Background & Summary

Species Distribution Models (SDMs; also commonly referred to as Habitat Suitability Models HSMs) have become essential tools in biodiversity and conservation science, generalizing information on species’ habitat suitability across geographic space and providing insights into the environmental factors at play1,2,3. Habitat suitability maps derived from SDMs have proven valuable resources across multiple domains of ecological research and conservation applications, such as ecological infrastructure planning, habitat restoration, invasive species management, and climate change adaptation4,5,6.

Standardized habitat suitability maps, developed using state-of-the-art methods and covering an extensive range of species, are essential for informed biodiversity assessments6,7,8. In response, several countries have launched initiatives to create comprehensive and standardized national databases of species’ habitat suitability maps. Examples of such national-level initiatives include the effort by NatureServe to map habitat suitability for 2,216 at-risk species across the contiguous United States (https://www.natureserve.org/), and a nationwide database in Australia covering 1,872 terrestrial and freshwater vertebrate species9. At a continental scale, the European Breeding Bird Atlas provides modelled habitat suitability maps for over 500 bird species across Europe10. However, even in well-studied countries in Western Europe and North America, these efforts often focus on a few well-known taxonomic groups, while many others remain underrepresented or entirely omitted. Switzerland is a typical example of this issue: despite considerable efforts by the Swiss Species Information Centers (InfoSpecies; www.infospecies.ch) to compile extensive species records and display maps based on observational data, many taxonomic groups remain only partially mapped or entirely unmapped.

By generalizing information from observational data, even when sparse, modelled habitat suitability maps provide a crucial complement. Moreover, when generated using a standardized and unified method, these maps enable systematic comparisons across species groups. Here, we demonstrate how such upscaling and generalization can be achieved at the national scale for thousands of species, using Switzerland as a case example. Specifically, we introduce SDMapCH (v1.3), a nationwide raster database of modelled species habitat suitability maps at a 25-meter resolution. SDMapCH provides individual maps for about 7,500 species, as well as aggregated maps for 11 major taxonomic groups, 26 ecological guilds, and 5 conservation status categories, under both present conditions and future climate change scenarios projected through the end of the century. SDMapCH was developed using the N-SDM modelling pipeline11, an end-to-end platform built on a spatially nested hierarchical framework and optimized for high-performance computing environments. N-SDM enables multi-level and multi-resolution integration of species and covariate data, addressing niche truncation issues12, and supporting parallel high-resolution modelling of numerous species within tight timeframes.

By providing standardized, high-resolution habitat suitability maps for diverse species across various taxonomic and functional groups, SDMapCH stands as a key resource for advancing scientific analyses and streamlining workflows for a wide range of applications in Switzerland and beyond. These applications include, but are not limited to, biological conservation in changing ecosystems13,14,15, biodiversity-energy development conflicts16,17,18, and evaluating nature’s contributions to people19,20,21. Since species, environmental data, and modelling techniques are constantly evolving, SDMapCH was structured to allow updates through the release of new versions. The technical characteristics and parameters specified in this article are those used for the first open version of SDMapCH (v1.3), following previous preliminary versions. The SDMapCH database and associated metadata, including the ODMAP reporting protocol22, detailed lists of species and candidate covariates, information on covariate selection and importance, and model evaluation results, are openly available on Dryad.

Methods

Species’ habitat suitability modelling

To model and map species’ habitat suitability, we used the N-SDM software11. N-SDM is an end-to-end SDM platform built around a spatially-nested framework. It combines a ‘global’ model, which quantifies the species’ response to bioclimatic conditions across its entire distributional range, with a ‘regional’ model fitted using finer-scale habitat covariates. Global and regional models were combined by using the ‘covariate’ nesting strategy, which consists in using the global model predictions (i.e., bioclimatic envelope) as a forced additional covariate for fitting the regional model. In the metadata.zip file accompanying the maps on Dryad, we provided both the N-SDM configuration file containing user-specific settings for all editable options, and the ODMAP reporting protocol22.

