3390/v13091717 Schleimer A, Richart L, Drygala F, et al (2022) Introgressive hybridisation between domestic pigs (Sus scrofa domesticus) and endemic Corsican wild boars (S. s. meridionalis): effects of human-mediated interventions. Heredity 128:279–290. https://doi.org/10.1038/s41437-022-00517-1 Schley L, Roper TJ (2003) Diet of wild boar Sus scrofa in Western Europe, with particular reference to consumption of agricultural crops. Mammal Rev 33:43–56. https://doi.org/10.1046/j.1365-2907.2003.00010.x Schwartz ALW, Shilling FM, Perkins SE (2020) The value of monitoring wildlife roadkill. Eur J Wildl Res 66:18. https://doi.org/10.1007/s10344-019-1357-4 Seiler A (2004) Trends and spatial patterns in ungulate-vehicle collisions in Sweden. Wildl Biol 10:301–313. https://doi.org/10.2981/wlb.2004.036 Sicacha-Parada J, Steinsland I, Cretois B, Borgelt J (2019) Accounting for spatial varying sampling effort due to accessibility in Citizen Science data: A case study of moose in Norway Stillfried M, Gras P, Börner K, et al (2017) Secrets of Success in a Landscape of Fear: Urban Wild Boar Adjust Risk Perception and Tolerate Disturbance. Front Ecol Evol 5:. https://doi.org/10.3389/fevo.2017.00157 Syfert MM, Smith MJ, Coomes DA (2013) The Effects of Sampling Bias and Model Complexity on the Predictive Performance of MaxEnt Species Distribution Models. PLOS ONE 8:e55158. https://doi.org/10.1371/journal.pone.0055158 Ten Caten C, Dallas T (2023) Thinning occurrence points does not improve species distribution model performance. Ecosphere 14:e4703 Thurfjell H, Ball JP, Åhlén P-A, et al (2009) Habitat use and spatial patterns of wild boar Sus scrofa (L.): agricultural fields and edges. Eur J Wildl Res 55:517–523. https://doi.org/10.1007/s10344-009-0268-1 Thurfjell H, Spong G, Ericsson G (2013) Effects of hunting on wild boar Sus scrofa behaviour. Wildl Biol 19:87–93. https://doi.org/10.2981/12-027 Vajas P, Von Essen E, Tickle L, Gamelon M (2023) Meeting the challenges of wild boar hunting in a modern society: The case of France. Ambio 52:1359–1372. https://doi.org/10.1007/s13280-023-01852-1 Valavi R, Elith J, Lahoz-Monfort JJ, Guillera-Arroita G (2019) blockCV: An r package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models. Methods Ecol Evol 10:225–232. https://doi.org/10.1111/2041-210X.13107 Vetter SG, Puskas Z, Bieber C, Ruf T (2020) How climate change and wildlife management affect population structure in wild boars. Sci Rep 10:7298. https://doi.org/10.1038/s41598-020-64216-9 Wang L, Diao C, Lu Y (2025) The role of remote sensing in species distribution models: a review. Int J Remote Sens 46:661–685. https://doi.org/10.1080/01431161.2024.2421949 Warren DL, Matzke NJ, Iglesias TL (2020) Evaluating presence-only species distribution models with discrimination accuracy is uninformative for many applications. J Biogeogr 47:167–180. https://doi.org/10.1111/jbi.13705 Welander J (2000) Spatial and temporal dynamics of wild boar (Sus scrofa) rooting in a mosaic landscape. J Zool 252:263–271. https://doi.org/10.1111/j.1469-7998.2000.tb00621.x Additional Declarations No competing interests reported. Supplementary Files MeyerAetal2026SupplementWildBoarSDMFranceEJWRv02.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 08 Apr, 2026 Reviewers agreed at journal 12 Mar, 2026 Reviewers invited by journal 23 Feb, 2026 Editor assigned by journal 23 Feb, 2026 Submission checks completed at journal 23 Feb, 2026 First submitted to journal 05 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8798859","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":596750554,"identity":"89b32822-4323-48e9-b814-d58ec57f7fdf","order_by":0,"name":"Adrian Ferdinand Meyer","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3UlEQVRIiWNgGAWjYFCCBIYDcPYHGzCbjXgtjDPSiNQCB8w8xGjhb899eODnDoZofonkY59tEu7kGRxgfvYAnxaJM88NDvaeYcidOSMteXZOwrNigwNs5gZ4rbmRxnCAt40hd8ONHGPm3B+HE7cd4GGTwKdDHqjl4F+glv038j8zWyQQocUAqOUw2BaJHGZmBmK0GJ55xnBYtk0id8aZZ8aMPUAt+w+zmeHVInc8jfnj2zab3P725McMP4BaZrY3P8OrBQqAagQSoGxmItRDAP8BopWOglEwCkbBCAMAj7JQZO7ILzIAAAAASUVORK5CYII=","orcid":"","institution":"University of Applied Sciences and Arts Northwestern Switzerland FHNW","correspondingAuthor":true,"prefix":"","firstName":"Adrian","middleName":"Ferdinand","lastName":"Meyer","suffix":""},{"id":596750561,"identity":"89964cb8-9bc8-4fa3-9af0-be43373b3c90","order_by":1,"name":"Théo Reibel","email":"","orcid":"","institution":"University of Applied Sciences and Arts Northwestern Switzerland FHNW","correspondingAuthor":false,"prefix":"","firstName":"Théo","middleName":"","lastName":"Reibel","suffix":""},{"id":596750569,"identity":"c67e9548-7524-4865-9e1a-ec5c0c89d899","order_by":2,"name":"Kevin Morelle","email":"","orcid":"","institution":"Max Planck Institute of Animal Behavior","correspondingAuthor":false,"prefix":"","firstName":"Kevin","middleName":"","lastName":"Morelle","suffix":""},{"id":596750577,"identity":"a07bb1f4-129d-45b9-bfe8-88ef8bcfd842","order_by":3,"name":"Mathias Kneubühler","email":"","orcid":"","institution":"University of Zurich","correspondingAuthor":false,"prefix":"","firstName":"Mathias","middleName":"","lastName":"Kneubühler","suffix":""},{"id":596750586,"identity":"53241342-905d-4a01-a1aa-26ab0c0d5700","order_by":4,"name":"Denis Jordan","email":"","orcid":"","institution":"University of Applied Sciences and Arts Northwestern Switzerland FHNW","correspondingAuthor":false,"prefix":"","firstName":"Denis","middleName":"","lastName":"Jordan","suffix":""}],"badges":[],"createdAt":"2026-02-05 15:25:48","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8798859/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8798859/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103520663,"identity":"8fb6aff4-a86e-4ae1-b896-4adcadee02d0","added_by":"auto","created_at":"2026-02-26 15:05:53","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4808562,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic processing overview. Datasets in italic font and rounded frames. Processes in white text on black rectangles.\u003c/p\u003e","description":"","filename":"Fig01Overview.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/ff1212c6cdc4d1c8fd5f6ec4.png"},{"id":104397861,"identity":"652ab20c-9d9f-4927-9d03-f60e83f721e8","added_by":"auto","created_at":"2026-03-11 11:58:18","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":8707719,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of geographic distributions of presence points from various data sources. A and B: All points prior to filtering; all available years combined for visualization. C and D: Remaining points after thinning; all available years combined for visualization.\u003c/p\u003e","description":"","filename":"Fig02PointThinning.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/cab7144abb873b97d45ddaa6.png"},{"id":104397500,"identity":"eb4e05c1-bfb3-4a8a-8f1c-0c40adec81da","added_by":"auto","created_at":"2026-03-11 11:49:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":7934139,"visible":true,"origin":"","legend":"\u003cp\u003eBias maps for background sampling. A is a composite example of seasonally explicit derivates of B, C and D. The latter three are shown here as combined maps over all seasons from 2017 to 2023.\u003c/p\u003e","description":"","filename":"Fig03BiasMaps.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/99893896f31e14871bc082aa.png"},{"id":103520671,"identity":"21aa11ad-9664-4f72-bca1-0eadd3bf9c4b","added_by":"auto","created_at":"2026-02-26 15:05:54","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":3772093,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of Environmental Feature Processing Workflow. Environmental features are preprocessed into six domains and assembled into a multitemporal feature stack. Before being trained in SDM pipelines the stack is analyzed for correlation and pruned.\u003c/p\u003e","description":"","filename":"Fig04FeaturePreprocessing.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/100deb5316f516f1921cbb1b.png"},{"id":104397640,"identity":"a86f90fc-1bab-4415-aa32-902185411332","added_by":"auto","created_at":"2026-03-11 11:53:38","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":11754131,"visible":true,"origin":"","legend":"\u003cp\u003eExample result of the ViT LC classification for the year 2020 for the whole study area (top) and a zoomed-in section in the central West (bottom). Example input False Color Infrared (FCIR) visualization of multitemporal HLS tile bottom left.\u003c/p\u003e","description":"","filename":"Fig05VisionTransformer.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/5d0d1411441229010e253550.png"},{"id":104397601,"identity":"ae959ae6-3269-4403-a1d9-10512fd448f8","added_by":"auto","created_at":"2026-03-11 11:52:41","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1425272,"visible":true,"origin":"","legend":"\u003cp\u003eWinter models: Partial importances of the 10 most important feature variables by permutation. Average over LOYO test data.\u003c/p\u003e","description":"","filename":"Fig06ImportanceWinter.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/e79505e375ae7125ac6fd00b.png"},{"id":104397501,"identity":"e5af142e-c561-4518-b210-208c020f1478","added_by":"auto","created_at":"2026-03-11 11:49:55","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1367077,"visible":true,"origin":"","legend":"\u003cp\u003eSummer models: Partial importances of the 10 most important feature variables by permutation. Average over LOYO test data.\u003c/p\u003e","description":"","filename":"Fig07ImportanceSummer.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/3a9a38a10e79d2f3b87314a4.png"},{"id":104397648,"identity":"91b8359b-8fe5-41f1-a940-54078470e045","added_by":"auto","created_at":"2026-03-11 11:53:45","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":18796911,"visible":true,"origin":"","legend":"\u003cp\u003eHabitat Suitability Index (HSI) by various modeling approaches with predictions on example LOYO data of the year 2019. This year was chosen for visualization because it is not characterized by any environmental extremes. The full LOYO time series over all available years can be found in SI S8. AL: Alps; AQ: Aquitaine; AV: Alsace/Vosges; BR: Bretagne; CH: Champagne; IF: Île de France; LO: Loire; NO: Normandie; PR: Provence-Côte d’Azur; PY: Pyrenees; RH: Rhône.\u003c/p\u003e","description":"","filename":"Fig08SDMResultMaps.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/8b3c508c24a8e5d2534106bb.png"},{"id":104398282,"identity":"49af0223-f2de-4a19-b12b-0e1d654155ea","added_by":"auto","created_at":"2026-03-11 12:01:16","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":7741006,"visible":true,"origin":"","legend":"\u003cp\u003eDifference map of multitemporal winter models highlighting ΔHSI variation based on input dataset composition. Examplary data for winter 2019. AQ: Aquitaine; BR: Bretagne; CR: Chartreuse; JU: Jura; LA: Languedoc; LI: Limousin-Indre; MC: Massif Central; PL: Pays-de-la-Loire; PR: Provence-Côte d’Azur; SA: Savoies.\u003c/p\u003e","description":"","filename":"Fig09DifferenceMap.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/0242e66c7f9955c086b49a59.png"},{"id":104398192,"identity":"92492fab-7075-451d-aef6-16c2406a100e","added_by":"auto","created_at":"2026-03-11 12:00:24","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":2719764,"visible":true,"origin":"","legend":"\u003cp\u003eTop: Novelty Pixels flagged for NT1 if MESS \u0026lt; -10 and for NT2 if ExDet \u0026gt; 1 (area percentage in logarithmic scale) on MaxEnt LOYO test years. Bottom: Example maps of the spatial distribution of novelty components NT1 and NT2. Other years can be found in SI S9.\u003c/p\u003e","description":"","filename":"Fig10ExtrapolationNovelty.png","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/1b6a3b22adacbe109303ffa4.png"},{"id":104407614,"identity":"04028334-3d3c-41f9-8fd1-39aa1fb66308","added_by":"auto","created_at":"2026-03-11 12:39:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":77581119,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/668df773-639f-417e-aed2-ba4d63443a7b.pdf"},{"id":103520673,"identity":"767ce270-90a2-43ab-a2fb-ebbef5250879","added_by":"auto","created_at":"2026-02-26 15:05:54","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":30710516,"visible":true,"origin":"","legend":"","description":"","filename":"MeyerAetal2026SupplementWildBoarSDMFranceEJWRv02.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8798859/v1/c9c2d373d380ed3805994a85.