Machine‑learning‑driven high‑accuracy spatial reconstruction of global wetlands

Using a machine learning model (Random Forest) incorporating anthropogenic features, we reconstructed global wetland distributions under future SSP126, SSP245, SSP370, and SSP585 scenarios based on outputs from seven climate models (ACCESS-ESM1-5, BCC-CSM2-MR, CanESM5, CESM2-WACCM, CMCC-ESM2, CNRM-CM6-1, and GFDL-ESM4). The spatial resolution is 0.25°, with global wetland distribution maps (in GeoTIFF format) generated every five years from 2015 to 2100, and annually from 2086 to 2100. Due to the fact that the training of version v1 did not account for the influence of spatial autocorrelation, the model performance on the test set may have been substantially overestimated. Version v2 therefore adopts a rigorous spatial cross-validation strategy.