Background & Summary

Endangered terrestrial vertebrates face severe threats from land use and land cover (LULC) changes subject to intensified climate change and human activities1,2. In particular, persistent LULC changes have led to significant transformations in over 50% of terrestrial habitats since the early 20th century3,4. To effectively protect these endangered species, it becomes critical to not only document current species distributions but also to understand their spatial dynamics and predict how ongoing and future LULC changes may impact habitat extent and quality5,6, This understanding is essential for effective conservation planning, as habitat degradation and loss can accelerate species extinction processes7. While the International Union for Conservation of Nature (IUCN) Red List as a widely-used dataset8 provides distribution maps of a wide range of species for the historical periods9,10, its range polygons often entail substantial uncertainty and inaccuracies despite their relatively fine spatial delineation11,12 and therefore limits its effectiveness for reliably identifying endangered species’ hotspots at high spatial fidelity. Meanwhile, a fine-scale dataset for the spatial evolution of endangered species’ habitat range under future warming scenarios is still lacking, which may impede national and regional policy-making towards natural habitat planning and protection13.

AOH (Area of Habitat) serve as an effective indicator to reflect the spatial survival pressure of endangered species14. AOH can be generated by combining the LULC types of suitable habitats for endangered species with spatially heterogeneous LULC data15, which can more precisely represent the heterogeneity of natural habitats as compared to the IUCN’s expert range maps. Recently, Lumbierres et al. applied this LULC-based extraction framework that identifies suitable habitat areas by subtracting unsuitable LULC types from IUCN expert range maps, generating gridded habitat maps for terrestrial birds and mammals under historical and current conditions16. Additionally, integrating AOH maps with elevation data has also proven to be an effective method of habitat extraction17. By integrating both habitat preferences and elevation ranges recorded in the IUCN Red List to exclude unsuitable land cover categories and elevations, Brooks et al. further evaluated the applicability of terrestrial habitats for endangered species in the IUCN data10. While these studies have contributed valuable knowledge to understanding the past and current existential pressures on biodiversity, the endangered terrestrial vertebrates that are mostly sensitive to LULC changes are not well covered in the abovementioned built datasets18. Moreover, these datasets do not take into account the spatial heterogeneity of future land cover changes, which largely limits its application to pinpointing the threats of land use changes to endangered species. Based on the above methods, our study further integrates species-specific elevation preferences and applies projected LULC dynamics under future climate scenarios to refine the AOH mapping of endangered species.

Identifying climate and anthropogenic threats to endangered species in the future requires an accurate depiction of the dynamic evolution of species habitats’ heterogeneous distribution under future warming and socioeconomic scenarios. Although existing research has attempted to predict species habitat changes under different future scenarios19, most of these studies used relatively low spatial resolutions (e.g., 0.5°) that are insufficient for accurately capturing species’ survival pressures at a fine scale. This may largely increase the uncertainty of the results since they might not capture the impact of habitat fragmentation on species survival20, especially for endangered species21,22. To address this, we used a comprehensive LULC dataset at 1 km resolution that we recently developed, covering 2020–2100 across multiple warming scenarios. Our approach integrates outputs from the Global Change Assessment Model (GCAM) with the Patch-generating Land Use Simulation (PLUS) model, a patch-level cellular automata framework that incorporates rule-based land expansion mechanisms and spatial neighborhood effects (see the introduction file of the repository)23. This method takes into account the effect of the simulation results of the previous year on the temporally adjacent period and has demonstrated its capability in capturing the heterogeneity of land use changes at the patch level and the robustness in depicting future change dynamics. This has therefore provided a well-established LULC product to be fully matched with suitable habitat types for endangered species. However, we recognize that climate can exert direct pressures on species independent of land use, including physiological stress, phenological disruptions, and range-limiting constraints, etc. These factors are fundamental in defining species’ realized ecological niches and have been emphasized in previous research as critical components for anticipating biodiversity responses to future climate scenarios24,25,26. Therefore, we focused on studying the indirect impact of climate change on species AOH, which was transformed into AOH changes mediated by LULC dynamics27. Nonetheless, future integration of direct climatic stressors and refined species-specific climate suitability information holds great promise for enhancing the ecological fidelity of biodiversity projections under global change.