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Wild Boar Collision Data and Satellite Computer Vision Refine Habitat Suitability Mapping across France","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAcross Europe, expanding ungulate populations and the proliferation of infrastructure and agriculture have significantly intensified human-wildlife conflicts (HWCs). Especially wildlife-vehicle collisions (WVCs), crop damage, and disease transmission in wild boar \u003cem\u003e(Sus scrofa)\u0026nbsp;\u003c/em\u003eexemplify these conflicts (Langbein et al. 2011; Massei et al. 2015). Due to its rapid population growth and ecological adaptability, wild boar increasingly exploits heterogeneous, human-dominated landscapes, including agricultural mosaics, suburban areas, and transport networks (Morelle et al. 2014; Stillfried et al. 2017; Barasona et al. 2021). These conflicts highlight transport safety concerns, but elevated interactions between wild boar and humans also underline broader implications for veterinary epidemiology (e.g. African Swine Fever, ASF), agriculture, and wildlife management (Sauter-Louis et al. 2021). Modeling the spatiotemporal extent of wild boar distribution represents an expedient monitoring approach.\u003c/p\u003e\n\u003cp\u003eSpecies distribution models (SDMs) provide a practical way to translate heterogeneous data into spatial predictions of relative suitability or occurrence, and they are not only widely used in wildlife ecology and management (Elith et al. 2011), but can also be specifically applied to the wild boar context at national scale (Bosch et al. 2014). Effectively managing wild boar populations and conflicts requires dynamic models capable of capturing spatiotemporal shifts in habitat use. Wild boar movements and habitat preferences vary seasonally with resource availability and anthropogenic pressures such as hunting (Thurfjell et al. 2013; Morelle and Lejeune 2015). These seasonal variations complicate static modeling approaches that rely on outdated or coarse-resolution environmental predictors, risking misrepresentation of actual habitat use and conflict hotspots (Culshaw et al. 2021). Moreover, most existing wild boar SDMs are based on static land cover (LC) maps, such as Corine Land Cover, that fail to reflect dynamic landscape changes driven by crop rotations, fruitification events, climatic extremes, and human disturbances (Pettorelli et al. 2005).\u003c/p\u003e\n\u003cp\u003eRecent advances in SDMs and remote sensing (RS) offer pathways toward a more robust, multitemporal understanding of habitat suitability (Wang et al. 2025). Dynamic SDMs, which explicitly incorporate temporally resolved satellite-derived environmental predictors, significantly enhance ecological realism (He et al. 2015). The lack of seasonally explicit SDMs for wild boar remains a gap in Europe, limiting precise, targeted management interventions in agriculture, transportation, disease management, and conservation (Enetwild Consortium et al. 2024). France provides an ideal case study due to its diverse landscapes, growing wild boar populations involved in frequent HWCs, and extensive national-scale monitoring data, yet it currently lacks a consistent, seasonal SDM framework. Additionally, wild boar have been observed to increase their home range during winter over multiple French climate zones (Calenge et al. 2002). Especially the rapidly rising wild boar incidents on the French railway network (Soci\u0026eacute;t\u0026eacute; nationale des chemins de fer fran\u0026ccedil;ais, SNCF) motivate a national-level investigation (Santos et al. 2017).\u003c/p\u003e\n\u003cp\u003eIn this study, we address this critical gap by developing an innovative multitemporal SDM pipeline that integrates modern RS approaches, advanced machine learning (ML), and carefully curated occurrence data for continental France. Specifically, we leverage novel LC classification derived from a state-of-the-art Vision Transformer (ViT) deep learning architecture fine-tuned on high-resolution multispectral satellite imagery (Jakubik et al. 2023). By training this model on multitemporal RS stacks, we produce highly detailed annual LC predictors at 30 m resolution that accurately capture subtle phenological and anthropogenic land-use dynamics crucial to wild boar ecology such as crop rotation or forest mast. Multistage satellite time-series data-fusion ML frameworks combining state-of-the-art computer vision with environmental proxy features based on expert annotation have proven particularly effective for such LC paradigms (Meyer et al. 2025).\u003c/p\u003e\n\u003cp\u003eOur methodological framework builds on structured WVC reports from SNCF and roadkill reports by the R\u0026eacute;seau Routier National (RRN) databases alongside opportunistic observations by the Global Biodiversity Information Facility (GBIF). Each presence-only dataset provides distinct ecological insights but carries inherent spatial biases. We explicitly apply spatial bias-correction strategies to prevent SDMs from inadvertently learning observational patterns instead of true ecological relationships (Baker et al. 2024). Importantly, rather than treating these dataset differences as nuisances, we leverage them strategically: WVC datasets uniquely contribute systematic, spatially structured observations reflecting animal movements relative to linear infrastructure, while GBIF enriches ecological breadth and increases geographical coverage.\u003c/p\u003e\n\u003cp\u003eTherefore, we address the following research objectives: (i) Do the seasonally explicit habitat suitability maps produced by our SDM approach provide ecologically meaningful proxies suitable for stakeholder use? (ii) Does multitemporal SDM calibration improve temporal transferability compared with a monotemporal \u0026ldquo;average year\u0026rdquo; baseline? (iii) Does integrating rail and road WVC data enhance the ecological realism of predicted management-relevant interface areas over classical \u0026ldquo;GBIF-only\u0026rdquo; models?\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis work provides a rigorous, data-driven foundation for targeted interventions in transportation planning, agriculture, veterinary epidemiology, and wildlife management. By highlighting the value of integrating multitemporal RS data and structured collision records, our framework demonstrates the feasibility of modern ML modeling pipelines. Moreover, as the developed LC classification approach leverages universally available satellite data, our methodology offers broad global transferability beyond the French context and can inform similar predictive modeling efforts for other conflict-prone ungulates.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eProcess Pipeline Overview\u003c/p\u003e\n\u003cp\u003eOur methodological framework (Fig. 1) involves several integrated analytical stages (detailed in section 2.2-2.7) to rigorously model seasonal habitat suitability for wild boar over the years 2017 to 2023 at a 1\u0026nbsp;km\u003csup\u003e2\u003c/sup\u003e resolution pixel grid covering continental France as our study area. Corsica is excluded due to its insular topography boasting strong variations in wild boar behavior such as heavy hybridization with domestic pigs (Iacolina et al. 2018) and documented endemic subspecies formation (Schleimer et al. 2022) beyond the scope of this study.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBriefly, we (i) assemble three complementary presence-only streams (rail and road collision records plus GBIF observations, Fig. 1 top-left), (ii) reduce clustering and redundant sampling in environmental space to mitigate pseudo-replication (Fig. 1 dark-red pathway), (iii) model source-specific observation effort and use it to weight background sampling (Fig. 1 violet pathway), (iv) construct a predictor stack that combines annually updated LC maps from the satellite-based ViT with seasonal remote-sensing indices, climate, and anthropogenic layers (Fig. 1 central grey and yellow pathways), (v) fit and compare monotemporal versus multitemporal SDMs for two seasons (Fig. 1 bottom), and (vi) evaluate temporal transferability using leave-one-year-out (LOYO) validation and presence-only diagnostics (Fig. 1 bottom-right).\u003c/p\u003e\n\u003cp\u003eThe pipeline produces annual 30 m LC maps explicitly tailored to wild boar ecology capturing critical resources and shelter habitats as an intermediate product (Fig. 1 yellow pathway). Subsequently we derive predictive seasonal habitat-suitability index layers at 1 km\u0026sup2; resolution that remain robust and interpretable across interannual variability, providing an essential resource for operational planning, targeted management and early-warning applications.\u003c/p\u003e\n\u003cp\u003ePresence Data and Thinning\u003c/p\u003e\n\u003cp\u003eWe compile our comprehensive presence-only dataset of wild boar occurrences across France for the years 2017\u0026ndash;2023 from three primary sources combined to maximize coverage of wild boar occurrences in different landscapes (Fig. 2): (i) Railway collision reports provided by SNCF R\u0026eacute;seau (the French rail network infrastructure operator) documenting wild boar struck by trains extracted from the \u0026ldquo;Brehat\u0026rdquo; database (Arthaud et al. 2024), (ii) Roadkill carcass records from the toll-free national road network (RRN) reporting vehicle collisions with wild boar provided by the French Ministry for Ecological Transition (DGITM/DMR/SAM4, Sept 2024) and (iii) combined wild boar observations aggregated through GBIF, which include opportunistic sightings, various published scientific observation collections and citizen-science derived reports by the general public for example through mobile apps (Beck et al. 2012; Enetwild Consortium et al. 2018).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe only include points with verified coordinate positions (error smaller than 0.5 km) and time stamps (temporal resolution at least one month) and filter out all duplicates in the same 1 km\u003csup\u003e2\u003c/sup\u003e cell. Each record is assigned to one of two \u0026ldquo;seasons\u0026rdquo; based on the month of observation, allowing us to model habitat selection separately for \u0026ldquo;summer\u0026rdquo; (Mar\u0026ndash;Aug) and \u0026ldquo;winter\u0026rdquo; (Sep\u0026ndash;Feb). This split is chosen to largely cover the French hunting season within the winter dataset: While the exact dates and special regulations can vary by department according to the federal biodiversity office OFB the hunting period usually opens in early September and ends by the end of February (Calenge et al. 2002; Vajas et al. 2023); stalking hunt can be allowed with a special permit in summer (Maillard et al. 2010).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePresence-only datasets are typically spatially clustered, which can bias model fitting toward densely sampled environments. We therefore applied environmental-space thinning to reduce redundant presences while preserving the breadth of occupied environmental conditions (Aiello-Lammens et al. 2015). We opted for subsampling occurrences using greedy farthest-point sampling (Chamon and Ribeiro 2017) in a decorrelated Mahalanobis-distance representation of the predictor space (de Oliveira et al. 2014) for each given season/year/dataset combination. In practice, we keep ~40% of the original records (Fig. 2) to balance reducing the impact of pseudoreplication (Fourcade et al. 2014) while retaining an ecologically meaningful signal which preserves the fundamental ecological niche of the wild boar as our target species (Jim\u0026eacute;nez and Sober\u0026oacute;n 2022). Algorithmic details are provided in Supplementary Information (SI) S1.\u003c/p\u003e\n\u003cp\u003eBias Correction of Background Points\u003c/p\u003e\n\u003cp\u003eBackground point sampling and selection is a key step in presence-only based SDMs, because it can significantly distort suitability predictions unless sampling effort is explicitly modeled (Lobo et al. 2010; Barbet-Massin et al. 2012). Uneven observer effort and accessibility can lead models to interpret sampling patterns as habitat preference requiring effective spatial correction strategies (Baker et al. 2024). To mitigate such effects, we construct spatially explicit sampling bias surfaces (Fig. 3) for each occurrence source and each year/season to model observation effort as a proxy for realistic sampling intensity (Lobo et al. 2010). This increasingly endorsed practice in presence-only SDMs aligns background point generation with actual detectability (Beck et al. 2014). Syfert et al. (2013) found that incorporating a sampling bias grid significantly enhanced goodness-of-fit metrics, even though it did not completely eliminate bias effects. Meta-analyses and simulations also support modest but meaningful gains from bias correction on independent test data (Baker et al. 2024). Implementation approach including algorithmic details and background information is described in SI S2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eGBIF occurrence data\u003c/strong\u003e are available in large quantities for wild boar in Europe (Croft et al. 2025), but are known to suffer from strong spatial biases such as over-representation near roads, urban centers and accessible sites (Beck et al. 2014; Meyer et al. 2016). These biases may correlate with environmental predictors and lead to misleading niche estimations (Baker et al. 2022; Dubos et al. 2022). To mitigate such effects, recent work recommends modeling sampling effort explicitly (Sicacha-Parada et al. 2019) and implementing density surfaces or accessibility proxies to weight background sampling (Barber et al. 2022). We modeled observer effort using kernel-density estimation of reporting intensity combined with an accessibility proxy (Human Footprint Index) to account for the tendency of observations to cluster near roads, settlements, and other accessible areas (Sanderson et al. 2002; Hughes et al. 2021) (Fig. 3B). For implementation details see SI S2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eWVCs and traffic infrastructure data\u003c/strong\u003e can serve as a highly valuable monitoring prior (Schwartz et al. 2020), with railway operations boasting a particularly high scheduled regularity (Jasińska et al. 2019). Because WVCs are exclusively reported in the immediate vicinity of infrastructure networks we assume that higher traffic volume generally leads to a higher likelihood of detecting collisions (Morelle et al. 2013), although we acknowledge that traffic density is only one of several interacting factors influencing WVCs (Benard 2023). We modeled detection effort as a functional proxy of network extent and traffic intensity, using annual average daily traffic for roads (\u0026ldquo;TMJA\u0026rdquo; dataset obtained from the French Ministry for Ecological Transition, Aug 2025) and annual average daily train frequency for rail sections (proprietary dataset provided by SNCF R\u0026eacute;seau, Mar 2025) (Fig. 3C and D). For implementation details see SI S2.