Advancing coal fire detection model for large-scale areas based on RS indices and machine learning

Abstract

Coal fires present significant global environmental and energy challenges, posing substantial barriers to achieving carbon-neutral goals. Thermal Infrared Remote Sensing (TIRS) technology, which is used to retrieve land surface temperatures, plays a crucial role in detecting coal fires. However, its accuracy suffers from solar radiation interference. In addition, there is limited research focused specifically on detecting coal fires over large areas. In this paper, thermal anomaly indices (TAIs), derived from short-wave infrared and near-infrared data, were selected for coal fire detection due to their relatively low sensitivity to solar radiation. Using these TAIs alongside other remote sensing (RS) indices, a coal fire detection model (CFDM) was developed and trained using the AutoGluon machine learning (ML) framework. The model is capable of identifying large-scale coal fire target areas without relying on deformation associated with coal fires. CFDM outperformed other ML algorithms, achieving Recall, Precision, F1-score, and Kappa coefficient values of 0.89, 0.94, 0.93, and 0.92, respectively. Shapley Additive Explanations (SHAP) were used to evaluate the importance of different features, validating the model’s reliability and interoperability. The model’s robustness has been further demonstrated using observed coal fire points over Xinjiang, China, and Jharkhand, India. A T-test confirms that the proposed CFDM is significantly superior to TAIs-based methods, offering better differentiation of coal fires from other thermal anomalies and reducing commission errors.