A spatially validated explainable framework for large-scale wildfire susceptibility mapping using deep sequence learning and gradient-boosting ensembles

Wildfire susceptibility mapping requires predictive frameworks that are accurate, spatially transferable, interpretable, and computationally efficient. This study developed a arge-scale explainable modeling framework for Madhya Pradesh, India. Fifteen topographic, climatic, surface-energy, biophysical, hydrological, land-cover, and anthropogenic conditioning factors were harmonized to a 1-km grid. High-confidence MODIS active-fire observations were clustered using spatio-temporal DBSCAN, and vegetation-constrained pseudo-absences were generated using a 1-km exclusion distance. Six models were evaluated: XGBoost, LightGBM, Gradient Boosting, a stacked ensemble, a Temporal Convolutional Network (TCN), and TCN–BiLSTM. The neural models operated on ordered predictor sequences rather than true chronological time series. Performance was assessed using an independent 70:30 holdout split, bootstrap confidence intervals, repeated random splits, and five-fold 50-km spatial block cross-validation. The stacked ensemble achieved the highest observed held-out ROC–AUC of 0.800 (95% CI: 0.785–0.814), followed closely by XGBoost (0.795) and LightGBM (0.790), whereas TCN and TCN–BiLSTM reached 0.773 and 0.674, respectively. Spatial validation reduced mean AUC to 0.735–0.752. Susceptibility classes derived from training probabilities were independently supported by increasing held-out fire rates from 11.4% in the lowest class to 82.2% in the highest. Explainability analyses consistently identified NDVI and hydroclimatic–energy variables as major predictive contributors. Pseudo-absence distance materially affected discrimination, whereas predictor decorrelation caused only a small AUC reduction. These findings highlight the importance of geographic independence in model evaluation and show that increased architectural complexity does not necessarily improve discrimination. The framework provides a transparent and computationally practical basis for regional wildfire susceptibility prioritization, while not representing real-time forecasting, fire spread, or future climate projections.