\u003c/p\u003e\n\u003cp\u003eThe three bias layers were normalized and combined into a composite bias surface per season (Fig. 3A), which then defined the probability of drawing background points. This composite weighting retains the footprint of each observation stream while reducing the risk of spurious correlations driven purely by accessibility or infrastructure placement. Mathematical definitions including blending parameters and interpolation of missing traffic counts, as well as a full list of bias maps are reported in SI S2.\u003c/p\u003e\n\u003cp\u003eLC Predictors and ViT\u003c/p\u003e\n\u003cp\u003eA key innovation of this study is the use of annually updated high-resolution land-cover (LC) predictors designed around wild boar resources and shelter (Fig. 4, left three columns). To represent the dynamic shift of LC more faithfully than standardized static products, we produced annual LC maps at 30 m resolution for 2017\u0026ndash;2023 using a multispectral ViT semantic-segmentation foundation model \u0026ldquo;Prithvi-100M\u0026rdquo;\u0026nbsp;(Jakubik et al. 2023)\u0026nbsp;fine-tuned on harmonized Landsat\u0026ndash;Sentinel-2 surface reflectance imagery (HLS). HLS is provided by NASA and combines the four satellite missions Landsat-8/9 and Sentinel-2A/B into a common atmospherically and radiometrically corrected virtual constellation\u0026nbsp;(Claverie et al. 2018).\u003c/p\u003e\n\u003cp\u003eTraining labels were compiled from multiple high-quality national and international GIS resources to provide robust training labels. The 12 chosen classes (see SI S3a) directly correspond to wild boar habitat and foraging opportunities, given their opportunistic omnivorous diet primarily consisting of vegetation and agricultural crops (Brandt et al. 2006; Ballari and Barrios-Garc\u0026iacute;a 2014). Wild boar frequently exploit energy-rich cultivated crops when available, shifting to natural resources such as acorns, chestnuts, roots, and tubers in forests or outside the cropping season. Brandt et al. (2006) highlight the significant dietary contributions of tree mast (oak and beech nuts) and cultivated crops to the French wild boar diet.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAnnotations were sourced for crop types from the Registre Parcellaire Graphique (RPG) (Cantelaube and Carles 2014), forest types from the French national forest inventory (IGN BD For\u0026ecirc;t v2) augmented with annual forest-loss updates (Hansen et al. 2013), and additional land-use categories from Theia LC products (Inglada et al. 2017; Puissant et al. 2019), complemented by Corine Land Cover and OpenStreetMap for built-up and water classes. We trained with an auxiliary \u0026lsquo;Other/NoData\u0026rsquo; class for segmentation stability, which was excluded from the subsequent workflow. Details on class composition can be found in SI S3a.\u003c/p\u003e\n\u003cp\u003eFor SDM integration, the 30 m LC maps were aggregated to the 1 km\u0026sup2; modeling grid as (i) fractional cover per class and (ii) distance-to-class layers capturing edge and proximity effects, which are known to be important for wild boar use of forest-agriculture interfaces (Thurfjell et al. 2009). Model training details, hyperparameters, hardware requirements and inference configuration are provided in SI S3b.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEnvironmental Predictor Stack\u003c/p\u003e\n\u003cp\u003eBeyond LC composition, we included dynamic and static predictors capturing vegetation condition, climate, terrain, fragmentation, and anthropogenic pressures (Fig. 4). RS time series can substantially improve SDMs by representing within- and between-year variation in habitat conditions (Pettorelli et al. 2005; He et al. 2015). We derived spectral indices from HLS mosaics at key phenological stages (April, July, October) and assigned them seasonally (spring/summer indices to summer models; autumn indices to winter models). To reduce redundancy and multicollinearity, predictors were pruned using variance inflation factor (VIF) filtering following established practice in wild boar SDMs (Bosch et al. 2014).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eClimate predictors combined dynamic seasonal anomalies from ERA5-Land reanalysis (temperature, precipitation, soil moisture, solar radiation; plus snow and minima for winter) (Mu\u0026ntilde;oz-Sabater et al. 2021) with higher-resolution long-term bioclimatic normals from WorldClim to represent microclimatic gradients relevant to wild boar persistence and movement (Enetwild Consortium et al. 2019; Vetter et al. 2020). Terrain variables (elevation, slope, aspect) were derived from the \u0026ldquo;SRTM DEM\u0026rdquo; dataset (obtained from USGS, see SI S4) to account for topographic structure and known wild-boar specific constraints (Acevedo et al. 2006).\u003c/p\u003e\n\u003cp\u003eWe further included landscape fragmentation resampled from the \u0026ldquo;effective mesh density\u0026rdquo; dataset (EEA/FOEN 2011) and the \u0026ldquo;Copernicus small woody features\u0026rdquo; dataset (Faucqueur et al. 2019) as proxies for permeability, shelter and corridor structure, which can shape wild boar movement and conflict risk (Welander 2000; Ficetola et al. 2014). Wetlands were included from the \u0026ldquo;INPN Zones humides\u0026rdquo; dataset (Rapinel et al. 2023) as seasonal resources and refuges (Barasona et al. 2021). Finally, we integrated coarse regional hunting-bag statistics (departmental counts by OFB, Aug 2025) as a spatial proxy for hunting pressure and associated behavioral shifts, activated only in winter models (Calenge et al. 2002; Keuling et al. 2008; Thurfjell et al. 2013). All predictors were harmonized to the 1 km\u0026sup2; grid. Data provenance, processing details and correlation matrices as basis for collinearity feature pruning are reported in SI S4.\u003c/p\u003e\n\u003cp\u003eSDM Modeling Strategy\u003c/p\u003e\n\u003cp\u003eWe modeled habitat suitability separately for summer and winter using two complementary SDM engines: MaxEnt and gradient-boosted decision trees (GBM). MaxEnt is a widely used presence-only approach with strong theoretical grounding and a long record of robust performance when sampling bias is addressed and model complexity is controlled (Phillips et al. 2006; Elith et al. 2011). GBMs can capture nonlinearities and interactions and provide a flexible benchmark against a regularized, more interpretable model class. We implemented MaxEnt (Phillips et al., 2006) through the \u003cem\u003eelapid\u003c/em\u003e Python 3 interface (Anderson 2023), while the kernel is based on \u003cem\u003eglmnet\u0026nbsp;\u003c/em\u003eensuring compatibility and comparability of models and results to the popular open-source R release \u003cem\u003emaxnet\u003c/em\u003e (Phillips et al. 2017). We implemented GBM in Python 3 using the \u003cem\u003eLightGBM\u003c/em\u003e library (Ke et al. 2017).\u003c/p\u003e\n\u003cp\u003eMultitemporal calibration of SDMs has often been suggested as a pathway for overcoming existing SDM limitations (Reside et al. 2010; Mart\u0026iacute;nez-Minaya et al. 2018; Eduardo et al. 2022). To test whether exposing the learner to interannual environmental variability improves robustness, we compared (i) a multitemporal calibration in which training data retain year-specific predictors, and (ii) a monotemporal baseline in which predictors are averaged across training years into an \u0026ldquo;average year\u0026rdquo; representation comparable to the monotemporal approach of Bosch et al. (2014) or Guisan \u0026amp; Thuiller (2005).\u003c/p\u003e\n\u003cp\u003eWe evaluated temporal transferability using a leave-one-year-out (LOYO) design: models were trained on all but one year and then projected to the held-out year using its corresponding predictor stack. For each LOYO fold, we sampled 15,000 background points per year from the composite bias surface and fitted models to presences and weighted background points. Hyperparameters were tuned using automated search (Optuna) within the training data, with spatial blocking used during tuning to reduce optimistic bias from spatial autocorrelation (Roberts et al. 2017; Valavi et al. 2019). Full search spaces, convergence settings, and calibration details are reported in SI S5.\u003c/p\u003e\n\u003cp\u003eModel outputs are transformed to a continuous habitat-suitability index (0\u0026ndash;1) using the cloglog transformation commonly used for MaxEnt-style presence-only predictions, facilitating interpretation as relative suitability as an approximate probability of presence\u0026nbsp;(Phillips and Dud\u0026iacute;k 2008).\u003c/p\u003e\n\u003cp\u003eEvaluation and Uncertainty Quantification\u003c/p\u003e\n\u003cp\u003eFor each year/season combination run we calculate the \u003cem\u003eContinuous Boyce Index\u003c/em\u003e (CBI, Spearman variant) (Boyce et al. 2002; Hirzel et al. 2006) as our main presence-only diagnostic quantifying how much the predicted suitability values deviate from a random expectation, given the distribution of presence points across prediction bins. We complementarily calculate model discrimination capability expressed as the \u003cem\u003eArea under the Receiver Operating Characteristic Curve\u003c/em\u003e (AUC). Because the wild boar is a widespread generalist with few clearly unsuitable areas, rank-based presence-only metrics are often more informative than discrimination metrics relying on pseudo-absences (see chapter Discussion: SDM Metrics). We report AUC for comparability with earlier SDM literature and for computing permutation-based feature importance. We additionally calculate variable response curves by systematically varying one predictor at a time while keeping other predictors constant at their mean values (SI S6).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe quantify environmental novelty and projection extrapolation capability of our models with two complementary indices: (i) For novelty type 1 (NT1) \u0026ndash; values outside the univariate training ranges \u0026ndash; we compute the Multivariate Environmental Similarity Surface (MESS) (Elith et al. 2010), which evaluates each pixel against the full distribution of training values per predictor and reports the minimum similarity across variables. (ii) For novelty type 2 (NT2) \u0026ndash; novel combinations of predictors even when each falls within its marginal range \u0026ndash; we compute ExDet-NT2 (Mesgaran et al. 2014) as a Mahalanobis distance-based measure relative to the training mean covariance structure (scaled by the maximum training distance). Thresholds are specified in Fig. 10.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eVision Transformer\u003c/p\u003e\n\u003cp\u003eFine-tuning the ViT on multispectral satellite imagery successfully predicted high-resolution (30 m) annual LC maps for France, distinguishing 12 ecologically relevant habitat classes with strong metrics (Table 1): Mean Intersection-over-Union (IoU) = 47.7 %, mean pixelwise class accuracy (mAcc) = 71.8 %. Achieved values slightly surpassed the multitemporal crop segmentation achieved in the original reference (Jakubik et al. 2023). Detailed results, confusion matrices and an evaluation against 1k manually annotated points are available in SI S3c.