Predictive modeling study demonstrates accurate forest fire risk forecasting by coupling fuel and climate dynamics, indicating improved early warning for catastrophic wildfires.
Key Points
To develop and validate a physically aligned deep learning framework, FWI-MSNet, that couples fuel and climate interactions to improve forest fire risk forecasting across multiple temporal scales.
Extracted 21 physically relevant indicators from 18 years of synchronized observation data at the Huitong Ecological Station in China, incorporating Forest Fire Weather Index (FWI) metrics.
Engineered the FWI-MSNet architecture using parallel multiscale 1D-CNNs for daily-to-seasonal dynamic feature extraction, coupled with Gated Recurrent Units and Transformer networks for temporal and global integration.
Evaluated performance against seven baseline models (including XGBoost, LSTM, CNN, and four ablation variants) and validated trajectory tracking on the 2020 Australian wildfires.
Decreased the composite mean of prediction errors (RMSE, MAE, and MAPE) by 56.1% relative to baseline models.
Increased the coefficient of determination (R²) to 0.9251 compared to a baseline model average of 0.223.
Successfully reproduced the risk evolution trajectory during the 2020 catastrophic forest fires in Australia.