Deep Learning Modeling of Human Activity Affected Wildfire Risk by Incorporating Structural Features: A Case Study in Eastern China

Wildfire risk prediction is a critical component of disaster prevention and mitigation, often closely associated with local human activities in most regions. Recent studies demonstrate that employing joint modeling techniques using diverse datasets alongside Convolutional Neural Networks-Long Short-Term Memory Networks (CNN-LSTM) produces favorable predictive results. This approach effectively tackles certain drawbacks of fire weather indices (FWI), notably the insufficient consideration of surface coverage and coarse resolution. However, previous research inadequately explored variations in the impact of influencing factors across different categories and spatial orientations, neglecting the internal structural features within the samples. This study focuses on the six eastern provinces of China, utilizing a multi-source dataset comprising satellite-monitored wildfire products from 2012 to 2022, along with terrestrial ecology, terrain, and simulated meteorological elements. By introducing channel and spatial attention mechanisms, high-resolution imagery, and visual transformer model, this research optimizes the CNN-LSTM wildfire prediction model. Results indicate a noteworthy enhancement, elevating accuracy, Kappa coefficient, and AUC of ROC curves from 91.15%, 80.87%, and 97.01% to 93.30%, 85.63%, and 98.15%, respectively. This refined model not only refines high-risk prevention areas highlighted by FWI but also enhances understanding of mountain trails in hilly terrains. Consequently, it reduces false alarms in regions such as non-harvesting agricultural fields, reinforcing predictive risk assessment concerning potential human activities within forested areas. Sensitivity analysis reveals that while the impact of internal sample structural features on wildfire risk prediction is lower than meteorological elements, it surpasses the influence of terrain and terrestrial ecology elements. Thus, this study has developed a methodology integrating multiple attention mechanisms and sample structural features, furnishing high-precision daily kilometer-level wildfire risk prediction products. This approach holds substantial promise for the precise prevention and control of regional wildfires.