STFI: a spatio-temporal deep fusion architecture for high-accuracy fire spot identification near power transmission lines

The increasing frequency and intensity of wildfires have led to frequent power transmission line outages, posing critical threats to the operational security of modern power grids. However, conventional satellite-based wildfire detection methods often fail to accurately identify small-scale, localized fire events near transmission infrastructure due to limited spatiotemporal resolution and generalization capabilities. To address these challenges, this study proposes a novel spatio-temporal deep fusion-based fire spot identification (STFI) framework that leverages high-frequency Himawari-8/9 geostationary satellite data, with an emphasis on the simultaneous and hierarchical extraction of multi-spatiotemporal wildfire features at multiple levels and their deep fusion. The proposed framework integrates three core components: (1) a deep-select-conv block (DSCB) for multi-level spatial and spectral feature extraction, (2) a dynamic-temporal transformer block (DTFB) for modeling long-range temporal dependencies, and (3) a dual dynamic fusion head (DDFH) for adaptive deep fusion and classification of heterogeneous spatiotemporal features. The experiment results demonstrate that the STFI-based dual-branch architecture, termed ST-DC&DT-Net, achieves a high classification accuracy of 96.01%, with a significantly reduced miss rate of 3.71%. These findings validate the effectiveness and reliability of the proposed framework for real-time, fine-grained wildfire monitoring in high-risk areas near power transmission lines.