• Mapped wildfire susceptibility at finer spatio-temporal scales. • Augmented sparse positive samples using the proposed STS-PSA strategy. • Improved sample imbalance in fine-scale wildfire susceptibility studies. • Achieved substantial performance improvements across various machine learning models. In wildfire susceptibility studies, reliance on sparse satellite-detected hotspots often results in insufficient sample representativeness, severe class imbalance, and information loss. To address this, the present study considers the geomorphological and land-cover heterogeneity across different responsibility areas, as well as the seasonal variation in wildfire-driving conditions, and develops a refined spatio-temporal susceptibility assessment framework that matches regional and seasonal differences in wildfire drivers and hazard effects. Given that such refinement further exacerbates the scarcity of positive samples, we propose a Burned Area Spatio-Temporal Stratification-Based Positive Sample Augmentation (STS-PSA) Strategy that integrates high-confidence fire points with historical burned area data to enhance sample completeness and balance. Using California as a case study, this study demonstrates that the augmented samples show high consistency with original fire points in feature space while improving class separability, and significantly improve predictive performance across Random Forest (RF), Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), and Convolutional Neural Network (CNN) models. Improvements are most pronounced in data-scarce strata: for example, in Summer–LRA, AUC increases from 0.909 to 0.983 for RF, from 0.878 to 0.980 for XGB, from 0.819 to 0.969 for MLP and from 0.727 to 0.939 for CNN. These findings demonstrate the effectiveness and generalizability of the STS-PSA strategy and provide a new pathway for fine-grained wildfire susceptibility modeling under data-limited conditions.
He et al. (Tue,) studied this question.