Another common pattern observed is that species may simultaneously experience both increases and losses in AOH. This is largely due to the spatial heterogeneity of LULC changes across different regions, where AOH is restored or expanded in certain areas. However, existing research primarily focuses on the risks of reduced AOH for endangered species due to agricultural expansion28, urbanization29, and infrastructure development30, demonstrating the vulnerability of biodiversity to habitat degradation and reduction31,32. The future evolution of a species’ AOH often involves both habitat loss and habitat gain due to changing LULC and climate conditions. These opposing trends may offset each other33,34, and therefore, assessments focusing solely on habitat loss may underestimate the overall dynamics and actual risks to species’ suitable habitats35,36, thereby limits the accuracy of assessments of the survival pressures on endangered species under future warming scenarios37,38. Additionally, many existing studies evaluate habitat changes using species distribution models (SDMs) that require extensive input data and species occurrence records39. While SDMs are widely applied at regional scales, their application to global, multi-species AOH mapping can be constrained by data availability and high computational demands, limiting their practicality in large-scale biodiversity assessments. As noted in existing studies, it is essential to map AOH changes into LULC changes to ensure accuracy as well as to enable comparisons across different scenarios and timespan, thus generating a temporal-spatially consistent dataset.

This study aimed to generate a dataset of global 1-km habitat distribution for endangered terrestrial vertebrates and their spatial evolution under diverse warming scenarios from 2020–2100. Specifically, we integrated our recently developed LULC dataset40 covering diverse scenarios of future climate and socioeconomic changes41,42, with the IUCN expert range maps and habitat and elevation preference data43 for endangered species covering 2,571 amphibians, 617 birds, 1,280 mammals, and 1,456 reptiles. We established a global habitat suitability range distribution dataset for endangered terrestrial vertebrates at a 1 km resolution for the baseline year 2020 and future warming scenarios (including SSP126, SSP245, SSP370, and SSP585 scenarios for the year 2030, 2050, 2070, and 2100). The accuracy of our dataset was validated through actual observation point data. The newly built dataset can help pinpoint species’ AOH under future warming and socioeconomic scenarios and provide important implications for endangered species conservation.

Methods

Figure 1 illustrates the workflow for dataset creation, which can be divided into four key stages. The first stage included organizing and analyzing our future multi-scenario, multi-timepoint LULC data (https://doi.org/10.6084/m9.figshare.23542860)40, obtaining expert distribution maps corresponding to the IUCN Red List of endangered terrestrial vertebrates (https://www.iucnredlist.org/resources/spatial-data-download), and collecting elevation data sourced from EarthEnv-DEM90 (http://www.earthenv.org/DEM)44. Supplementary Figures S1, S2 show examples of elevation and LULC data. In the second stage, we focused on determining species’ suitable habitat land use types and elevation preferences based on constraints provided by the IUCN, which could help identify the ecological niches vital for species survival. The third stage integrated habitat and elevation preferences with the collected LULC data to generate AOH distribution maps, which include both the species scale and the category scale under multiple future scenarios. By applying grid-based spatial analysis techniques, we quantitatively assessed how AOH for different species evolve across time under multiple land use and climate scenarios, producing fine-resolution AOH maps that reflect both spatial and temporal dynamics. In the fourth stage, we establish a baseline using the 2020 AOH distribution to calculate the evolution of AOH under different scenarios. This process identified hotspots of future species changes and revealed areas that are predicted to undergo significant changes in AOH.

Fig. 1
figure 1

Overall methodological framework for generating the AOH maps.

Full size image

Moreover, we designed a three-step validation framework to ensure the dataset’s reliability. This process utilizes real observation point data to assess the dataset’s accuracy. Details of this validation methodology are described in the technical validation section. The dataset production process is complex, involving extensive interactions among large volumes of data. Each stage of the workflow creates substantial data outputs that should be meticulously managed, analyzed, and integrated to ensure consistency in the final output data. To streamline analysis and improve computational efficiency, we used Python to efficiently process large raster datasets and vector data, supported by key packages such as GeoPandas, Rasterio, GDAL, and so on. Moreover, we implemented parallel processing techniques to optimize analysis performance.

