Evaluating the Contribution of Temporal and Meteorological Features to Daily Hotspot Prediction Using XGBoost with SHAP-Based Interpretation

Forest and land fires remain a recurring environmental problem in Indonesia, highlighting the need for reliable approaches to hotspot prediction. This study evaluates the contribution of temporal and meteorological features to daily hotspot prediction in Riau Province using XGBoost, with SHAP employed to interpret feature contributions to model predictions. VIIRS hotspot data from NASA FIRMS and daily meteorological data from BMKG covering 2013–2020 were used. Two modeling scenarios were compared: a meteorological-only model and a model incorporating temporal features, including hotspot lags, rolling meteorological averages, and seasonal information. The meteorological-only model achieved an RMSE of 655.23 and an R² of 0.279, whereas the inclusion of temporal features reduced RMSE to 227.61 and increased R² to 0.724. This corresponds to an approximately 65% reduction in RMSE, indicating a substantial improvement in predictive performance. SHAP analysis identified the previous day’s hotspot count (hotspot_lag1) as the feature with the largest contribution to model predictions, followed by other temporal and meteorological variables. These findings demonstrate the predictive value of incorporating recent hotspot history alongside meteorological conditions and provide a basis for further development of data-driven hotspot monitoring approaches.