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

Area of Science:

• Environmental Science and Engineering
• Computer Science and Artificial Intelligence
• Remote Sensing and Geospatial Technology

Background:

• Wildfires present a significant global threat, exacerbated by climate change, leading to ecological damage and economic losses.
• Existing wildfire detection methods struggle with real-time accuracy and early warning capabilities.
• Unmanned Aerial Vehicles (UAVs) combined with Artificial Intelligence (AI) offer a promising avenue for enhanced wildfire surveillance.

Purpose of the Study:

• To introduce ResCaps-TC-Attn-Fire, a novel deep learning framework for real-time forest fire detection using UAVs.
• To enhance the accuracy, speed, and reliability of wildfire detection systems.
• To provide a robust solution for early wildfire detection and monitoring, aiding in mitigation efforts.

Main Methods:

• Development of ResCaps-TC-Attn-Fire, integrating Residual-Capsule Networks, Threshold Convolution, and Attention Mechanisms.
• Utilizing a comprehensive dataset of 14,140 UAV-sourced images for model training and evaluation.
• Comparative analysis against existing methods such as YOLOv3, ABi-LSTM, and Enhanced YOLOv8n.

Main Results:

• ResCaps-TC-Attn-Fire achieved superior performance with 99.78% accuracy, 99.7% precision, and 99.8% recall.
• The model demonstrated significantly faster early detection (3.2s faster than YOLOv3) and a low false alarm rate (0.1%).
• Fire intensity estimation was achieved with a Mean Absolute Error (MAE) of 0.15.

Conclusions:

• ResCaps-TC-Attn-Fire represents a highly effective and reliable AI-driven solution for real-time UAV-based wildfire detection.
• The framework shows significant potential for wildfire mitigation, outperforming current state-of-the-art methods.
• Future work should focus on optimizing computational cost and power consumption for broader deployment.