\u003c/p\u003e\n\u003cp\u003eThe resulting LC maps demonstrate a clear and detailed delineation of regional landscape patterns and class-specific distributions (Fig. 5). On a national scale, broad spatial gradients are captured accurately (Fig. 5 top right). The zoomed-in section (Fig. 5 bottom right) illustrates precise classification of fine-scale features, such as river corridors and crop-field mosaics. Overall, the annual LC maps effectively capture the interannual and seasonal variability in vegetation dynamics and crop rotations, supporting nuanced habitat assessments at scale. Producing such maps requires significant computational resources (~50 GPU-hours per annual inference run), indicating a clear trade-off between spatial detail and operational efficiency. Nonetheless, the achieved classification quality and ecological relevance justify the computational cost, providing robust inputs for subsequent wild boar habitat suitability modeling.\u003c/p\u003e\n\u003cp\u003ePer-class performance varied (Table 1), with notably high IoU for water (83.4 %), wheat (61.6 %), rapeseed (57.9 %), urban areas (53.1 %) and coniferous forest (47.3 %). Conversely, mixed and beech forests showed lower IoU scores (26.9 % and 21.3 %, respectively), indicating intra-domain variability (evident when assessing the confusion matrix, see SI S3c). Accuracy was consistently high (\u0026gt;78 %) for water, rapeseed, corn/maize, and coniferous forests, reflecting strong model differentiation potential.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 1 - Classification metrics after fine-tuning the ViT for 120 Epochs.\u003c/p\u003e\n\u003ctable border=\"0\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eClass\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eIoU [%]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy [%]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eWheat\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e61.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e78.32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCorn / Maize\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e55.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e84.11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eBarley\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e38.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e72.85\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eRapeseed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e57.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e85.24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eGeneral Agriculture \u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e32.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e38.25\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eGrassland\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e52.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e66.71\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eOak Grove\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e34.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e70.21\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eConiferous Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e47.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e84.61\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eMixed Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e26.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e29.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eBeech Grove\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e21.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e80.45\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eUrbanization\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e53.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e79.77\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eWater\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e83.37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e91.04\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eTotal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e47.69\u003cbr\u003e\u003c/strong\u003e(mIoU)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e71.79\u0026nbsp;\u003c/strong\u003e(mAcc)\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;60.25\u0026nbsp;\u003c/strong\u003e(aAcc)\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eMultitemporal Model Performance\u003c/p\u003e\n\u003cp\u003eAcross all LOYO folds, MaxEnt models exhibited consistently higher CBI values than GBM models under every configuration (Table 2). MaxEnt\u0026rsquo;s mean CBI (Spearman) was very high for both summer and winter, typically ~0.88\u0026ndash;0.98, whereas GBM\u0026rsquo;s mean CBI ranged lower (~0.56\u0026ndash;0.83). MaxEnt also showed much greater stability across years, with low standard deviation (often 0.02\u0026ndash;0.10) indicating consistent performance on each held-out year. In contrast, GBM\u0026rsquo;s CBI was highly variable (SD up to ~0.5), and in some test years the GBM model\u0026rsquo;s CBI dropped to near zero or even negative. For example, the winter monotemporal GBM had one fold with a negative CBI (\u0026asymp; \u0026ndash;0.35), indicating worse-than-random predictions for that year\u0026rsquo;s data. Notably, MaxEnt never produced a negative CBI in any fold, and even its worst-case test fold remained moderately high (CBI \u0026asymp; 0.67). These results demonstrate that the MaxEnt engine achieved stronger generalization capacity than GBM across temporal folds.\u003c/p\u003e\n\u003cp\u003eTable 2 \u0026ndash; Summary of SDM results on test data from LOYO temporal CV. High values for CBI \u0026gt; 0.8 printed bold.\u003c/p\u003e\n\u003ctable border=\"0\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eData Input\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSeason\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTemporal Setup\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel Engine\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMean CBI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn; \u003cstrong\u003eSD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMax CBI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.943\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.991\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.756\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.418\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e1.000\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.885\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.119\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.984\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.602\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.338\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.977\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.967\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.028\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.989\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.653\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.548\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.985\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.947\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.093\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.993\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eWVCs+GBIF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.560\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.314\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.904\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.941\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.045\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.991\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.694\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.385\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.986\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.914\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.071\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.985\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eSummer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.603\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.393\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.967\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.977\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.016\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.991\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMonotemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.131\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.961\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eMaxEnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.976\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.991\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 82px;\"\u003e\n \u003cp\u003eGBIF_only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003eWinter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 111px;\"\u003e\n \u003cp\u003eMultitemporal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003eGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.831\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 64px;\"\u003e\n \u003cp\u003e\u0026plusmn;\u0026nbsp;0.197\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.981\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eDespite these stark differences in CBI, the Area Under the Curve (AUC) on withheld-year data was acceptably high for both modeling approaches (full fold-wise AUC results are provided in SI S7) MaxEnt models achieved test AUCs typically in the 0.60\u0026ndash;0.72 range for both summer and winter scenarios, while GBM models\u0026rsquo; test AUC ranged roughly 0.62\u0026ndash;0.75, occasionally reaching up to ~0.80 in winter multitemporal runs. In other words, both engines appeared to perform well by the AUC metric on validation data. However, the CBI reveals a crucial difference: GBM\u0026rsquo;s high AUC did not translate into ecologically reliable predictions. GBM models often attained very high training AUC (\u0026asymp; 0.93), indicating a near-perfect fit to the training presences, yet their validation CBI was poor. This pattern reflects overfitting in the GBM models: they fit the training data extremely well but also failed to rank habitats correctly for unseen test data years. MaxEnt, by contrast, maintained only moderate training AUC (\u0026asymp; 0.70), suggesting stronger regularization, but it retained high CBI on validation folds. Thus, MaxEnt models generalized better, yielding higher validation CBI despite similar (or slightly lower) AUC than GBM \u0026ndash; a sign that MaxEnt predictions were more ecologically consistent.\u003c/p\u003e\n\u003cp\u003eIn terms of seasonal and temporal settings, monotemporal vs. multitemporal models performed similarly overall. Any differences due to temporal setup were small compared to the large engine effect. MaxEnt\u0026rsquo;s CBI remained high in both monotemporal and multitemporal modes (e.g. winter MaxEnt CBI \u0026asymp;0.95\u0026ndash;0.98 in both setups), and GBM\u0026rsquo;s performance, while lower, followed the same pattern (its best CBI occurred in monotemporal summer runs, but even there mean CBI was only ~0.75). Data input (WVCs+GBIF vs. GBIF_only) had minimal impact on these rank-based metrics. Models trained on the full presence dataset versus only GBIF records yielded very similar CBI outcomes (Table 2). This suggests that adding the extra WVC occurrence points (from sources beyond GBIF) did not drastically alter the models\u0026rsquo; overall skill in predicting habitat suitability rankings. For MaxEnt in particular, CBI stayed consistently high regardless of data source.\u003c/p\u003e\n\u003cp\u003eFeature Importance\u003c/p\u003e\n\u003cp\u003eBecause of the consistently higher CBI of MaxEnt over GBM models, permutation feature importance was only calculated for MaxEnt. Across seasons, the ten most important predictors consistently grouped into three clear ecological categories (Fig. 6 and 7): (i) proximity to mature broadleaf forest (particularly beech and oak), (ii) LC structure (fractional grassland and general agriculture; crop fractions and distances), and (iii) weather/energy constraints from ERA-5 (winter solar radiation; summer air temperature and shallow soil moisture/temperature).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eWinter\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eProximity to mature beech forests consistently emerged as the strongest predictor (approx. 15% across models). Initially, monotemporal models also strongly emphasized fractional grassland (approx. 11%) and general agriculture (approx. 11%). However, their importance decreased substantially under multitemporal calibration (grassland: \u0026Delta; \u0026asymp; -4%; general agriculture: \u0026Delta; \u0026asymp; -3%), indicating that their predictive value varies considerably across years. Small woody features may serve as a more relevant refuge in winter (up to 10%) than in summer (approx. 