Data collection and processing

A comprehensive set of base geographic data was assembled, including our recently developed global 1-km resolution LULC dataset covering 235 watersheds. This dataset integrates future climate change, socioeconomic dynamics, and the influence of prior-year simulation results on adjacent periods. By incorporating climate and socioeconomic factors from future scenarios, we differentiated LULC suitability probabilities between historical and future periods. Using an enhanced cellular automata model, PLUS, we simulated patch-level changes of various land cover types at the watershed scale. The dataset demonstrates high simulation accuracy (Kappa = 0.94, OA = 0.97, FoM = 0.10), Furthermore, to provide a more balanced evaluation of classification performance, we calculated class-wise True Skill Statistic (TSS) values based on pixel-wise comparison between the real and simulated LULC maps45. The resulting macro-average TSS was 0.85, the detailed TSS values for each class are provided in Table S5. This dataset provides a reasonable approximation of the spatiotemporal dynamics of global LULC changes influenced by climate and socioeconomic drivers, which can support habitat suitability refinement efforts for endangered species.

Additionally, high-resolution digital elevation model data from the EarthEnv-DEM9044 provided important information on topographical features (see Supplementary Figure S1). We also compiled an inventory of endangered terrestrial vertebrates, which includes spatial distribution maps and informative details on suitable habitats. More detailed distribution maps for birds are provided in the work by Jetz et al.46, which applies a widely recognized modeling framework for global avian distributions. In contrast, the distribution maps for amphibians, mammals, and reptiles were obtained from the IUCN, accessible via their official website at http://www.iucnredlist.org. Through these sources, we acquired expert range maps for a total of 2,571 amphibians, 617 birds, 1,280 mammals, and 1,456 reptiles with their preference descriptions. Notably, all these species are classified under the IUCN’s vulnerable, endangered, or critically endangered categories, which emphasize the urgency of conservation efforts.

The IUCN Red List is a valuable resource that provides approximate spatial distributions and habitat and elevation preferences for each endangered species. Specifically, we extracted detailed textual habitat and range descriptions from the IUCN database for each species and parsed them using a rule-based model that identifies keywords, habitat qualifiers, and geographic constraints. These preferences are categorized according to the IUCN’s habitat classification scheme, as detailed from https://www.iucnredlist.org/resources/habitat-classification-scheme. However, it is noted that there is a gap between habitat preferences and actual LULC dataset classifications, which poses challenges for practical applications in conservation planning. To address this limitation, we built upon the translation table recently published by Lumbierres et al.47, and translating their LULC classification into the classification system used in our LULC dataset48. This allowed us to align IUCN habitat categories with our LULC types in a consistent and reproducible manner (Supplementary Table S3). This improvement allowed for a more accurate mapping of species habitat preferences to actual land cover types and deepened our understanding of the AOH for various endangered species.

Mapping of habitat suitable range distribution and change

The process of generating AOH distribution maps utilized a two-step filtering approach based on expert range maps of endangered species, by integrating LULC and digital elevation model data. This approach aimed to refine the extraction of suitable habitats for specific endangered species, thereby providing a robust framework for conservation and management efforts. We started with the expert range maps for each species, which are preliminary estimations of potential AOH based on expert knowledge and historical data. While these maps serve as a valuable starting point, they often contain unsuitable areas that do not meet the ecological requirements of the species. Therefore, further filtering is essential to improve the accuracy of the AOH extraction.

In the first step, we overlaid the expert range maps with LULC data to exclude areas that did not match the habitat preferences of the species. This overlay analysis allowed us to identify and exclude land use types that are considered unsuitable for certain species. For instance, natural areas are preferred while urbanized areas or agricultural lands are ignored because they may not provide the necessary resources or conditions for some species. By establishing specific filtering criteria based on LULC requirements, we retained only those areas that matched the habitat preferences of each species. The second step was to further refine the filtering criteria by integrating the DEM data. Species often have distinct elevation preferences that play a crucial role in the AOH selection for many species. We analyzed the elevation preferences specific to each species and applied these insights to the preliminary suitable habitat areas. By overlaying the DEM data, we excluded areas where elevation did not correspond to the requirements of the species, ultimately producing the AOH distribution for specific species.

