Multiscale wildfire and smoke detection in complex drone forest environments based on YOLOv8.

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Climate change intensifies forest fires, necessitating advanced monitoring.
• Traditional methods like manual inspection and satellite remote sensing have limitations in real-time forest fire detection.
• Drones combined with artificial intelligence (AI) are emerging as a mainstream solution for effective fire monitoring.

Purpose of the Study:

• To develop an improved AI model for enhanced forest fire detection using drones.
• To address limitations of existing models in accuracy, efficiency, and robustness against environmental variations.
• To improve early warning systems and emergency response for forest fires.

Main Methods:

• An improved YOLOv8-based deep learning model was developed.
• Incorporated local convolution and the EMA module to enhance feature interaction and reduce complexity.
• Introduced AgentAttention (combining Softmax and linear attention) in the Backbone for robust feature extraction.
• Designed the BiFormer module for adaptive fusion of global and local features to improve multi-scale and multi-angle detection.

Main Results:

• The improved model achieved 93.57% Precision and 88.51% Recall, outperforming the original model.
• Demonstrated significant improvements in efficiency with a 14.3% increase in FPS, 25% reduction in GFLOPs, and 19.7% reduction in Params.
• The model shows enhanced accuracy and robustness in detecting flames and smoke under various conditions.

Conclusions:

• The proposed AI model offers a significant advancement in drone-based forest fire detection.
• The enhancements lead to higher accuracy, increased efficiency, and better robustness, crucial for real-time monitoring.
• This research provides strong technical support for forest fire early warning, emergency response, and resource protection.