Study areas

Following the spatially-nested framework of N-SDM, we distinguished between regional- and global-level study areas. The regional-level area (projection extent) included all of Switzerland, with a total area of about 41,000 km². For the global-level area (calibration extent), we used a bounding box approximately 250 times larger, covering the European continent and a portion of Northern Africa (ranging from 32.60 °N to 71.70 °N, and from 28.56 °W to 40.21 °E). The projection and calibration extents were the same for all species, selected as a tradeoff to encompass most species’ climatic niches while keeping computational costs manageable. Moreover, incorporating data from distant continents might increase uncertainty due to data consistency issues and taxonomic differences23,24,25. This extent also balanced spatial scale and model accuracy26,27,28, preventing niche truncation for widely distributed species and maintaining accuracy for those limited to specific locations.

Species data

Two sets of species occurrence records were used: one for the regional level (within Switzerland) and one for the global level (outside Switzerland). For Switzerland, validated occurrence records for 21,290 species aggregated at a 25-m resolution and covering the 1980–2021 period were provided by the Swiss Species Information Center InfoSpecies (https://doi.org/10.15468/x9z635)29. To avoid modelling issues related to low sample size (e.g., model convergence, number of covariates), only species with more than 50 occurrence records after spatial thinning (described below) were retained for modelling, resulting in 7,508 species (species list with detailed taxonomic information available in the metadada.zip file accompanying the maps on Dryad). Species were classified into 11 major taxonomic groups: amphibians (16 species modelled out of 21 known to be part of Switzerland’s native biodiversity30), arthropods (2,266/32,685; including 296 Ephemeroptera, Plecoptera, Trichoptera, and Odonata (EPTO)), breeding birds (175/205), ferns and mosses (589/1,236), fishes (30/71), fungi and lichens (1,700/11,121), mammals (77/86), molluscs (157/270), reptiles (14/16), and vascular plants (2,096/2,536). The 378 remaining modelled species were nonnative to Switzerland, including 314 vascular plants, 19 arthropods, and 13 fungi and lichens. In addition, when available, InfoSpecies provided information on the ecological guild of each species, totaling 26 guilds (i.e., groups of species with similar ecological needs, based on the typology of Swiss natural habitats31). Red List conservation status for each species in Switzerland30 were also provided by InfoSpecies. Occurrence records for outside of Switzerland for matching species and the same 1980–2021 period were obtained from the Global Biodiversity Information Facility (GBIF; https://www.gbif.org/) (https://doi.org/10.15468/dl.fktyas). The extracted GBIF records were tested and filtered for coordinate validity, which included identifying equal latitude/longitude values, rounded coordinates, and zero values32. Additionally, we filtered for spatial and temporal outliers by extracting occurrences only within the area defined by a shapefile covering the target European domain and specifying the same period as the Swiss data (1980–2021). Further details on GBIF data processing and matching with InfoSpecies species names are provided in the metadada.zip file accompanying the maps on Dryad. Sampling bias was corrected using a spatial filter approach that thinned the occurrence records, mitigating observation clusters33,34,35. We applied level-specific minimum distances between two observations for spatial thinning: 500 m for the regional level and 1 km for the global level. Ideally, these values should be based on species biology, environmental conditions, and previous research on the study system34,36,37. However, such comprehensive information was not available for all species. Therefore, we opted for a tradeoff, considering the spatial resolution of the covariate data (25 m within Switzerland and 30 arcseconds outside) and choosing conservatively larger distances to ensure efficient spatial disaggregation. Finally, for each species and occurrence set, 10,000 background points were randomly generated, using a single background approach for efficiency since differences between methods are usually minimal38,39,40.