4%).\u003c/p\u003e\n\u003cp\u003eIn contrast, dynamic climatic predictors gained importance under multitemporal modeling. ERA-5 surface solar radiation, reflecting dynamic seasonal weather conditions, increased dramatically from around 4% in monotemporal models to 16% in multitemporal models (\u0026Delta; \u0026asymp; +12%), becoming a critical predictor of interannual variation in winter habitat suitability.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMoreover, incorporating structured WVC data notably reduced importance of features potentially influenced by observer bias: small woody features decreased from around 10% (GBIF_only monotemporal) to about 5% (WVC+GBIF multitemporal; \u0026Delta; \u0026asymp; -5%), and dense urbanization (approx. 3%) dropped out entirely. Overall, multitemporal calibration combined with WVC data increased the cumulative predictive power of the top features from 70% to 73% (\u0026Delta; \u0026asymp; +3%), underscoring improved focus on ecologically robust variables (Fig. 6).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSummer\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDuring summer, multitemporal calibration emphasized dynamic climatic constraints even more distinctly. ERA-5 mean summer temperature rose markedly from around 6% importance in monotemporal models to about 15% in multitemporal models (\u0026Delta; \u0026asymp; +9%), highlighting the influence of dynamic climatic variability on habitat selection. Fractional grassland decreased in importance under multitemporal models (from around 12% monotemporal to around 6% multitemporal; \u0026Delta; \u0026asymp; -6%), suggesting high year-to-year variability in grassland suitability.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u0026nbsp;Furthermore, a notable ecological shift occurred with multitemporal modeling: the importance of proximity to maize (approx. 8% monotemporal) was replaced by proximity to oak stands (approx. 8% multitemporal; \u0026Delta; \u0026asymp; +8%), suggesting that once interannual variability is integrated, proximity to mast-producing broadleaf stands may supersede crop adjacency as a predictor. Proximity to beech similarly from about 8% monotemporal to around 4% multitemporal (\u0026Delta; \u0026asymp; -4%), indicating further seasonal differences in resource reliance. After mast year events oak stands might offer extended availability of acorns into the following spring and summer compared to beech nuts which spoil more rapidly (Brandt et al. 2006).\u003c/p\u003e\n\u003cp\u003eWheat forms a stable summer resource (approx. 6-8%). Small woody features and natural heath and altitude provide modest, steady contributions (approx. 3-5% each). Overall, cumulative predictive importance slightly decreased from monotemporal (approx. 69%) to multitemporal (approx. 61%, \u0026Delta; \u0026asymp; ‑8%), reflecting a more balanced distribution of predictor importance, clearly distinguishing stable landscape resources from dynamically varying climatic variables.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eEcological Inference\u003c/p\u003e\n\u003cp\u003eVisual inspection of the seasonal predictions (Fig. 8) reveals concordant hotspots in both summer and winter: Provence-C\u0026ocirc;te d\u0026rsquo;Azur, the Pyrenees, the forests of \u0026Icirc;le-de-France, and the Alsace/Vosges chain. Large riparian corridors (e.g. Loire and Rh\u0026ocirc;ne) consistently exhibit elevated suitability. Persistently poor habitats include the high-altitude Alps and the open arable plains of Champagne, which lack shelter and diverse forage.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSeason-specific shifts were evident: The winter maps (Fig. 8 top) uniquely showed increased suitability in the milder Atlantic regions of Bretagne and Normandie, suggesting these areas might serve as important winter refuges. Conversely, the extensive coniferous forests of Aquitaine exhibited depressed suitability in winter compared to summer (Fig. 8 bottom), indicating that some large forest tracts are less utilizable during winter months (possibly due to reduced forage or ongoing disturbances such as hunting).\u003c/p\u003e\n\u003cp\u003eElevated suitability in peri-urban belts persisted even after bias correction, mirroring empirical reports of risk-tolerant behavior and stable urban or suburban boar groups in European cities (Stillfried et al. 2017; Marin et al. 2024). Dense centers remained less suitable than surrounding suburban greenbelt mosaics. Specifically, our models suggest elevated suitability hotspots in the wider suburban areas of the South-West (Marseille, Toulon, Nice, Montpellier, Narbonne, Perpignan) as well as Paris, Bordeaux, Toulouse and Nantes during both summer and winter. In the Upper Rhine valley (Strasbourg, Mulhouse) peri-urban suitability is specifically strong in the winter months.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe note that generally most of continental France has been consistently mapped as an at least moderately suitable habitat for wild boar. A stronger concentration of hotspots during winter can be attributed to coastal zones, heterogenous lowlands with a moderate to warm climate, as well as more urbanized areas. Summer signal spreads out more evenly favoring some additional forest-agriculture interfaces reducing the least suitable areas too very singular stretches.\u003cbr\u003e\u0026nbsp;\u003cbr\u003eThese interpretations are consistent with known wild boar ecology: preferred use of forest-agriculture ecotones, mast-producing broadleaf stands, riparian corridors, and an increased reliance on peri-urban resources (Schley and Roper 2003; Brandt et al. 2006; Thurfjell et al. 2009; Stillfried et al. 2017; Marin et al. 2024).\u003c/p\u003e\n\u003cp\u003eSDM Metrics\u003c/p\u003e\n\u003cp\u003eAcross all LOYO folds, MaxEnt produced higher and markedly more stable CBI than GBM, and its suitability maps were spatially coherent and ecologically interpretable (Fig. 8). This pattern is expected for presence-only data when models are strongly regularized in the feature space and evaluated with a rank-based metric (Phillips and Dud\u0026iacute;k 2008; Elith et al. 2011). By contrast, boosted trees can achieve very high in-sample discrimination yet exhibit poor temporal transfer if capacity is not tightly constrained relative to sample size and predictor cardinality. In our case, GBM frequently reached training AUCs \u0026gt; 0.9 but yielded mediocre or unstable CBIs on withheld years, whereas MaxEnt retained moderate training AUCs (~0.7) and high validation CBI values.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis divergence illustrates why the CBI is the primary criterion for presence-only SDMs: CBI asks whether observed presences accumulate in high-suitability bins more than expected by chance and does not rely on a comparison against pseudo-absences (Hirzel et al. 2006; Di Cola et al. 2017). We therefore adhere to recommendations preferring the CBI for model comparison and ecological interpretation because the CBI does not require systematically collected absence data and is comparatively insensitive to prevalence and background sampling definition (Lobo et al. 2008; Jim\u0026eacute;nez-Valverde 2012; Jim\u0026eacute;nez and Sober\u0026oacute;n 2020).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe AUC, although still widely reported, remains sensitive to the geographic extent of the background, species prevalence, and pseudo-absence design, and it can inflate apparent skill especially for widespread generalists or penalize ecologically sensible maps\u0026nbsp;(Lobo et al. 2008; Jim\u0026eacute;nez-Valverde 2012; Warren et al. 2020). Additionally, generalist species like wild boar, by definition, exhibit broader ecological niches, occupying a wide range of environmental conditions, and typically have few clearly unsuitable areas in France. AUC relies on sharply contrasting presence-background separation and tends to be less sensitive in such a context\u0026nbsp;(Warren et al. 2020). CBI instead quantifies how well the model\u0026rsquo;s predicted suitability corresponds to actual species use patterns across continuous gradients of suitability, providing a more nuanced and ecologically meaningful assessment of predictive accuracy\u0026nbsp;(Paudel et al. 2015; Milano et al. 2024).\u003c/p\u003e\n\u003cp\u003eThe added benefit of WVC datasets\u003c/p\u003e\n\u003cp\u003eRail and road collision records supplied systematic, high-resolution presences along linear infrastructure \u0026ndash; an observation process that complements the observer-biased reach of GBIF (Seiler 2004; Langbein et al. 2011; Jasińska et al. 2019; Schwartz et al. 2020).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe winter 2019 difference map (Fig. 9) illustrates the practical effect of the integration of WVC datasets: Generally, consensus between WVCs+GBIF and GBIF_only models is high with \u0026Delta;HSI between -0.2 and 0.2 for over 96.5% of the study area (mean 0.007 \u0026plusmn; 0.094 SD).\u003c/p\u003e\n\u003cp\u003eWhen WVCs were included, predicted suitability increases within the conifer-dominated forests of Aquitaine and in mountainous regions like the Pyrenees, the Massif Central, Provence-C\u0026ocirc;te d\u0026rsquo;Azur, Languedoc and the pre-Alps (Chartreuse, Savoies, Jura). Inflated GBIF signals in the West (Bretagne, Pays-de-la-Loire, Limousin-Indre) were down-regulated, most notably in a heavily sampled zone northwest of Nantes. This is also reflected by the disappearance of the \u0026ldquo;dense urbanization\u0026rdquo; proxy from Top10 feature importance, once WVCs are included.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWVCs emphasize habitats where wild boar actually move across landscapes, capturing dispersal and risk corridors (e.g., beech and oak proximity, fragmentation, grassland-agriculture mosaics). Consequently, forest-edge and movement-related predictors (distance to beech, ERA-5 energy variables) gain relative weight, while purely observational correlates (small woody features, urban bias) diminish. These corrections are ecologically coherent: in winter, boar movements and resource tracking bring animals into contact with transport corridors and montane forest refugia, which collision data detects but GBIF_only underrepresents. WVCs also yielded a sharper more concentrated feature signal in both seasons, indicating a model that explains variance through a smaller, more coherent and interpretable predictor subset. Conversely, GBIF potentially overemphasizes easily accessible zones (Beck et al. 2014; Hughes et al. 2021). While both WVCs+GBIF as well as GBIF_only models yield high scalar metrics, the net effect is not a dramatic performance change but a material improvement in the spatial realism of predicted conflict corridors and seasonal refugia (Morelle et al. 2013, 2014).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMultitemporal Modeling Advances Extrapolation Capability\u003c/p\u003e\n\u003cp\u003eBecause of consistently more robust performance of MaxEnt models over GBM we restricted the environmental novelty analysis to MaxEnt, calculating flagged fractional pixel count of the total AoI: NT1 if MESS \u0026lt; -10, representing pixels where at least one predictor exceeded its training range (univariate extrapolation), and NT2 if ExDet \u0026gt; 1, indicating novel predictor combinations outside training ranges (Fig. 10 top).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOverall, multitemporal models consistently reduced novelty (NT1) compared to monotemporal approaches, especially under typical climatic conditions. During the winters 2017\u0026ndash;2022 monotemporal models generally showed moderate NT1 novelty, ranging from 1\u0026ndash;6 %, but spiked sharply to 10% in the anomalously warm winter of 2023. In contrast, multitemporal winter models remained consistently low in novelty (\u0026lt;0.2 %), rising only modestly to 2% in 2023.\u003c/p\u003e\n\u003cp\u003eSummer scenarios exhibited more pronounced variability. Monotemporal NT1 novelty remained relatively low (~3%) from 2018 to 2020, but surged dramatically to 32% during the anomalous summer of 2021. It subsequently decreased to 5% in 2022, stabilizing around 9% in 2023. Multitemporal summer models maintained substantially lower novelty overall (baseline 0.02\u0026ndash;0.04%), but still showed brief increases during anomalies (5% in 2021, 1% in 2022) before returning to baseline levels (0.07 %) in 2023. NT2 novelty, representing uncertain predictor combinations, remained consistently low (0.4\u0026ndash;1 %) across all scenarios and training schemes.