Wildfire detection near transmission lines using geostationary meteorological satellite data

Wildfires pose a severe threat to both ecological systems and critical infrastructure, particularly power transmission lines traversing high-risk areas such as forests and grasslands. We focus on wildfire detection near transmission lines in Yunnan Province, China, using geostationary meteorological satellite data from Himawari-8 and Geo-Kompsat-2A (GK2A). A machine learning-based approach is proposed that integrates spectral, spatial, and temporal features to improve detection accuracy. First, a spatial-contextual fire detection method tailored for the study area is developed, achieving an F1-Score of 0.59 for Himawari-8 data and 0.64 for GK2A data. Second, three feature input strategies are designed to exploit the high temporal resolution of geostationary satellites. On the independent 157-event evaluation set, the eXtreme Gradient Boosting (XGBoost) model with fused spectral-spatial-temporal features using GK2A data achieves a Precision of 0.80, a Recall of 0.74, and an F1-Score of 0.77. Third, the SHapley Additive exPlanations (SHAP) framework is employed to interpret model outputs and identify key features, enhancing model transparency. On this event set, the proposed method achieves higher recall and lower omission than the official Wild Land Fire (WLF) and Forest Fire (FF) products, while producing more false-positive alerts. Ultimately, the main finding indicates that fusing high-frequency temporal dynamics with local spatial-contextual features mitigates the traditional limitations of static thresholding.