To illustrate the entire process, we used Dyacopterus brooksi as a case study to demonstrate the transformation of the suitable habitat range before and after the filtering steps. This species primarily inhabits tropical rainforests and secondary forests in Indonesia, and is classified as endangered due to habitat loss, deforestation, and human disturbances. Figure 2 visually depicts how the integration of LULC and DEM data systematically removes unsuitable areas from the broader expert range map of Dyacopterus brooksi, resulting in a more refined AOH. This approach refines the spatial representation of species’ AOH by narrowing expert-derived range maps to more ecologically plausible habitat areas, based on land use and elevation constraints. While we acknowledge inherent limitations due to data availability, we provide validation analyses in the technical validation section to assess the reliability of our dataset. Through this two-step filtering process, we successfully integrated expert knowledge with geographic information data, enabling the AOH distribution maps to more accurately reflect the actual suitability of habitats for each species within their ecological context.

Fig. 2
figure 2

AOH extraction process using Dyacopterus brooksi (a large fruit bat primarily distributed in Southeast Asian forests). This species plays a crucial role in the ecosystem by dispersing seeds through fruit consumption, thus maintaining forest health and diversity. However, it may face survival pressure due to habitat loss and human activities) as an example: (a) IUCN expert range map; (b) elevation distribution; (c) spatial distribution of LULC types; (d) spatial distribution of the extracted AOH after two-step filtering.

Full size image

The analysis of AOH changes over time is critical for understanding the dynamic evolution of endangered species habitats in response to environmental impacts. To achieve this, we developed a systematic approach for identifying AOH change areas by subtracting the baseline AOH distribution from the projected future AOH distribution. The baseline AOH distribution was established for the year 2020, providing a reference point for subsequent analyses. First, we defined the baseline AOH distribution based on the previously generated maps. This baseline serves as the foundational snapshot of the suitable habitats available to species in the year 2020. Second, following the establishment of the baseline AOH distribution, we then projected future AOH distributions under various environmental scenarios. This included modeling potential changes in AOH based on projected land use changes. The future AOH distributions were generated for a specific target year, which allow us to visualize how habitat suitability might evolve. Finally, after establishing the baseline and future AOH distributions, we employed a GIS-based approach to perform a spatial subtraction of the baseline AOH distribution from the spatial distribution under different SSP-based land use and climate pathways. These progressive procedures effectively highlighted the changes in habitat suitability by identifying areas where habitats are expected to be lost and regions where AOH is projected to expand.

Through this subtraction method, we categorized the changes into three distinct types:

  1. (1)

    Loss Areas: Areas of future AOH distribution loss as compared to baseline.

  2. (2)

    Increased Areas: Areas of future AOH distribution increase as compared to baseline.

  3. (3)

    Unchanged Areas: Areas where habitat suitability remains unchanged between the baseline and future distributions.

The results of the AOH evolution analysis were visually represented in Fig. 3, which illustrates the spatial dynamics of AOH changes for selected endangered species. The figure provides a comprehensive overview of the areas experiencing habitat loss, increase, and unchanged, thereby facilitating a clearer understanding of the future challenges and opportunities for conservation efforts. By identifying and visualizing AOH change areas, this approach provides valuable insights into the evolving spatial dynamics of endangered species habitats, which helps to identify areas that need to be prioritized. Furthermore, understanding these spatiotemporal habitat changes contributes to a broader knowledge base regarding species’ potential responses and resilience in the face of ongoing LULC dynamics.

Fig. 3
figure 3

The extraction process of AOH change areas using Dyacopterus brooksi as an example: (a) AOH spatial distribution in 2020 as the baseline; (b) AOH spatial distribution in 2100 under the SSP585 scenario; (c) spatial distribution of AOH change areas, including loss, unchanged, and increased areas.

Full size image

Projections of biodiversity maps and their changes

We used raster summation to measure species diversity at the raster scale after mapping AOH for all species globally, as shown in Fig. 4(a–d). This analysis is crucial for understanding the ecological richness and the potential vulnerabilities of different areas to habitat changes. By aggregating individual AOH rasters, we generated species richness maps for amphibians, birds, mammals, and reptiles under four representative future scenarios: SSP126, SSP245, SSP370, and SSP585. These scenarios cover a range of possible futures, reflecting different socio-economic trajectories of LULC change and their corresponding impacts on biodiversity. The AOH rasters were calculated for different time points: 2030, 2050, 2070, and 2100. The resulting species diversity rasters thus provide a comprehensive snapshot of biodiversity for each scenario and time point. These results are detailed in the introduction file of the repository, which illustrate variations in species diversity across both spatial and temporal dimensions.

Fig. 4
figure 4

Gobal species diversity maps (a–d) and global AOH change maps (e–h) based on the 2020 baseline, using the 2050 results from the SSP245 scenario as an example.

