LSTM neural network integrating remote sensing and meteorological data reveals forest fire risk trajectories in China over the past 100 years

Quantifying current forest fire risk is essential to ecological security and sustainable resource management, while reconstructing historical fire risk helps understand the impacts of climate change and human activities on forest fire regimes, enabling effective lessons from past forest management practices. However, the lack of long-term forest fire risk assessment studies (e.g., 1921 to 1960) is due to insufficient reliable spatio-temporally explicit forest fire records, stemming from the unavailability of spatial information technology for modeling fire occurrences. This study presented a retrospective fire risk modelling approach using deep learning, meteorological station observations, and spatial interpolation to overcome this limitation. By modifying forest fire danger index (MFFDI) and integrating it with a multi-step prediction Long Short-Term Memory model, historical station-wise MFFDI data for 1921–1960 was generated, leading to the creation of a forest fire risk distribution map via spatial interpolation that accounts for terrain and human influences. The results indicated that the model-predicted MFFDI for 1921 to 1960 demonstrated high accuracy, with a mean R² of 0.77 and a mean MSE of 0.24 from an independent validation subset of the 778 meteorological stations across China. For the independent validation of the model-generated forest fire risk maps, 193 of the 265 historical fire events occurred in high or relatively high risk zones, and 72 in moderate risk zones. Spatial analysis revealed that over the past 100 years, high and relatively high risk areas were primarily located in southwest, northwest, and north China, comprising 19 % to 26 % of the total forest fire risk area. After 1980, high and relatively high risk zones gradually concentrated in the central and southwestern regions, while moderate risk zones shifted from the southwest to the southeast. The area of high risk zones was constantly stable, below 8 % over time. This study reconstructs historical fire risk maps, highlighting the century-long dynamics of forest fire risk in China, which is essential for formulating scientific fire control strategies.