A Novel Feature Extraction Method of Environmental Factors for Forest Fire Risk Modeling Based on Adaptive Time Window

A feature set that can fully reflect information regarding the cumulative dryness state (CDS) of forest fuels is crucial in forest fire risk modeling. Due to the uneven spatial and temporal distribution of rainfall, the CDS information often exhibits significant spatial heterogeneity. Current feature extraction methods for environmental factors based on fixed time windows struggle to capture this spatial heterogeneous information accurately. This paper proposes an adaptive time window-based method for extracting forest environmental factors features. By using precipitation as a constraint, this method adaptively constructs dynamic time windows for each pixel, thereby obtaining finer CDS information. The random forest (RF) and support vector machine (SVM) algorithms were used to construct the fire risk models, and both showed improvements in overall accuracy, indicating the effectiveness of the proposed method. The improvement performance of the RF model was better than that of the SVM model, and the overall accuracy can be improved by 5% to 8% under appropriate precipitation constraint settings.