Residual capsule network with threshold convolution and attention mechanism for forest fire detection using UAV imagery.

Wildfires pose a severe threat to ecosystems, economies, and human lives, exemplified by the 2019 Australian bushfires, which devastated 46 million acres, destroyed thousands of structures, and caused USD 148.5 billion in economic losses, alongside profound ecological damage. With climate change intensifying the frequency and severity of such events, there is a pressing need for advanced, real-time wildfire detection systems. Unmanned Aerial Vehicles (UAVs) integrated with remote sensing and Artificial Intelligence (AI) offer a promising solution for early detection and continuous monitoring. This paper introduces ResCaps-TC-Attn-Fire, a novel deep learning framework tailored for UAV-based forest fire detection, combining Residual-Capsule Networks, Threshold Convolution, and Attention Mechanisms. Residual-Capsule Networks enhance the capture of spatial hierarchies and inter-feature relationships, improving robustness to diverse fire characteristics, while Threshold Convolution filters irrelevant features to boost generalization and efficiency. The Attention Mechanism prioritizes critical fire-related regions, ensuring precise detection. Evaluated on a comprehensive UAV-sourced dataset of 14,140 images, ResCaps-TC-Attn-Fire achieves an impressive 99.78% accuracy, 99.7% precision, and 99.8% recall, surpassing existing methods like YOLOv3 (85.2%), ABi-LSTM (96.2%), and Enhanced YOLOv8n (99.0%). With early detection (3.2 s faster than YOLOv3), a 0.1% false alarm rate, and fire intensity estimation (MAE 0.15), this model demonstrates its potential as a reliable, real-time solution for wildfire mitigation, despite challenges like higher computational cost (12 s inference, 15W power), paving the way for future optimizations and broader deployment.