Full size image

To further analyze the dynamics of habitat suitability changes, we overlaid the individual AOH change maps for each species, as shown in Fig. 4(e–h). This step enabled us to produce global AOH change maps, which are further illustrated in the introduction file of the repository. These change maps highlight areas of species loss and regions where AOH is projected to expand at the category level. By visualizing these changes, we can identify hotspots where conservation efforts are urgently needed due to significant AOH losses. Our analysis revealed an important correlation: regions characterized by higher species diversity tended to experience more pronounced changes in AOH. This finding suggests that areas supporting a greater variety of species are more sensitive to AOH alterations, making them potential critical points for conservation interventions.

Data Records

The data (including tables and maps) are stored in the Dryad Open Access Repository and are publicly available at https://doi.org/10.6084/m9.figshare.2798553849. All the spatial data are stored in raster GeoTIFF format and tabular data are stored in csv format. The dataset is categorized into two main types: species scale and category scale. Initially, the data were classified into AOH maps and future change maps (folders named species scale AOH, species scale change, category scale AOH, and category scale change). Within the species scale dataset, raster values are binary [0, 1], with 1 indicating the presence of AOH. The data are organized into three levels on the basis of development pathway, year, and species category. The name formats for AOH maps is: AOH/SSPxxx/yyyy/c/s.tif, while the name formats for AOH change maps is: Change/SSPxxx/yyyy/c/s.tif. A total of 2,571 amphibian, 617 bird, 1280 mammal, and 1456 reptile species are included in the AOH and change data. In the category scale dataset, the data are organized into two levels based on the diverse warming scenarios and years (including the SSP126, SSP245, SSP370, and SSP585 scenarios for the year 2030, 2050, 2070, and 2100). The name format for species diversity maps is: AOH/xxx/yyyyy/c/xxx_yyyyy_c.tif, and the name format for species diversity change maps is: Change/xxx/yyyyy/c/xxx_yyyyy_c.tif. Species diversity was calculated by overlaying each species’ AOH map, with overlapping areas summed50. For the change maps, raster values indicate the magnitude of species diversity change in a given area. Positive values indicate regions with increasing species diversity, with higher values indicating greater increases, whereas negative values reflect decreasing richness, and zero values denote no change in species diversity.

In the file name, xxx represents SSP scenarios; yyyy represents the project year; c represents species category; and s represents the species name. Additionally, we have included a separate folder containing our validation data and results. We also compiled statistics on the AOH change areas for each species category under different scenarios and time points (Supplementary Table S2), which were also summarized at the continental scale (Supplementary Table S3). Furthermore, we calculated the percentage of AOH loss for each endangered species at the species scale for future scenarios (Supplementary Table S1).

To account for potential future shifts in species ranges beyond current IUCN boundaries, we additionally provided potential AOH maps that assume ± 10% range expansion and contraction. These data are organized in separate folders with consistent naming conventions, and they are intended to support sensitivity analyses or species-specific customization.

Technical Validation

Reliability of input data

At the first step, the reliability of the data sources is checked, particularly the accuracy of the IUCN expert range maps data51,52. Therefore, we relied on independently observed species occurrence data to assess their reliability and confirm their credibility. Given that the observed occurrence data are only available for the present time, we validated the overall framework and results based on the data for the year 2020. In this context, the reliability of the dataset under future scenarios is influenced not only by the resolution and accuracy of LULC simulations but also by the limitations in accounting for species-specific direct climate sensitivities. Our current framework incorporates climate change indirectly through LULC projections that are themselves driven by future climate scenarios. The LULC datasets used in this study reflect anticipated changes in climate variables (e.g., temperature and rainfall) as modeled under various SSP pathways, which enables partial integration of climate effects into habitat projections. This enables us to regard the dataset as a “dynamic map” that can be iteratively updated as more refined LULC models and species-specific climate response data become available.