\u003c/p\u003e\n\u003cp\u003eDistinct spatial patterns of extrapolation further support these findings (Fig. 10, bottom panel):\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003eIn the anomalous summer of 2021, severe flooding in northeastern France triggered high NT1 novelty in monotemporal models, notably along the Moselle and Meuse valleys. In contrast, multitemporal models showed only isolated novelty patches.\u003c/li\u003e\n \u003cli\u003eDuring winter 2023/24, the warmest global winter on record, monotemporal models again indicated widespread NT1 novelty across northern France due to extreme temperatures. Conversely, multitemporal models exhibited minimal extrapolation, underscoring their robustness in capturing broader climatic variability.\u003c/li\u003e\n \u003cli\u003eNT2 novelty (novel predictor combinations) was consistently limited to small, localized clusters, primarily in mountainous areas and densely populated city centers, and remained nearly identical across all seasons and modeling approaches.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eComparing predictive performance by CBI, differences between mono-\u0026nbsp;and multitemporal models were modest on average (Table 2), yet their consequences were substantial for robustness. Novelty diagnostics showed that multitemporal calibration consistently curtailed true extrapolation both spatially and quantitatively. These findings support the argument that exposing the learner to multi-year variability widens the realized environmental space, improves temporal transfer, and increases resilience to non-stationarity\u0026nbsp;(Franklin 2010; Mart\u0026iacute;nez-Minaya et al. 2018; Eduardo et al. 2022). We therefore argue that multitemporal calibration of generalist species SDMs is becoming an increasingly necessary property under future rapid climate variability scenarios.\u003c/p\u003e\n\u003cp\u003eLimitations and Outlook\u003c/p\u003e\n\u003cp\u003eDespite robust methodological approaches, several important limitations warrant consideration. Our models currently estimate habitat suitability, but we do not directly translate our results into abundance or density metrics. To calibrate HSI into actual population sizes future modeling approaches can bridge this gap by integrating occupancy-detection frameworks (MacKenzie et al. 2017), distance sampling (Buckland et al. 2015), or spatial capture-recapture methods (Royle et al. 2013). For instance, camera trapping setups or GPS collar telemetry combined with capture-recapture approaches provide the necessary spatially explicit capture histories, enabling the conversion of habitat suitability into population density estimates at fine spatial scales. While we are conducting a subsequent camera trap study alongside infrastructure corridors, we argue that the rapid reproductive capacity and high ecological plasticity of wild boar typically result in swift occupancy of suitable niches. Consequently, our HSI can be considered directly informative and operationally useful for guiding proactive management decisions, even in the absence of direct density estimates.\u003c/p\u003e\n\u003cp\u003ePresence-only inference remains sensitive to residual bias and spatial dependence even under weighting; CBI as an evaluation method as well as bias correction techniques mitigate but cannot completely eliminate these issues (Syfert et al. 2013; Dubos et al. 2022; Baker et al. 2024). The benefits of such corrections are context-dependent: They can reduce redundancy and pseudo-replication, but point thinning always discards information and does not necessarily guarantee better transferability or faithful response curves (Fourcade et al. 2014; Aiello-Lammens et al. 2015; Ten Caten and Dallas 2023). A higher spatial resolution might increase environmental coverage at the potential cost of loss of generalizability. The spatial resolution of the current predictor stack, aggregated to a 1 km\u0026sup2; grid, may also overlook finer-grained ecological processes or landscape elements critical to wild boar resource use, such as small habitat patches, agricultural edges, or localized human disturbances. Hence, leveraging finer-resolution predictors or hierarchical multi-scale modeling approaches (Johnson et al., 2004; DeCesare et al., 2012) could further enhance ecological realism. For example, models may not fully account for rapidly changing anthropogenic factors such as intensified hunting pressures, agricultural shifts, rapid urban expansion, or sudden infrastructure developments, all of which can dynamically alter habitat suitability beyond the captured predictor variability. Our extrapolation novelty analysis shows that integrating multitemporal environmental scenarios can in principle capture a wider range of such potential scenarios and therefore can offer a suitable pathway for navigating future environmental or anthropogenic extremes.\u003c/p\u003e\n\u003cp\u003eAdditionally, integrating Artificial Neural Networks (ANNs), particularly Convolutional Neural Networks (CNNs) or ViTs, as an SDM engine represents a natural and beneficial progression in our framework. Neural networks excel in capturing complex nonlinear relationships, interactions, and spatiotemporal patterns that conventional SDMs might overlook. Early fusion approaches (such as integrating LC maps and satellite-derived phenological indicators at high resolution directly into a fully trainable ANN SDM) could maximize the information content and predictive accuracy. These networks might efficiently leverage multi-scale spatial information, analyzing both the immediate and contextual landscape configurations around occurrence points, offering substantial ecological interpretability improvements. However, robust ANN implementations potentially require even larger and richer datasets. As WVCs involving wild boar are steadily rising and reporting protocols by SNCF and other agencies are improving, we anticipate that future data will enhance spatiotemporal resolution and comprehensiveness, strengthening the predictive capability and applicability of such models.\u003c/p\u003e\n\u003cp\u003eApplying our seasonal suitability layers makes it possible to time and place interventions with greater precision: (i) Transport agencies can prioritize winter-risk segments for fencing, signage and crossing structures where models reveal recurrent WVC corridors (Seiler 2004; Langbein et al. 2011). Agricultural services can anticipate summer hotspots around energy-rich crops and deploy deterrents or target hunting accordingly (Ficetola et al. 2014). (ii) Veterinary authorities can focus surveillance and biosecurity along predicted movement corridors and refugia relevant to ASF spread (Morelle et al. 2020; Sauter-Louis et al. 2021). (iii) Urban planners should explicitly consider the suburban greenbelts as functional habitat and manage attractants, connectivity, and public communication accordingly (Stillfried et al. 2017; Marin et al. 2024).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBecause the pipeline can digest new satellite data and occurrence streams as soon as they become available, it is generally fast and flexible. On the other hand it also needs to be rerun periodically to operationally track short-term changes and provide early warning under evolving climatic or LU conditions.\u0026nbsp;\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eOur results confirm that large portions of France provide highly suitable habitat for wild boar, underscoring the species' exceptional adaptability and generalist behavior. Although this adaptability poses considerable management challenges \u0026ndash; especially regarding WVCs, agricultural conflicts, and urban encroachment \u0026ndash; the presented modeling workflow offers a robust framework to identify and proactively mitigate conflict hotspots. Technically, the analysis demonstrates that MaxEnt-based multitemporal SDMs substantially outperform monotemporal models in terms of predictive robustness. They offer reduced extrapolation under future extreme climatic conditions and high spatial coherence of predicted wild boar habitat suitability. Incorporating rail and road collision datasets alongside GBIF observations increases data quantity, mitigates spatial biases and enhances realism. Ecologically, the persistent suitability of peri-urban and suburban greenbelts, even after rigorous bias correction, highlights urban habitats as a highly important refuges and resource-rich environments for wild boar populations in France. Although current limitations \u0026ndash; such as spatial precision of collision data, absence of direct abundance estimation, and computational intensity of the ViT pipeline \u0026ndash; persist, our approach offers concrete pathways for spatially and temporally explicit predictions on wild boar occurrence distribution.\u003c/p\u003e \u003cp\u003eCollectively, this study underscores the critical importance of multitemporal data collection and modeling, careful bias correction, and diversified data integration. Our research has broad implications for wildlife conservation, transportation safety, agricultural management, veterinary epidemiology and environmental conflict mitigation. By bridging SDMs, RS, and ML, this study provides a data-driven foundation for evidence-based decision-making. Our research is not only applicable in France but largely builds on open satellite data which is universally available across diverse ecological and infrastructural landscapes worldwide. In the future the methodological insights from this research can be transferred further to other ungulate species that are frequently involved in HWCs and specifically WVCs, such as red deer (\u003cem\u003eCervus elaphus\u003c/em\u003e), roe deer (\u003cem\u003eCapreolus capreolus\u003c/em\u003e), reindeer (\u003cem\u003eRangifer tarandus\u003c/em\u003e), elk (\u003cem\u003eAlces alces\u003c/em\u003e), offering a transferable approach and generalizable pathway to data-driven wildlife management in human-modified landscapes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAuthor contributions\u003c/p\u003e\n\u003cp\u003eA.M. and D.J. conceived the study. A.M. led the methodological development, performed the analyses, and drafted the manuscript. D.J. provided methodological input for modeling and statistics. T.R. supported data preprocessing and model implementation, especially for the Vision Transformer. M.K. contributed to the remote‑sensing components. K.M. contributed to ecological interpretation and management relevance. D.J. provided supervision and project oversight. All authors contributed critically to the drafts and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eData and Code availability\u003c/p\u003e\n\u003cp\u003eGBIF occurrence records used in this study are publicly available from the Global Biodiversity Information Facility (GBIF). Railway collision records (SNCF R\u0026eacute;seau) and road collision records (R\u0026eacute;seau Routier National) were used under data-sharing agreements and are not publicly available; access may be granted by the respective data owners upon reasonable request. HLS satellite datasets are available globally and free of charge from NASA. Pretrained ViT model weights are available through huggingface. Processed predictor layers, trained model weights, and model outputs can be obtained from the corresponding author upon reasonable request. Access to the private code repository is restricted because of the intermediate data products necessary to run the scripts but may be granted from the corresponding author upon reasonable request.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis study was co-funded by SNCF R\u0026eacute;seau (Paris, France) and the Institute Geomatics, FHNW (Muttenz, Switzerland).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003eEthical approval\u003c/p\u003e\n\u003cp\u003eThis study did not involve capture, handling, or experimental procedures on animals.