A Similarity-Constrained Sampling Framework for Early-Stage Forest Fire Susceptibility Mapping in Sample-Scarce Regions

Abstract

Forest fire events in coastal eastern Guangdong are strongly associated with human activities and often occur as small-scale, short-lived early-stage fire events. Given uneven thermal monitoring infrastructure, a key challenge is leveraging data-rich local areas to identify high-susceptibility areas in data-sparse regions. Focusing on Shanwei City, this study used fire data from thermal imaging in the Shenzhen–Shanwei Special Cooperation Zone. By integrating biophysical covariates and anthropogenic factors, we developed a Randomized Iterative Sampling–Similarity Assessment–Model Selection (RISAMS) framework to optimize negative-sample selection and identify a robust classifier. Shapley Additive exPlanations (SHAP) was implemented in Python 3.13.5 using the SHAP package (version 0.48.0) and was used to interpret drivers of susceptibility. In the optimal iteration, mean covariate similarity between training samples and the Shanwei macro-environmental background reached 0.94, with a minimum of 0.72. The optimal Random Forest classifier achieved an AUC of 0.92 on the training set and 0.91 on the test set, with Cohen’s Kappa of 0.78 for both. High-susceptibility zones clustered along forest edges and wildland–urban interfaces in northern and central Shanwei. SHAP identified road density, elevation, and slope as the primary drivers, with a cumulative relative importance of 61.87%. This framework supports early-stage forest fire susceptibility mapping under uneven monitoring resources and improves early warning of early-stage fires in human-dominated humid coastal forests.