Forest Fire Segmentation via Temporal Transformer from Aerial Images

Forest fires are among the most critical natural tragedies threatening forest lands and resources. The accurate early detection of is essential to reduce losses improve firefighting. Conventional firefighting techniques, based on ground inspection limited by field-of-view, lead insufficient monitoring capabilities for large areas. Recently, due their excellent flexibility ability cover regions, unmanned aerial vehicles (UAVs) have been used combat fire incidents. An step an autonomous system that monitors situations first locate in a video. State-of-the-art forest-fire segmentation methods vision transformers (ViTs) convolutional neural networks (CNNs) use single image. Nevertheless, has inconsistent scale form, small from long-distance cameras lack salient features, so image challenging. In addition, techniques CNNs treat all pixels equally overlook global information, limiting performance, while ViT-based suffer high computational overhead. To address these issues, we proposed spatiotemporal architecture called FFS-UNet, which exploited temporal information combining transformer into modified lightweight UNet model. First, extracted keyframe two reference frames using three different encoder paths parallel obtain shallow features perform feature fusion. Then, deep temporal-feature extraction, enhanced learning made extraction more robust. Finally, combined de-convolution decoder path via skip-connections segment fire. We evaluated empirical outcomes UAV-collected video Corsican Fire datasets. FFS-UNet demonstrated performance with fewer parameters achieving F1-score 95.1% IoU 86.8% video, 91.4% 84.8% dataset, were higher than previous techniques. Therefore, suggested model effectively resolved fire-monitoring issues UAVs.