Covariate data

To model the habitat suitability of each individual species, we used a suite of 877 candidate covariates from eight main categories (bioclimatic, edaphic, hydrologic, land use and cover, population density, topographic, transportation, and vegetation; detailed covariate list available in the metadada.zip file accompanying the maps on Dryad). Only bioclimatic covariates (n = 19) were used for global-level models, while all other covariates were used for regional-level models. Regional-level covariates were extracted from the 25-m resolution SWECO25 raster database41. To capture environmental conditions across a broader area surrounding the occurrence records, habitat covariates were retrieved using 13 focal moving windows available in SWECO25, each computing the average pixel value within circular neighborhoods of increasing size, with radii ranging from 25 m to 5 km. For global-level covariates, to benefit from best-available products, we used specific data sources for inside and outside of Switzerland. For Switzerland, bioclimatic covariates were extracted from the CHclim25 dataset (https://doi.org/10.5281/zenodo.10635681), which gather Swiss-wide climate layers downscaled at 25-m resolution. For outside of Switzerland, we used 30-arcsecond CHELSA layers42. To project potential future habitat suitability, we retrieved projection data of all 19 bioclimatic covariates from the CHclim25 dataset for the four coupled global and regional climate models (GCM-RCM) available (CLMCOM-CCLM4/HADGEM, DMI-HIRHAM/ECEARTH, MPICSC-REMO2/MPIESM, and SMHI-RCA/IPSL), three future intervals (2020–2049, 2045–2074, and 2070–2099), and two greenhouse gas concentration trajectories or representative concentration pathways (RCPs)43, RCP4.5 (“Low Carbon”) and RCP8.5 (“High Carbon”). All covariates were standardized to zero mean and unit variance.

Covariate selection

As a preliminary step, candidate covariates identified by experts as being likely irrelevant for modelling a species (or group of species) were discarded by providing N-SDM with a so called ‘expert table’ (available in the metadata.zip file accompanying the maps on Dryad) filled in collaboration with species experts. For instance, edaphic covariates, such as soil pH or aeration, were not used for modelling butterflies and reptile species, and hydrological covariates related to stream morphology were discarded for modelling fungi or bat species. By doing so, we ensured that only biologically realistic and relevant variables were considered. The two-step automated covariate selection procedure included in N-SDM, ‘covsel’44, was then applied for each species at both global and regional levels, with default values. In short, covsel first reduces the dimensionality of the candidate set by eliminating less informative covariates among collinear pairs (default Pearson correlation coefficient: |r| > 0.70) and remaining covariates are used to fit models with algorithm-specific embedded regularization procedures. Covariates are ranked on the basis of their importance scores and the top k covariates are retained, with default k = round(log2(number of occurrences)). To avoid models ending with too many covariates, the default maximum possible value for k is 12. Since the covariate nesting strategy was used, the global-model output was forced as an additional covariate in all regional models. Relative covariate importance, ranging between 0 and 1, was automatically outputted by N-SDM (see help files of the ‘nsdm.varimp’ function for more details). Finally, in an attempt to mitigate potential observational biases in occurrence data, we used the N-SDM option that allows setting a zero value to the set of selected covariates expected to be related to these biases, including distance from the transportation network or settlements (see the N-SDM configuration file available in the metadada.zip file accompanying the maps on Dryad for the full list of observational covariates). This approach can help reduce potential observer bias in the prediction area, provided that the selected covariates effectively capture the spatial patterns of observer activity45,46,47.

Model fitting and selection

The five modelling algorithms available in N-SDM – Generalized Linear Model (GLM)48, Generalized Additive Model (GAM)49, Maxnet (MAX)50, Random Forest (RF)51, and light Gradient Boosted Machine (GBM)52 – were fitted using their default N-SDM values for hyperparameter tuning. To account for class imbalance, background points and occurrences were weighted equally in the models39. Model accuracy was evaluated using a split-sample approach repeated 100 times with 30% of the data kept for validation2. For each model, the best combination of hyperparameters was identified using the average “Score”11 of three evaluation metrics including the re-scaled Area Under the Curve′ (AUC′; or Somers’ D, such as \({AU}{C}^{{\prime} }={AUC}\ast 2-1\))53, the maximized True Skill Statistic (maxTSS)2, and the Continuous Boyce Index (CBI)54. All these evaluation metrics ranged between −1 and +1.

