Daily Wildfire Risk Prediction by Mining Global and local spatio-temporal dependency

With the influence of climate change, the trend of wildfires is becoming increasingly frequent, resulting in many adverse effects during major wildfires. Currently, in the field of wildfire prediction, basic machine learning models are generally used to explore latent features only in spatial or temporal dimensions, without analyzing the combined effects of time and space on wildfire prediction from both aspects of the dataset and prediction model. To address these issues, we choose Yunnan Province, China, with its diverse climate and topography, as the study area. We expand the adjacent temporal domain to create spatio-temporal datasets based on spatial datasets and propose a daily wildfire risk prediction model by mining global and local spatio-temporal dependency (GLSTD). Specifically, we construct temporal, spatial and spatio-temporal datasets to verify the effectiveness of the model. Unlike conventional Pearson correlation analysis of continuous variable correlations, Spearman correlation analysis, which can analyze both discrete variables and continuous variables, is used to analyze the redundancy of wildfire factors. We embed multi-head self-attention blocks into the spatio-temporal feature extraction module to explore the global and local spatio-temporal correlation features between the ignitions and the related factors, effectively predicting wildfire risk daily. The results of ablation study and comparison experiment show that GLSTD can effectively explore the global and local spatio-temporal correlation features between the ignitions and the related factors, and outperform random forests, especially outperforming the most commonly used convolutional neural network models by 10% points in accuracy. In addition, through Spearman correlation analysis, it is found that the maximum temperature, average temperature, minimum temperature in Yunnan Province, China, are highly correlated with each other. SHapley Additive exPlanation (SHAP) importance analysis shows the relative humidity, elevation, and precipitation variables have a significant impact on model prediction, providing valuable and practical information for local wildfire prevention in the study area. Source code has been available at https://github.com/Bin447/Wildfire-Predict.git.