Identification and Observation of Spatiotemporal Evolution Features of Coalfield Fire Areas Using Landsat Time Series Decomposition

Globally uncontrolled and persistent coal fires cause substantial resource depletion and long-term environmental and safety risks. However, their long-term detection is hindered by complex subsurface dynamics, combustion behaviour and insufficient data from traditional monitoring methods. This study developed an integrated framework for the dynamic characterisation of fire hazards. The approach combines time-series decomposition, abrupt change analysis, and robust regression, using Landsat-derived land surface temperature (LST) time-series data (2013-2024) for the Miquan coalfield, Xinjiang, China. The LST was first retrieved from Landsat thermal infrared imagery via the Radiative Transfer Equation model. A hybrid Seasonal-Trend decomposition using Loess (STL) and Breaks for Additive Season and Trend (BFAST) model then separated trend, seasonal, and residual components, enabling the detection and quantification of abrupt LST changes related to coal fire dynamics. Thermal trends and anomalies were further quantified using the Mann–Kendall test, Sen’s slope, and a novel Annual Anomaly Deviation (AAD) metric. Crucially, using RANSAC-based regression, persistent potential fire pixels were extracted and categorised into four frequency regimes, with high robustness to noise and data gaps. The resulting coal fire frequency maps achieved over 77% spatial accuracy relative to field records, demonstrating enhanced continuity and reliability in long-term coal fire monitoring. The framework provides a versatile, pixel-level, and analytical paradigm for understanding the spatio-temporal evolution of these challenging landscape phenomena.