Spatial predictions

We predicted habitat suitability values for each species and five modelling algorithms for the baseline period and three future intervals × two RCPs. Results from the five modelling algorithms were individually mapped over a 25-m resolution grid of 64,007,390 cells covering Switzerland (CH1903+/ LV95 geographic coordinate system), and ensembled together by computing the mean of the five maps. Aggregated maps for the 11 major taxonomic groups, 26 ecological guilds, and 5 conservation status categories were computed by taking the mean of the maps corresponding to the subset of species within each group. Maps were provided for both mean values and the coefficient of variation.

Data Records

The SDMapCH (v1.3) database is publicly available on Dryad55. It consists of five main map archives (tar.zst) and one companion metadata archive (.zip). SDMapCHv1_3_current.tar.zst contains GeoTIFF (.tif) maps for the current period. SDMapCHv1_3_2020_2049.tar.zst, SDMapCHv1_3_2045_2074.tar.zst, and SDMapCHv1_3_2070_2099.tar.zst contain maps for the three future periods (2020–2049, 2045–2074, 2070–2099) under the two RCP scenarios (RCP4.5 and RCP8.5). SDMapCHv1_3_aggregates.tar.zst contains aggregated maps for species grouped by ecological guild, Swiss Red List status, major taxonomic group, and all species combined, for the current and three future periods. Finally, metadata.zip contains supporting tabular and text information in CSV and PDF format.

Data Overview

Individual habitat suitability maps were compiled in SDMapCH (v1.3)55 for 7,425 species out of 7,508 modelled, excluding 83 sensitive species (1.1%) for which the dissemination of distribution data posed conservation or legal concerns. These concerns relate to the risk of collection, destruction, or disturbance of rare and endangered species, such as certain orchids and reptiles, their potential exploitation through the international illegal wildlife trade, and the requirements of Swiss biodiversity conservation law. However, individual habitat suitability maps for all these sensitive species are available upon formal request to InfoSpecies (Swiss Species Information Centre) by contacting info.species@unine.ch, in accordance with their standard access procedure. In addition to the individual maps, aggregated versions (mean and coefficient of variation) for the 11 major taxonomic groups, 26 ecological guilds, and 5 conservation status categories were included, resulting in a total size of 487 GB. We illustrated with maps obtained for individual species (Fig. 1) and major taxonomic groups (Fig. 2) for the current period and future projections.

Fig. 1
figure 1

SDMapCH (v1.3) habitat suitability maps for example species for current and future (2070–2099) periods under the low (RCP4.5) and high (RCP8.5) greenhouse gas scenarios. For each species, values were rescaled to a 0–100 range for data storage purposes and improved readability, with, from top to bottom: Acanthis flammea (bird), Eryngium alpinum (vascular plant), Lacerta agilis (reptile), Macrolepiota procera (fungus), Parnassius phoebus (arthropod), and Rana temporaria (amphibian). For the use and validation of individual habitat suitability maps, please note: (i) The habitat suitability maps were primarily developed for research purposes; (ii) These maps were not validated for each individual species; therefore, interpretation at the species level should be complemented with additional taxonomic or faunistic expertise; (iii) The results of these models were not systematically validated by InfoSpecies; (iv) Individual maps should not be used directly for practical applications or management decisions without consultation with experts in the relevant taxonomic groups. The authors disclaim responsibility for the accuracy of individual maps.

Full size image
Fig. 2
figure 2

SDMapCH (v1.3) (a) mean habitat suitability and (b) coefficient of variation maps for major taxonomic groups for current and future (2070–2099) periods under the low (RCP4.5) and high (RCP8.5) greenhouse gas scenarios. For each taxonomic group, mean values were rescaled to a 0–100 range for data storage purposes and improved readability, and coefficient of variation values were capped at 1,000 for readability. Arthropods*: aquatic arthropods including Ephemeroptera, Plecoptera, Trichoptera, and Odonata (EPTO). Arthropods**: All other arthropods excluding EPTO. Number of modelled species per group: amphibians (18), arthropods (2,285, including 296 EPTO) breeding birds (179), ferns and mosses (591), fishes (37), fungi and lichens (1,713), mammals (83), molluscs (168), reptiles (15), and vascular plants (2,410). For the use and validation of aggregated habitat suitability maps, please note: (i) The habitat suitability maps were primarily developed for research purposes and must be validated for case-specific conservation purposes; (ii) The maps represent aggregations of taxonomic groups for which species were not validated individually; therefore, interpretation should be complemented with additional taxonomic or faunistic expertise; (iii) The results of these models were not systematically validated by InfoSpecies; (iv) The maps should not be used directly for practical applications or management decisions without consultation with experts in the relevant taxonomic groups.

