Coal Fire Identification Based on Multisource Remote Sensing Time-Series Features and Hierarchical Feature Optimization: A Case Study in the Zhundong Coalfield

This is the authors' abstract. We don't add key points for this paper.

Coal fires are combustion phenomena of underground coal seams associated with natural or coal mining activities and have received widespread attention in many countries and regions. Accurately identifying coal fires is crucial for environmental safety and resource management. However, this task is challenging due to complex surface backgrounds, redundant high-dimensional features, and the lack of systematic feature selection strategies. In this study, an integrated coal fire identification framework is proposed for the Zhundong coalfield, China. This framework combines multisource remote sensing time-series feature extraction with a hierarchical feature selection strategy. A comprehensive spatiotemporal feature set was constructed from Landsat-8, Sentinel-1, and DEM data. This set includes multitemporal statistical indicators of tasseled cap transform components, thermal infrared temperatures, and surface deformation features extracted using the coherent enhancement improved phase gradient stacking method. It also includes topographic attributes derived from DEM data. To capture the complex spatial heterogeneity within coal fire areas, a window-based enhancement strategy was implemented, applying local spatial statistical operators across all feature domains. A hierarchical optimization pipeline was used to address feature redundancy. This pipeline first included an initial screening stage that integrated physical–statistical scoring with SHAP attribution analysis. Then, a genetic algorithm–simulated annealing metaheuristic refinement was used to optimize the feature subset. An Optuna-tuned XGBoost classifier was finally employed for automated coal fire identification. Rather than maximizing raw accuracy, the framework prioritizes compactness and interpretability, attaining an F1-score of 0.8974 with only 19 features. This represents a 54.8% reduction in dimensionality relative to the recursive feature elimination with cross-validation-based strategy, with correspondingly lower inference cost, at a modest accuracy cost. The selected subset is derived from repeated independent optimization runs rather than a single search and remains discriminative when transferred to other tree-based classifiers. Overall, the framework achieves effective integration of temporal and spatial features, reduces redundancy through a hierarchical selection and optimization architecture, and provides an interpretable and scalable solution for large-scale coal fire monitoring and automated detection.