Decadal Wildfire Risk Projection in Guangxi, China, Using a Two-Stage Deep Learning Framework with Recursive Environmental Factor Forecasting

Multi-year wildfire risk projection requires environmental drivers to be estimated beyond the observational period. This study develops a two-stage framework for Guangxi, China. FactorLSTM recursively estimates nine time-varying environmental factors, and an LSTM–Transformer combines these estimates with four projection-static factors to generate monthly wildfire probability estimates on a 1 km grid for 2019–2028. Observed-driver classification and recursive, predicted-driver evaluation are distinguished. An additional retrospective experiment using a separately trained classifier with a positive-class weight of 10 and Platt calibration yielded a pooled ROC-AUC of 0.7835 and average precision of 0.001475 for available valid observations in 2019–2025. However, its mean predicted probability was approximately 4.50 times the observed positive fraction, indicating residual overestimation. Seasonal maps summarise monthly outputs for January–April, May–September, and October–December. The ten-year outputs are conditional projections under extrapolated environmental drivers and fixed projection-stage land cover. The results identify limitations in probability calibration, rare-event detection, and observation coverage; they do not establish reliable monthly forecasting across the complete ten-year horizon.