Specifically, we obtained species occurrence records (Fig. 5) for all terrestrial endangered species (as identified by the IUCN Red List) observed between 2019 and 2024 from the Global Biodiversity Information Facility (GBIF) and eBird (https://ebird.org/)53. To ensure the accuracy of the independently observed species occurrence data, we used only high-precision records and applied multiple filtering criteria. For birds, we set a coordinate uncertainty threshold of fewer than 30 meters, whereas for other species categories, we retained only records with coordinate uncertainties below 300 meters16,54. Through these stringent filters, we successfully obtained a large set of species occurrence data, including 112 amphibians (4.4% of 2,571), 290 birds (47.0% of 617), 259 mammals (20.2% of 1,280), and 220 reptiles (15.1% of 1,456). Furthermore, we excluded occurrence points outside the IUCN expert range maps, as such mismatches may result from various sources, including species misidentification, migration events, taxonomic inconsistencies, or inaccuracies and incompleteness in the IUCN range data itself54. To reduce the influence of species with very large observation counts, we implemented a subsampling strategy: for species with more than 20 valid points within the IUCN range, we randomly selected 20 points to ensure comparability across taxa and minimize sampling bias. As a result, we obtained species occurrence data for 69 amphibians (61.6% of 112), 222 birds (76.6% of 290), 177 mammals (68.3% of 220), and 133 reptiles (60.5% of 220). This directly supports the reliability of the IUCN expert range maps, which underpinned the implementation of our methodology and technical validation. These statistics are summarized in the repository.

Fig. 5
figure 5

The spatial distribution of observed species occurrence data. The subplots visualize the observed species occurrence data, IUCN expert range maps, and the AOH distribution we extracted for different species: (a) Amphibians (e.g., Rana iberica); (b) Birds (e.g., Buceros bicornis); (c) Mammals (e.g., Leopardus tigrinus); (d) Reptiles (e.g., Carinascincus greeni).

Full size image

Accuracy of dataset

Previous AOH maps were validated using only species observation points and occurrence polygons55,56. However, this validation approach may fall short in capturing the spatial heterogeneity of habitat suitability within the AOH extent. A more robust validation should assess whether species occurrence is significantly concentrated within predicted suitable areas, rather than being randomly or uniformly distributed across the expert range. Therefore, we used a three-step validation framework to ensure the reliability of the dataset. The first stage involved the use of the method developed by Dahal et al. to evaluate whether the distribution patterns in the AOH were significantly better than random distribution patterns54. We compared the proportion of observed points falling within the AOH with the proportion of the AOH area relative to the expert range map area. Additionally, we assessed the probability of actual observation point data falling within species diversity maps holistically, thereby validating the entire dataset. Considering the human annotation of the textual information, we also conducted comparative experiments to identify systematic errors in the AOH map creation process10,57. The spatial example is shown in Fig. 5:

Species observation point validation

We used species observation data to validate the AOH maps for each species. Given our 1-kilometer grid resolution, ensuring that individual species points fall precisely within the AOH map is crucial for validating the distribution pattern comparison. This enabled us to perform a quantitative comparison between the AOH distribution and random distribution patterns. As a result, 69 amphibians, 222 birds, 177 mammals, and 133 reptiles were included in the comparative validation process. By comparing the proportion of observation points within the AOH to the proportion of the AOH area to the expert range map area, we determine whether the AOH exceeded the random distribution. If the proportion of observation points within the AOH exceeded the AOH area to the expert range map area, the reliability of our AOH identification is greater than that of a random distribution. We found that 94% of amphibians, 94% of birds, 95% of mammals, and 91% of reptiles presented distribution patterns within our AOH maps that outperformed the random distribution (Supplementary Table S4). On average, the proportion of observations within the AOH exceeded the expected random proportion by 34.54% for amphibians, 38.53% for birds, 22.87% for mammals, and 33.03% for reptiles. These consistent positive deviations confirm the effectiveness of AOH in spatially refining species’ likely presence areas.

Dataset-wide validation

We introduce dataset-wide validation to ensure robustness and representativeness. Specifically, we overlaid the occurrence points onto the corresponding species diversity maps (i.e., raster layers showing the total number of overlapping AOH per pixel) and used spatial extraction (i.e., the “extract values to points” method) to determine whether the raster value at each point was greater than zero. This method allowed us to evaluate the spatial pattern accuracy of the dataset comprehensively, especially in cases where point data were sparse or unevenly distributed58. The spatial database of the validation points has also been uploaded to the repository for transparency and reuse. Although it is unrealistic to achieve 100% agreement due to inherent uncertainties in the occurrence point data, such as potential geolocation errors, temporal inconsistencies, and observation biases, a high proportion of validation points were found within areas of non-zero species diversity. Validation results for the entire dataset indicated that 2,384 amphibian species points (86.53%), 85,718 bird species points (81.58%), 63,751 mammal species points (87.81%), and 3,582 reptile species points (84.84%) that fell within their respective species diversity maps. These results provide strong evidence supporting the spatial reliability of our species diversity maps.