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAcevedo P, Escudero MA, Muńoz R, Gort\u0026aacute;zar C (2006) Factors affecting wild boar abundance across an environmental gradient in Spain. Acta Theriol (Warsz) 51:327\u0026ndash;336. https://doi.org/10.1007/BF03192685\u003c/li\u003e\n\u003cli\u003eAiello-Lammens M, Boria R, Radosavljevic A, et al (2015) spThin: An R package for spatial thinning of species occurrence records for use in ecological niche models. Ecography 38:541\u0026ndash;545. https://doi.org/10.1111/ecog.01132\u003c/li\u003e\n\u003cli\u003eAnderson CB (2023) elapid: Species distribution modeling tools for Python. J Open Source Softw 8:4930. https://doi.org/10.21105/joss.04930\u003c/li\u003e\n\u003cli\u003eArthaud F, Lecoeur G, Pierre A (2024) Transformers \u0026agrave; Grande Vitesse. J Rail Transp Plan Manag 29:100418. https://doi.org/10.1016/j.jrtpm.2023.100418\u003c/li\u003e\n\u003cli\u003eBaker DJ, Maclean IMD, Gaston KJ (2024) Effective strategies for correcting spatial sampling bias in species distribution models without independent test data. Divers Distrib 30:e13802. https://doi.org/10.1111/ddi.13802\u003c/li\u003e\n\u003cli\u003eBaker DJ, Maclean IMD, Goodall M, Gaston KJ (2022) Correlations between spatial sampling biases and environmental niches affect species distribution models. Glob Ecol Biogeogr 31:1038\u0026ndash;1050. https://doi.org/10.1111/geb.13491\u003c/li\u003e\n\u003cli\u003eBallari SA, Barrios-Garc\u0026iacute;a MN (2014) A review of wild boar Sus scrofa diet and factors affecting food selection in native and introduced ranges. Mammal Rev 44:124\u0026ndash;134. https://doi.org/10.1111/mam.12015\u003c/li\u003e\n\u003cli\u003eBarasona JA, Carpio A, Boadella M, et al (2021) Expansion of native wild boar populations is a new threat for semi-arid wetland areas. Ecol Indic 125:107563. https://doi.org/10.1016/j.ecolind.2021.107563\u003c/li\u003e\n\u003cli\u003eBarber RA, Ball SG, Morris RKA, Gilbert F (2022) Target-group backgrounds prove effective at correcting sampling bias in Maxent models. Divers Distrib 28:128\u0026ndash;141. https://doi.org/10.1111/ddi.13442\u003c/li\u003e\n\u003cli\u003eBarbet-Massin M, Jiguet F, Albert CH, Thuiller W (2012) Selecting pseudo-absences for species distribution models: how, where and how many? Methods Ecol Evol 3:327\u0026ndash;338. https://doi.org/10.1111/j.2041-210X.2011.00172.x\u003c/li\u003e\n\u003cli\u003eBeck J, Ballesteros-Mejia L, Buchmann CM, et al (2012) What\u0026rsquo;s on the horizon for macroecology? Ecography 35:673\u0026ndash;683. https://doi.org/10.1111/j.1600-0587.2012.07364.x\u003c/li\u003e\n\u003cli\u003eBeck J, B\u0026ouml;ller M, Erhardt A, Schwanghart W (2014) Spatial bias in the GBIF database and its effect on modeling species\u0026rsquo; geographic distributions. Ecol Inform 19:10\u0026ndash;15. https://doi.org/10.1016/j.ecoinf.2013.11.002\u003c/li\u003e\n\u003cli\u003eBenard A (2023) Road ecology: toward a predictive model of wildlife-vehicle collisions. Univ Claude Bernard Lyon 1\u003c/li\u003e\n\u003cli\u003eBosch J, Mardones F, P\u0026eacute;rez A, et al (2014) A maximum entropy model for predicting wild boar distribution in Spain. Span J Agric Res 12:984\u0026ndash;999. https://doi.org/10.5424/sjar/2014124-5717\u003c/li\u003e\n\u003cli\u003eBoyce MS, Vernier PR, Nielsen SE, Schmiegelow FKA (2002) Evaluating resource selection functions. Ecol Model 157:281\u0026ndash;300. https://doi.org/10.1016/S0304-3800(02)00200-4\u003c/li\u003e\n\u003cli\u003eBrandt S, Baubet E, Vassant J, Servanty S (2006) R\u0026eacute;gime alimentaire du sanglier en milieu forestier de plaine agricole. Faune Sauvage 273:20\u0026ndash;27\u003c/li\u003e\n\u003cli\u003eBuckland ST, Rexstad EA, Marques TA, Oedekoven CS (2015) Distance sampling: methods and applications. Springer\u003c/li\u003e\n\u003cli\u003eCalenge C, Maillard D, Vassant J, Brandt S (2002) Summer and hunting season home ranges of wild boar (Sus scrofa) in two habitats in France. Game Wildl Sci 19 (4):281\u0026ndash;301\u003c/li\u003e\n\u003cli\u003eCantelaube P, Carles M (2014) Le registre parcellaire graphique : des donn\u0026eacute;es g\u0026eacute;ographiques pour d\u0026eacute;crire la couverture du sol agricole. NOVAE - Ing\u0026eacute;nierie Savoir-Faire Innovants 58\u0026ndash;64\u003c/li\u003e\n\u003cli\u003eChamon LF, Ribeiro A (2017) Greedy sampling of graph signals. IEEE Trans Signal Process 66:34\u0026ndash;47\u003c/li\u003e\n\u003cli\u003eClaverie M, Ju J, Masek JG, et al (2018) The Harmonized Landsat and Sentinel-2 surface reflectance data set. Remote Sens Environ 219:145\u0026ndash;161. https://doi.org/10.1016/j.rse.2018.09.002\u003c/li\u003e\n\u003cli\u003eCroft S, Warren D, Blanco-Aguiar J, Smith G (2025) Predicting the distribution of common wild mammal species across Europe - are there sufficient occurrence data? Eur J Wildl Res 71:133. https://doi.org/10.1007/s10344-025-02014-2\u003c/li\u003e\n\u003cli\u003eCulshaw V, Mairal M, Sanmart\u0026iacute;n I (2021) Biogeography Meets Niche Modeling: Inferring the Role of Deep Time Climate Change When Data Is Limited. Front Ecol Evol 9:. https://doi.org/10.3389/fevo.2021.662092\u003c/li\u003e\n\u003cli\u003ede Oliveira G, Rangel TF, Lima-Ribeiro MS, et al (2014) Evaluating, partitioning, and mapping the spatial autocorrelation component in ecological niche modeling: a new approach based on environmentally equidistant records. Ecography 37:637\u0026ndash;647. https://doi.org/10.1111/j.1600-0587.2013.00564.x\u003c/li\u003e\n\u003cli\u003eDi Cola V, Broennimann O, Petitpierre B, et al (2017) ecospat: an R package to support spatial analyses and modeling of species niches and distributions. Ecography 774\u0026ndash;787. https://doi.org/10.1111/ecog.02671\u003c/li\u003e\n\u003cli\u003eDubos N, Pr\u0026eacute;au C, Lenormand M, et al (2022) Assessing the effect of sample bias correction in species distribution models. Ecol Indic 145:109487. https://doi.org/10.1016/j.ecolind.2022.109487\u003c/li\u003e\n\u003cli\u003eEduardo AA, Liparini A, Martinez PA, et al (2022) Assessing multitemporal calibration for species distribution models. Ecol Inform 71:101787. https://doi.org/10.1016/j.ecoinf.2022.101787\u003c/li\u003e\n\u003cli\u003eEEA/FOEN (2011) Landscape fragmentation in Europe. European Environment Agency (EEA), Swiss Federal Office for the Environment (FOEN), Copenhagen, Denmark\u003c/li\u003e\n\u003cli\u003eElith J, Kearney M, Phillips S (2010) The art of modelling range-shifting species. Methods Ecol Evol 1:330\u0026ndash;342. https://doi.org/10.1111/j.2041-210X.2010.00036.x\u003c/li\u003e\n\u003cli\u003eElith J, Phillips SJ, Hastie T, et al (2011) A statistical explanation of MaxEnt for ecologists. Divers Distrib 17:43\u0026ndash;57. https://doi.org/10.1111/j.1472-4642.2010.00725.x\u003c/li\u003e\n\u003cli\u003eEnetwild Consortium, Croft S, Blanco-Aguiar JA, et al (2024) Modelling wild boar abundance at high resolution. EFSA Support Publ 21:8965E. https://doi.org/10.2903/sp.efsa.2024.EN-8965\u003c/li\u003e\n\u003cli\u003eEnetwild Consortium, Croft S, Smith G, et al (2018) Wild boar in focus: Review of existing models on spatial distribution and density of wild boar and proposal for next steps. EFSA Support Publ 15:1490E. https://doi.org/10.2903/sp.efsa.2018.EN-1490\u003c/li\u003e\n\u003cli\u003eEnetwild Consortium, Croft S, Smith G, et al (2019) Wild boar in focus: initial model outputs of wild boar distribution based on occurrence data and identification of priority areas for data collection. EFSA Support Publ 16:1533E\u003c/li\u003e\n\u003cli\u003eFaucqueur L, Morin N, Masse A, et al (2019) A new Copernicus high resolution layer at pan-European scale: small woody features. In: Remote Sensing for Agriculture, Ecosystems, and Hydrology XXI. SPIE, pp 268\u0026ndash;278\u003c/li\u003e\n\u003cli\u003eFicetola GF, Bonardi A, Mairota P, et al (2014) Predicting wild boar damages to croplands in a mosaic of agricultural and natural areas. Curr Zool 60:170\u0026ndash;179. https://doi.org/10.1093/czoolo/60.2.170\u003c/li\u003e\n\u003cli\u003eFourcade Y, Engler JO, R\u0026ouml;dder D, Secondi J (2014) Mapping Species Distributions with MAXENT Using a Geographically Biased Sample of Presence Data: A Performance Assessment of Methods for Correcting Sampling Bias. PLOS ONE 9:e97122. https://doi.org/10.1371/journal.pone.0097122\u003c/li\u003e\n\u003cli\u003eFranklin J (2010) Moving beyond static species distribution models in support of conservation biogeography. Divers Distrib 16:321\u0026ndash;330. https://doi.org/10.1111/j.1472-4642.2010.00641.x\u003c/li\u003e\n\u003cli\u003eGuisan A, Thuiller W (2005) Predicting species distribution: offering more than simple habitat models. Ecol Lett 8:993\u0026ndash;1009. https://doi.org/10.1111/j.1461-0248.2005.00792.x\u003c/li\u003e\n\u003cli\u003eHansen MC, Potapov PV, Moore R, et al (2013) High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 342:850\u0026ndash;853. https://doi.org/10.1126/science.1244693\u003c/li\u003e\n\u003cli\u003eHe KS, Bradley BA, Cord AF, et al (2015) Will remote sensing shape the next generation of species distribution models? Remote Sens Ecol Conserv 1:4\u0026ndash;18. https://doi.org/10.1002/rse2.7\u003c/li\u003e\n\u003cli\u003eHirzel AH, Le Lay G, Helfer V, et al (2006) Evaluating the ability of habitat suitability models to predict species presences. Ecol Model 199:142\u0026ndash;152. https://doi.org/10.1016/j.ecolmodel.2006.05.017\u003c/li\u003e\n\u003cli\u003eHughes AC, Orr MC, Ma K, et al (2021) Sampling biases shape our view of the natural world. Ecography 44:1259\u0026ndash;1269. https://doi.org/10.1111/ecog.05926\u003c/li\u003e\n\u003cli\u003eIacolina L, Pertoldi C, Amills M, et al (2018) Hotspots of recent hybridization between pigs and wild boars in Europe. Sci Rep 8:17372. https://doi.org/10.1038/s41598-018-35865-8\u003c/li\u003e\n\u003cli\u003eInglada J, Vincent A, Arias M, et al (2017) Operational High Resolution Land Cover Map Production at the Country Scale Using Satellite Image Time Series. Remote Sens 9:95. https://doi.org/10.3390/rs9010095\u003c/li\u003e\n\u003cli\u003eJakubik J, Roy S, Phillips CE, et al (2023) Foundation Models for Generalist Geospatial Artificial Intelligence\u003c/li\u003e\n\u003cli\u003eJasińska KD, Żmihorski M, Krauze-Gryz D, et al (2019) Linking habitat composition, local population densities and traffic characteristics to spatial patterns of ungulate-train collisions. J Appl Ecol 56:2630\u0026ndash;2640. https://doi.org/10.1111/1365-2664.13495\u003c/li\u003e\n\u003cli\u003eJim\u0026eacute;nez L, Sober\u0026oacute;n J (2022) Estimating the fundamental niche: Accounting for the uneven availability of existing climates in the calibration area. Ecol Model 464:109823. https://doi.org/10.1016/j.ecolmodel.2021.109823\u003c/li\u003e\n\u003cli\u003eJim\u0026eacute;nez L, Sober\u0026oacute;n J (2020) Leaving the area under the receiving operating characteristic curve behind: An evaluation method for species distribution modelling applications based on presence-only data. Methods Ecol Evol 11:1571\u0026ndash;1586. https://doi.org/10.1111/2041-210X.13479\u003c/li\u003e\n\u003cli\u003eJim\u0026eacute;nez-Valverde A (2012) Insights into the area under the receiver operating characteristic curve (AUC) as a discrimination measure in species distribution modelling. Glob Ecol Biogeogr 21:498\u0026ndash;507. https://doi.org/10.1111/j.1466-8238.2011.00683.x\u003c/li\u003e\n\u003cli\u003eKe G, Meng Q, Finley T, et al (2017) LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In: Advances in Neural Information Processing Systems. Curran Associates, Inc.\u003c/li\u003e\n\u003cli\u003eKeuling O, Stier N, Roth M (2008) How does hunting influence activity and spatial usage in wild boar Sus scrofa L.? Eur J Wildl Res 54:729\u0026ndash;737. https://doi.org/10.1007/s10344-008-0204-9\u003c/li\u003e\n\u003cli\u003eLangbein J, Putman R, Pokorny B (2011) Traffic collisions involving deer and other ungulates in Europe and available measures for mitigation. In: Apollonio M, Andersen R, Putman R (eds) Ungulate Management in Europe: Problems and Practices. Cambridge University Press, Cambridge, pp 215\u0026ndash;259\u003c/li\u003e\n\u003cli\u003eLobo JM, Jim\u0026eacute;nez-Valverde A, Hortal J (2010) The uncertain nature of absences and their importance in species distribution modelling. Ecography 33:103\u0026ndash;114. https://doi.org/10.1111/j.1600-0587.2009.06039.x\u003c/li\u003e\n\u003cli\u003eLobo JM, Jim\u0026eacute;nez-Valverde A, Real R (2008) AUC: a misleading measure of the performance of predictive distribution models. Glob Ecol Biogeogr 17:145\u0026ndash;151. https://doi.org/10.1111/j.1466-8238.2007.00358.x\u003c/li\u003e\n\u003cli\u003eMacKenzie DI, Nichols JD, Royle JA, et al (2017) Occupancy estimation and modeling: inferring patterns and dynamics of species occurrence. Elsevier\u003c/li\u003e\n\u003cli\u003eMaillard D, Gaillard JM, Hewison M a. J, et al (2010) Ungulates and their management in France. Cambridge University Press, p 441\u003c/li\u003e\n\u003cli\u003eMarin C, Couderchet L, Le Campion G, Werno J (2024) Wildlife and the city. Modelling wild boar use of urban nature: Empirical contribution, methodological proposal. Urban Ecosyst 27:1291\u0026ndash;1312. https://doi.org/10.1007/s11252-024-01510-8\u003c/li\u003e\n\u003cli\u003eMart\u0026iacute;nez-Minaya J, Cameletti M, Conesa D, Pennino MG (2018) Species distribution modeling: a statistical review with focus in spatio-temporal issues. Stoch Environ Res Risk Assess 32:3227\u0026ndash;3244. https://doi.org/10.1007/s00477-018-1548-7\u003c/li\u003e\n\u003cli\u003eMassei G, Kindberg J, Licoppe A, et al (2015) Wild boar populations up, numbers of hunters down? A review of trends and implications for Europe. Pest Manag Sci 71:492\u0026ndash;500. https://doi.org/10.1002/ps.3965\u003c/li\u003e\n\u003cli\u003eMesgaran MB, Cousens RD, Webber BL (2014) Here be dragons: a tool for quantifying novelty due to covariate range and correlation change when projecting species distribution models. Divers Distrib 20:1147\u0026ndash;1159. https://doi.org/10.1111/ddi.12209\u003c/li\u003e\n\u003cli\u003eMeyer A, Milani G, Lack N, et al (2025) Machine