Improved Wildfire Detection and Segmentation Using a Global-Local Attention Mechanism for Doppler Radar Data

Effective detection and identification of wildfires are essential for efficient control and mitigation of their impacts. Satellite remote sensing is commonly used for hotspot detection, but its effectiveness is hindered in subtropical monsoon climates due to cloud and fog interference. Recently, smoke signals produced by wildfires have been successfully detected using weather Doppler radar, providing a valuable supplement to satellite-based monitoring. However, existing fire area segmentation techniques based on radar reflectivity data face significant challenges, including poor segmentation at target boundaries, limited adaptability to targets of varying sizes, and insufficient consideration of temporal correlations between data frames. To address these issues, we propose a novel wildfire segmentation approach that integrates a global-local attention mechanism with temporal correlation information. First, the Global-Local Attention (GPA) module is used to extract both key local features and global distribution patterns, thereby enhancing segmentation accuracy, particularly at target boundaries. Second, a Multi-Scale Fusion (MSF) module combines spatial features at multiple scales, enabling the model to better capture diverse spatial hierarchies of fire points and adapt to targets of varying sizes. Finally, a Temporal Feature Extraction (TEF) module is introduced to capture temporal dependencies, leveraging the correlations between consecutive data frames. Experimental results on the Fire-Radar Reflectivity (FRR) dataset demonstrate that our model outperforms baseline approaches. Compared to the Trans-UNet model, it improves pixel-level accuracy and target-level precision by approximately 3% and 4%, respectively, and by approximately 3% and 5%, respectively, compared to the state-of-the-art Evit-Unet model.