Wildfire susceptibility prediction using a multisource and spatiotemporal cooperative approach

Wildfire is one of the natural hazards that poses threats to the safety of forest ecological environment. It is very important to predict wildfire risk in the early stage. Most of the wildfire prediction research based on deep learning networks only extracts features on the spatial dimension. In this work, a deep learning model hybridizing 3D CNN and ConvLSTM (Convolutional Long short Term Memory) was proposed, where the strategies of multisource spatiotemporal cooperative feature fusion are adopted. Some redundant wildfire factors with high correlations by multiple collinear analysis and weight analysis were eliminated. Different from other methods, the daily weather forecast was used as the input of the study region, shortening the time prediction resolution from annual or quarterly to daily to achieve a more accurate prediction in time. Taking the daily ignition in Yunnan Province, China, as the research object, the experimental results showed that the proposed model performs well on the test dataset (AUC = 0.901 and accuracy = 0.912). Seven mainstream machine learning methods were employed for comparison with the proposed model. Ablation and comparison experiments show that the proposed model is a valid alternative tool for wildfire susceptibility prediction.