learning-based multimodal data fusion for the Swiss land use statistics. Eur J Remote Sens 58:2508420. https://doi.org/10.1080/22797254.2025.2508420\u003c/li\u003e\n\u003cli\u003eMeyer C, Weigelt P, Kreft H (2016) Multidimensional biases, gaps and uncertainties in global plant occurrence information. Ecol Lett 19:992\u0026ndash;1006. https://doi.org/10.1111/ele.12624\u003c/li\u003e\n\u003cli\u003eMilano F, Casazza G, Galimberti A, et al (2024) Combining distribution modelling and phylogeography to understand present, past and future of an endangered spider. BMC Ecol Evol 24:106. https://doi.org/10.1186/s12862-024-02295-2\u003c/li\u003e\n\u003cli\u003eMorelle K, Bubnicki J, Churski M, et al (2020) Disease-Induced Mortality Outweighs Hunting in Causing Wild Boar Population Crash After African Swine Fever Outbreak. Front Vet Sci 7:. https://doi.org/10.3389/fvets.2020.00378\u003c/li\u003e\n\u003cli\u003eMorelle K, Lehaire F, Lejeune P (2014) Is wild boar heading towards movement ecology? A review of trends and gaps. Wildl Biol 20:196\u0026ndash;205. https://doi.org/10.2981/wlb.00017\u003c/li\u003e\n\u003cli\u003eMorelle K, Lehaire F, Lejeune P (2013) Spatio-temporal patterns of wildlife-vehicle collisions in a region with a high-density road network. Nat Conserv 5:53\u0026ndash;73. https://doi.org/10.3897/natureconservation.5.4634\u003c/li\u003e\n\u003cli\u003eMorelle K, Lejeune P (2015) Seasonal variations of wild boar Sus scrofa distribution in agricultural landscapes: a species distribution modelling approach. Eur J Wildl Res 61:45\u0026ndash;56. https://doi.org/10.1007/s10344-014-0872-6\u003c/li\u003e\n\u003cli\u003eMu\u0026ntilde;oz-Sabater J, Dutra E, Agust\u0026iacute;-Panareda A, et al (2021) ERA5-Land: a state-of-the-art global reanalysis dataset for land applications. Earth Syst Sci Data 13:4349\u0026ndash;4383. https://doi.org/10.5194/essd-13-4349-2021\u003c/li\u003e\n\u003cli\u003ePaudel PK, Hais M, Kindlmann P (2015) Habitat suitability models of mountain ungulates: identifying potential areas for conservation. Zool Stud 54:e37. https://doi.org/10.1186/s40555-015-0116-9\u003c/li\u003e\n\u003cli\u003ePettorelli N, Vik JO, Mysterud A, et al (2005) Using the satellite-derived NDVI to assess ecological responses to environmental change. Trends Ecol Evol 20:503\u0026ndash;510. https://doi.org/10.1016/j.tree.2005.05.011\u003c/li\u003e\n\u003cli\u003ePhillips SJ, Anderson RP, Dud\u0026iacute;k M, et al (2017) Opening the black box: An open‐source release of Maxent. Ecography 40:887\u0026ndash;893. https://doi.org/10.1111/ecog.03049\u003c/li\u003e\n\u003cli\u003ePhillips SJ, Anderson RP, Schapire RE (2006) Maximum entropy modeling of species geographic distributions. Ecol Model 190:231\u0026ndash;259. https://doi.org/10.1016/j.ecolmodel.2005.03.026\u003c/li\u003e\n\u003cli\u003ePhillips SJ, Dud\u0026iacute;k M (2008) Modeling of species distributions with Maxent: new extensions and a comprehensive evaluation. Ecography 31:161\u0026ndash;175. https://doi.org/10.1111/j.0906-7590.2008.5203.x\u003c/li\u003e\n\u003cli\u003ePuissant A, Sell\u0026eacute; A, Baghdadi N, et al (2019) The \u0026lsquo;urban\u0026rsquo; component of the French Land Data and Services Centre (THEIA). In: 2019 Joint Urban Remote Sensing Event (JURSE). pp 1\u0026ndash;4\u003c/li\u003e\n\u003cli\u003eRapinel S, Panhelleux L, Lemercier B, et al (2023) Projet de cartographie nationale des milieux humides \u0026ndash; Rapport m\u0026eacute;thodologique de la pr\u0026eacute;localisation des milieux et des zones humides, de la cartographie des habitats et de l\u0026rsquo;\u0026eacute;valuation des fonctions. Patrinat (OFB-MNHN-CNRS-IRD)\u003c/li\u003e\n\u003cli\u003eReside AE, VanDerWal JJ, Kutt AS, Perkins GC (2010) Weather, Not Climate, Defines Distributions of Vagile Bird Species. PLOS ONE 5:e13569. https://doi.org/10.1371/journal.pone.0013569\u003c/li\u003e\n\u003cli\u003eRoberts DR, Bahn V, Ciuti S, et al (2017) Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40:913\u0026ndash;929. https://doi.org/10.1111/ecog.02881\u003c/li\u003e\n\u003cli\u003eRoyle JA, Chandler RB, Sollmann R, Gardner B (2013) Spatial capture-recapture. Academic press\u003c/li\u003e\n\u003cli\u003eSanderson EW, Jaiteh M, Levy MA, et al (2002) The Human Footprint and the Last of the Wild: The human footprint is a global map of human influence on the land surface, which suggests that human beings are stewards of nature, whether we like it or not. BioScience 52:891\u0026ndash;904. https://doi.org/10.1641/0006-3568(2002)052%5B0891:THFATL%5D2.0.CO;2\u003c/li\u003e\n\u003cli\u003eSantos SM, Carvalho F, Mira A (2017) Current knowledge on wildlife mortality in railways. Railw Ecol 11\u0026ndash;22\u003c/li\u003e\n\u003cli\u003eSauter-Louis C, Conraths FJ, Probst C, et al (2021) African Swine Fever in Wild Boar in Europe\u0026mdash;A Review. Viruses 13:1717. https://doi.org/10.3390/v13091717\u003c/li\u003e\n\u003cli\u003eSchleimer A, Richart L, Drygala F, et al (2022) Introgressive hybridisation between domestic pigs (Sus scrofa domesticus) and endemic Corsican wild boars (S. s. meridionalis): effects of human-mediated interventions. Heredity 128:279\u0026ndash;290. https://doi.org/10.1038/s41437-022-00517-1\u003c/li\u003e\n\u003cli\u003eSchley L, Roper TJ (2003) Diet of wild boar Sus scrofa in Western Europe, with particular reference to consumption of agricultural crops. Mammal Rev 33:43\u0026ndash;56. https://doi.org/10.1046/j.1365-2907.2003.00010.x\u003c/li\u003e\n\u003cli\u003eSchwartz ALW, Shilling FM, Perkins SE (2020) The value of monitoring wildlife roadkill. Eur J Wildl Res 66:18. https://doi.org/10.1007/s10344-019-1357-4\u003c/li\u003e\n\u003cli\u003eSeiler A (2004) Trends and spatial patterns in ungulate-vehicle collisions in Sweden. Wildl Biol 10:301\u0026ndash;313. https://doi.org/10.2981/wlb.2004.036\u003c/li\u003e\n\u003cli\u003eSicacha-Parada J, Steinsland I, Cretois B, Borgelt J (2019) Accounting for spatial varying sampling effort due to accessibility in Citizen Science data: A case study of moose in Norway\u003c/li\u003e\n\u003cli\u003eStillfried M, Gras P, B\u0026ouml;rner K, et al (2017) Secrets of Success in a Landscape of Fear: Urban Wild Boar Adjust Risk Perception and Tolerate Disturbance. Front Ecol Evol 5:. https://doi.org/10.3389/fevo.2017.00157\u003c/li\u003e\n\u003cli\u003eSyfert MM, Smith MJ, Coomes DA (2013) The Effects of Sampling Bias and Model Complexity on the Predictive Performance of MaxEnt Species Distribution Models. PLOS ONE 8:e55158. https://doi.org/10.1371/journal.pone.0055158\u003c/li\u003e\n\u003cli\u003eTen Caten C, Dallas T (2023) Thinning occurrence points does not improve species distribution model performance. Ecosphere 14:e4703\u003c/li\u003e\n\u003cli\u003eThurfjell H, Ball JP, \u0026Aring;hl\u0026eacute;n P-A, et al (2009) Habitat use and spatial patterns of wild boar Sus scrofa (L.): agricultural fields and edges. Eur J Wildl Res 55:517\u0026ndash;523. https://doi.org/10.1007/s10344-009-0268-1\u003c/li\u003e\n\u003cli\u003eThurfjell H, Spong G, Ericsson G (2013) Effects of hunting on wild boar Sus scrofa behaviour. Wildl Biol 19:87\u0026ndash;93. https://doi.org/10.2981/12-027\u003c/li\u003e\n\u003cli\u003eVajas P, Von Essen E, Tickle L, Gamelon M (2023) Meeting the challenges of wild boar hunting in a modern society: The case of France. Ambio 52:1359\u0026ndash;1372. https://doi.org/10.1007/s13280-023-01852-1\u003c/li\u003e\n\u003cli\u003eValavi R, Elith J, Lahoz-Monfort JJ, Guillera-Arroita G (2019) blockCV: An r package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models. Methods Ecol Evol 10:225\u0026ndash;232. https://doi.org/10.1111/2041-210X.13107\u003c/li\u003e\n\u003cli\u003eVetter SG, Puskas Z, Bieber C, Ruf T (2020) How climate change and wildlife management affect population structure in wild boars. Sci Rep 10:7298. https://doi.org/10.1038/s41598-020-64216-9\u003c/li\u003e\n\u003cli\u003eWang L, Diao C, Lu Y (2025) The role of remote sensing in species distribution models: a review. Int J Remote Sens 46:661\u0026ndash;685. https://doi.org/10.1080/01431161.2024.2421949\u003c/li\u003e\n\u003cli\u003eWarren DL, Matzke NJ, Iglesias TL (2020) Evaluating presence-only species distribution models with discrimination accuracy is uninformative for many applications. J Biogeogr 47:167\u0026ndash;180. https://doi.org/10.1111/jbi.13705\u003c/li\u003e\n\u003cli\u003eWelander J (2000) Spatial and temporal dynamics of wild boar (Sus scrofa) rooting in a mosaic landscape. J Zool 252:263\u0026ndash;271. https://doi.org/10.1111/j.1469-7998.2000.tb00621.x\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-wildlife-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejwr","sideBox":"Learn more about [European Journal of Wildlife Research](http://link.springer.com/journal/10344)","snPcode":"10344","submissionUrl":"https://submission.nature.com/new-submission/10344/3","title":"European Journal of Wildlife Research","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Wild Boar (Sus scrofa), Species Distribution Models, Multitemporal Modeling, Wildlife-Vehicle Collisions, Remote Sensing, Vision Transformer","lastPublishedDoi":"10.21203/rs.3.rs-8798859/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8798859/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWild boar \u003cem\u003e(Sus scrofa)\u003c/em\u003e populations have expanded rapidly across Europe, leading to escalating human-wildlife conflicts (HWCs), notably wildlife-vehicle collisions (WVCs), increased agricultural damage and disease transmission. In continental France these issues are compounded by the species\u0026rsquo; ecological adaptability exhibiting increasing overlaps with urbanization and transportation networks. In this study, we leverage time-series environmental data, computer vision and Species Distribution Modeling (SDM) to predict wild boar habitat suitability and investigate its spatiotemporal drivers.\u003c/p\u003e \u003cp\u003eWe use presence-only data from WVC reports on the national railway and road networks, along with publicly available \u003cem\u003eGBIF\u003c/em\u003e observation collections, to enhance the predictive power of our SDMs, while addressing inherent sampling biases of these datasets with tailored corrections. A key innovation of this study is the integration of large scale, very-high-resolution land cover predictors explicitly focused on wild boar resource preference. By fine-tuning a multitemporal \u003cem\u003eVision Transformer\u003c/em\u003e foundational AI model on multispectral satellite remote sensing imagery we capture subtle seasonal phenological differences.\u003c/p\u003e \u003cp\u003eOur results highlight clear spatial, seasonal and annual variations in wild boar habitat suitability. The multitemporal SDM pipeline offers improved ecological realism and resilience to climate extremes, yielding meaningful predictions when extrapolating to novel environmental scenarios. The methodological and ecological insights gained through this study provide actionable knowledge for French transportation planning, agriculture and wildlife management. Identifying regions with high seasonal habitat suitability can inform targeted and preventive interventions. More broadly, our results demonstrate that advanced, data-driven methods are becoming indispensable for proactively and sustainably addressing HWCs in an increasingly anthropogenic world.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e","manuscriptTitle":"Wild Boar Collision Data and Satellite Computer Vision Refine Habitat Suitability Mapping across France","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-26 15:05:48","doi":"10.21203/rs.3.rs-8798859/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-08T16:03:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"39476949098338034495828942355929479164","date":"2026-03-12T15:19:43+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-23T21:09:11+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-23T20:43:03+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-23T13:04:12+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Journal of Wildlife Research","date":"2026-02-05T15:05:47+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-wildlife-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejwr","sideBox":"Learn more about [European Journal of Wildlife Research](http://link.springer.com/journal/10344)","snPcode":"10344","submissionUrl":"https://submission.nature.com/new-submission/10344/3","title":"European Journal of Wildlife Research","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"c5c9890b-1b3f-4391-b3a4-0a48b40a25f1","owner":[],"postedDate":"February 26th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-02-26T15:05:48+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-26 15:05:48","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8798859","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8798859","identity":"rs-8798859","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}