Full size image

Technical Validation

Model evaluation

Figure 3a displays cross-validated evaluation metrics obtained for the final ensemble maps of the 7,508 individual species. The mean ± standard deviation Score value was 0.89 ± 0.05 indicating rather high predictive performances. We provided detailed model evaluation metrics in the metadata.zip file accompanying the maps on Dryad, including breakdowns of AUC′, maxTSS, CBI, and their average Score, summarized by taxonomic group and modelling algorithm, including the ensemble version. Across all metrics, GLM showed the lowest mean values and the highest standard deviations. MAX, RF, GBM, and RF achieved the highest mean values for AUC′, CBI, maxTSS, and Score, respectively. Group-wise, amphibians achieved the highest ensemble performance for AUC′, maxTSS, and Score, whereas birds had the highest CBI. Fungi and lichens tended to have the lowest ensemble scores across all metrics. For each group, we compared the ensemble score with the highest score obtained by any single algorithm. On average at the group level, ensembles performed close to the best single algorithm, with differences ranging from −0.001 to −0.009 for AUC′, −0.012 to −0.025 for maxTSS, −0.041 to −0.111 for CBI, and −0.017 to −0.034 for Score. These results indicate that ensembles deliver consistently high and stable performance across taxa, although in all cases a single algorithm slightly outperformed them.

Fig. 3
figure 3

SDMapCH (v1.3) model evaluation and covariate selection results. (a) Cross-validated model evaluation metrics (AUC’: Somers’ D Area Under the Curve, CBI: Continuous Boyce Index, maxTSS: maximized True Skill, and their average Score) for the 7,508 species and averaged across the five modelling algorithms (Generalized Linear Model, Generalized Additive Model, Maxnet, Random Forest, and Light Gradient Boosted Machine). Violin plots illustrate data distributions, showing the median (black dots) and the 2.5th to 97.5th percentile range (black lines). (b) Selection frequency and importance of the covariate categories for the 7,508 species. The covariate ‘count index’ (large colored bars) is the log of the count of covariates from a category selected in the final models divided by the overall number of covariates available in this category. The ‘importance’ (thin dark bars) is the average importance value of the covariates from a category in the final models across the 7,508 species and five algorithms.