We also conducted an analysis to investigate the mismatch of observation points that did not fall within the AOH map. Extracting the corresponding LULC types from these mismatched points (see Supplementary Table S3), it was found that the occurrence points not falling within the AOH map for reptiles, mammals, birds, and amphibians were distributed across these LULC categories, with the total proportion in cropland, urban areas, barren land, and water bodies reaching 0.53, 0.55, 0.62, and 0.44, respectively. This supports our understanding that real-world observation data is inherently affected by observation errors. After accounting for the observational errors and land-use misclassifications, we recalculated the AOH precision for the dataset. The accuracy for amphibians increased to 93.26%, for birds to 94.47%, for mammals to 93.20%, and for reptiles to 92.26%. In addition, we also compared our species richness maps with those of Lumbierres et al. and found that their dataset contained 60,451 mammal species points (83.27%) that fell within their species diversity map, providing a valuable benchmark for assessing the spatial consistency between the two datasets. Furthermore, although differences in resolution and species scope (e.g., the near-threatened species), the spatial correlation exceeded 0.75, indicating strong consistency.

Comparative experiment for systematic error identification

To identify systematic errors in the annotations of key habitat preference and elevation preference, we applied a keyword filtering algorithm. This algorithm extracts potentially relevant information from IUCN habitat preference descriptions (e.g., utilizing terms such as [‘forest’, ‘between’, ‘occurs’, ‘elevation’]). Five characters before and after the keyword were selected, and the result may be null. A random sample of the filtered data was compared with manually annotated preference data to pinpoint systematic errors in the annotation process. In addition to these validations, we also presented the maps to regional experts for feedback, During the filtering and validation process, we identified a small number of cases (<2% of species) where no extractable key habitat or elevation information could be found in the IUCN species descriptions. These were defined as systematic errors resulting from the limitations of our data-driven rule extraction step. Given their low proportion, we considered these omissions to fall within an acceptable margin of uncertainty.

Finally, using the actual land use data from 2020 and future simulations, we generated a corresponding dataset and conducted a three-step validation process, by taking 2020 as the baseline. These validation steps ensured that our framework and dataset accurately and reliably reflect future information regarding species AOH distributions and changes.

Usage Notes

This study produced a global dataset of endangered species habitats distributions and their spatial evolution under different climate warming scenarios from 2020-2010 at a 1 km resolution, which includes a baseline map for 2020 and projections for 2030, 2050, 2070, and 2100 under the SSP126, SSP245, SSP370, and SSP585 scenarios. The dataset offers valuable spatial insights at the species level, with validation results showing that our AOH maps provide a more accurate representation of endangered species’ habitats as compared to the random distribution patterns of the IUCN expert range maps. Our work not only provides a high-resolution geospatial data product for understanding habitat shifts and biodiversity conservation, but also fills the gap in high-resolution habitat change predictions for endangered species on a global scale.

Meanwhile, there also exist some limitations with our data product. The spatial precision of the LULC dataset is the main factor affecting the accuracy, as lower resolutions may suffer from implicit spatial details. In this study, we applied our newly developed 1-km LULC dataset40, which effectively captures current and future LULC changes under warming scenarios. Further improvements, such as the incorporation of higher-resolution LULC data (e.g., the 300-m ESA-CCI LULC dataset), can further improve accuracy and usability of our dataset59. In this work, we translated the projected impacts of climate change on species into land-use-driven habitat suitability changes, enabling a spatially explicit framework. While this approach offers practicality and broad applicability, it currently assumes static species distribution boundaries as defined by IUCN. However, to account for potential future shifts, we provide both 10% potential expansion and contraction buffers for each species range, as well as open-source code that allows users to customize species range assumptions based on updated data or expert knowledge. In future works, our goal is to further develop the dataset by incorporating additional climate change scenarios and more refined spatial patterns, while maintaining the established criteria for delineating endangered species habitats. By offering a dynamic and scalable tool for addressing the ongoing threats posed by climate change to endangered species and ecosystems, our dataset is expected to provide robust support for preserving endangered species and advancing research on human-environment interactions.