Integrating remote sensing and machine learning for forest fire susceptibility mapping in Peninsular Malaysia: the spatio-temporal sampling tree (STS-T)

The field of forest fire susceptibility mapping has long been constrained by a persistent trade-off between accuracy and interpretability: black-box ensemble algorithms deliver high accuracy but operate as opaque processes, while transparent decision tree models offer strong interpretability but fail to achieve sufficient accuracy. To fill this gap, this study builds on the forest fire dataset of Peninsular Malaysia spanning 2001–2023. Using recursive feature elimination, we extracted 49 key environmental features from 7349 original attributes and proposed a novel spatiotemporal sampling tree algorithm, STS-T, which has node-level adaptive temporal selection capability and incorporates historical meteorological lag terms into its decision-making process. Validated through stratified 10-fold cross-validation and a hold-out spatial test set from Johor, the pruned STS-T model fitted with a 1-year time window reached an accuracy of 98.48% and an ROC-AUC of 0.988.