Seasonal Wildfire Susceptibility Modelling Using Machine Learning with Multi-Source Satellite and Geospatial Data in Southern Riau, Indonesia

Forest and land fires in southern Riau vary seasonally, yet the benefits of separate seasonal susceptibility models and historical environmental trends remain uncertain. We compared all-season, dry-season, and rainy-season models using satellite and geospatial data for 2010–2024. Logistic Regression, Random Forest, XGBoost, and LightGBM were evaluated under condition-only and condition-change configurations with the area under the precision–recall curve (PR-AUC) as the primary selection metric. Under the original random split, LightGBM achieved PR-AUC values of 0.921, 0.905, and 0.966, respectively. Five-fold 50 km spatial validation reduced corresponding pooled PR-AUC values to 0.761, 0.717, and 0.836 with different leading algorithms in the dry- and rainy-season comparisons. Strict nested ablation indicated small effects of historical trends (+0.0045, +0.0077, and −0.0009 PR-AUC). All-season and dry-season maps were more spatially similar than either was to the rainy-season map, where wind speed, vegetation condition, and peat/lowland context were prominent in model attribution. Independent POLRI fire-presence records placed 77.60%, 76.79%, and 78.57% of matched incidents within High or Very High mapped classes, supporting spatial correspondence but not event-level prediction accuracy. Seasonal modeling therefore identified distinct relative susceptibility patterns; historical trends added modest, period-dependent information, and geographically separated validation gave more conservative estimates of discrimination.