Supplementary data for "Benchmarking forest fire occurrence prediction for real-world early warning: difficult negative sampling and spatiotemporal adaptive decision boundaries"

This record provides the processed source data supporting the main tables and figures reported in the manuscript “Benchmarking forest fire occurrence prediction for real-world early warning: difficult negative sampling and spatiotemporal adaptive decision boundaries”.

The dataset includes processed data for sample composition, model performance, ablation results, monthly evaluation metrics, and figure source data. These files support Tables 2–5 and Figs. 2–5 in the manuscript.

This record does not redistribute third-party raw datasets, full intermediate sample-level feature tables, trained model files, or gridded prediction products. Raw fire occurrence, meteorological, vegetation, terrain, land-cover, hydrological, road-network, population, and socioeconomic datasets should be obtained from their original providers under their respective licences and usage conditions.

The associated software repository used for difficult negative sample construction, feature processing, model training, evaluation, and figure/table generation is provided separately through the journal submission system.