A hybrid deep learning model for cross-seasonal forecasting of forest fire occurrence from long-term survey data

Climate change is exacerbating global forest fire regimes, creating an urgent demand for sophisticated predictive modeling tools. Focusing on the climatically sensitive and ecologically critical Daxing'anling Mountains in China, this study develops a hybrid deep learning framework termed CNN-ATT-SYNERGY. This model integrates Convolutional Neural Networks (CNN) with attention mechanisms to fuse multi-source environmental covariates, encompassing meteorological conditions, topographic features, lightning activity, anthropogenic disturbances, socioeconomic factors and historical fire records, thereby enabling robust forest fire occurrence assessment. Trained on a comprehensive dataset involving 3,368 fire incidents across 20 forestry bureaus, the proposed model delivers state-of-the-art performance, with an overall accuracy of 81.22% and an AUC of 87.97%, surpassing conventional benchmark models including Support Vector Machine (SVM) and Random Forest. In addition to its superior predictive capacity, the model effectively identifies dynamic spatiotemporal fire risk patterns: fire risks cluster in eastern regions during spring and autumn, migrate westward in summer, and localized high-risk zones are also detected in winter. Accordingly, the CNN-ATT-SYNERGY framework serves as a seasonally adaptive and scalable decision-support instrument for forest fire management, providing a refined technical solution for boreal forest regions grappling with intensifying wildfire risks.