Data explorer

Inspect the species of the dataset: You will see its native range, sampled cells, occurrence records, and model predictions by the Optimized DeepSDM. If you want to see how the choice of threshold influences the model performance or see at which locations the model ist most certain, switch to the thresholded predictions view. Move the slider to change the threshold and see how the predictions change. To keep the explorer lightweight, predictions for all species are shown at 0.25° (about 28 km). Higher-resolution maps can be generated with the codebase.

View

Native range

GBIF records

Evaluation plots

Benchmark setup

Data preparation

Occurrence records are aggregated to a 1 km equal-area grid to reduce spatial clustering. Records falling onto cities are removed. Training records come from GBIF and evaluation plots from sPlotOpen.

Range masking

Training and evaluation data is restricted to the species' native ranges from POWO. Out-of-range locations are ignored rather than counted as absences.

Predictors

Fifty-two environmental predictors are provided on the same 1 km grid.

Group	Features	Source
`chelsa` annual mean temperature, mean diurnal temperature range, isothermality, temperature seasonality, maximum temperature of warmest month, minimum temperature of coldest month, annual temperature range, mean temperature of wettest quarter, mean temperature of driest quarter, mean temperature of warmest quarter, mean temperature of coldest quarter, annual precipitation, precipitation of wettest month, precipitation of driest month, precipitation seasonality, precipitation of wettest quarter, precipitation of driest quarter, precipitation of warmest quarter, precipitation of coldest quarter	19	CHELSA bioclimatic variables
`soilgrids` BDTICM, BLDFIE, CECSOL, CLYPPT, ORCDRC, PHIHOX, SLTPPT, SNDPPT	8	SoilGrids soil properties
`topography` elevation, roughness, TRI, TPI, VRM, aspect cosine, aspect sine, slope, eastness, northness, profile curvature, tangential curvature, dx, dy, dxx, dyy	16	EarthEnv terrain and elevation derivatives
`human_footprint` Human Footprint, built-up areas, croplands, lights, navigable water, pasture, population density, railways, roads	9	Human Footprint Index components
`location` longitude, latitude	2	Optional location encoding, off by default

Predictions shown on the map

The surface and the scored plots above come from the Optimized DeepSDM — the DeepSDM baseline with the aggregation, subsampling, architecture and loss changes applied, the strongest deep model in the leaderboard. Its weights are in the released-models record on Zenodo, linked below.

Get the data

The Zenodo archive contains everything needed to train and evaluate a model. Per-species range polygons and non-aggregated records are available as optional extras.

mkdir -p /path/to/data && cd /path/to/data
wget -c https://zenodo.org/records/21297133/files/sage_benchmark_data.tar?download=1 \
     -O sage_benchmark_data.tar
tar xf sage_benchmark_data.tar

Data record on Zenodo Released models on Zenodo Full data documentation