Full size image

In addition to quantitative evaluation metrics, taxonomic experts from InfoSpecies centers qualitatively assessed the plausibility of habitat suitability maps for around 100 selected species across various taxonomic groups (e.g., birds, plants, reptiles). Overall, expert feedback highlighted that the models from SDMapCH v1.3 presented in this article were well-calibrated and performed better than previous preliminary versions (v1.1 and v1.2). As the modelling approach was not optimized for individual species or species groups, some models exhibited effects of sampling biases present in the original datasets, resulting in localized variations in predictions. For example, certain habitats (e.g., wetlands vs. forests) may be over- or under-represented in occurrence records for specific groups, leading to corresponding over- or under-predictions in habitat suitability56. More specifically, several reptile models showed underrepresentation of favorable areas in southwestern and southeastern parts of the country, with certain species like Natrix helvetica predicted too narrowly despite a known broader range. Bird models were identified by the experts as showing slight signals of detection bias in the raw observation data, with species generally being more elusive in their natural habitats (e.g., Accipiter gentilis, Accipiter nisus) being underpredicted, while birds that are much more commonly reported, such as from urban areas with higher observation effort (e.g., Acrocephalus palustris) being slightly overpredicted; in addition, probably due to missing key habitat covariates, reduced accuracy was expertized for wetland specialists (e.g., Acrocephalus scirpaceus, Actitis hypoleucos). Butterfly models were generally well-calibrated but exhibited localized overprediction for some species, such as Coenonympha glycerion in southwestern regions, and in some cases reflected biases linked to agricultural and garden habitats (e.g., Pieris napi). Plant models sometimes lacked environmental covariates for edaphic or alluvial specialists (e.g., Myricaria germanica, Heracleum mantegazzianum), leading to restrictive predictions for certain species (e.g., Prunus serotina, Vitis sylvestris) and overprediction in specific areas (e.g., Agrostis canina in southwestern regions). Mammal models were generally robust, but localized overprediction were reported for species such as Crocidura leucodon and Felis silvestris, particularly in southwestern regions. While refinements in data processing, covariate selection, and species-specific optimization could improve the ecological robustness of future versions, such efforts are beyond the scope of this work, which focuses on applying a consistent and coherent approach across all species, and not taxon-specific optimization. While such an individual optimization would offer the benefit of more finely tuned maps for each species, it comes at the cost of heterogeneous modelling approaches that hinder comparability. Therefore, the use of a harmonized methodology across over 7,500 species represents a key strength for broad-scale biodiversity analyses. For species-specific conservation planning, however, tailored refinements should be applied, ideally in collaboration with relevant taxonomic and faunistic experts.

Covariate selection and importance

All 19 candidate bioclimatic covariates at the global level were included in at least one model. BIO15 (precipitation seasonality), BIO1 (annual mean temperature), and BIO2 (mean diurnal range) were the three most frequently selected global-level covariates, by 7,452, 7,431, and 7,271 models, respectively. On average, BIO1 (annual mean temperature) was the global-level covariate with the highest relative importance (0.86 ± 0.13), followed by BIO4 (temperature seasonality; 0.73 ± 0.15), and BIO10 (mean temperature of warmest quarter; 0.65 ± 0.17).

Out of the 858 candidate habitat covariates available at the regional level, 645 were selected in at least one model. The global model output covariate was forced in all models and systematically outperformed the others in terms of relative importance (0.98 ± 0.05) (Fig. 3b). Among the other seven categories of regional-level covariates, ‘hydrologic’ covariates were the most often selected relative to the overall number of candidates in this category, whereas ‘topographic’ covariates were the least (Fig. 3b). Aside from the global model output, the ‘edaphic’ category had the highest average importance (0.20 ± 0.10), while the ‘population density’ category the lowest (0.11 ± 0.07) (Fig. 3b).

Averaged at the taxonomic group level, covariate category importance values varied among groups, except for bioclimatic covariates, which consistently exceeded 0.96 (on a scale from 0 to 1). Then, hydrologic covariates had the highest importance for arthropods (EPTO) (0.26), reptiles (0.20), and molluscs (0.19). Edaphic covariates reached their maximum importance values (0.19 to 0.21) for fungi & lichens, ferns & mosses, and molluscs, followed by vascular plants. Other covariate categories, including vegetation, topographic, transportation, and land use/cover, showed intermediate importance across all groups, with peaks of about 0.18 to 0.20 depending on the group. Population density covariates had consistently low importance across all groups, ranging from 0.05 for fishes to 0.14 for ferns & mosses. These patterns likely reflect differences in ecological requirements and habitat preferences among the major taxonomic groups assessed in this study. All detailed covariate selection and importance results are provided in the metadata.zip file that accompanies the maps on Dryad.

Integrity of output layers

Errors during the layer writing process can lead to unwanted artifacts in the output files. The integrity of SDMapCH output layers was checked by using a simple convolutional neural network (CNN) made of two convolutional blocks and two fully connected layers. The CNN was trained on a data set of about 5,000 image tiles (512 × 512 pixels), half of them with artifacts, and tested with an independent set of about 1,400 tiles per category. The classification accuracy on the test set was of 97%. All SDMapCH layers were checked by applying the CNN on 17 tiles sampled in a diagonal line across the layers. Layers with corrupted tiles were flagged, inspected manually, and re-written if necessary.