Enhanced Lightweight YOLOX for Small Object Wildfire Detection in UAV Imagery

Area of Science:

• Computer Vision
• Artificial Intelligence
• Forestry and Fire Management

Background:

• Unmanned aerial vehicle (UAV) imagery is crucial for forest fire detection, but current methods suffer from omission errors, low accuracy, and poor early warning capabilities.
• Existing target detection algorithms struggle with the complexities of fire environments and small target detection from UAV platforms.

Purpose of the Study:

• To develop an enhanced YOLOX network for rapid and accurate forest fire detection in UAV-captured aerial imagery.
• To improve the detection of small fire targets and reduce false positives in complex environments.

Main Methods:

• Implemented a novel multi-level-feature-extraction structure (CSP-ML) to improve small-target fire area detection.
• Integrated a CBAM attention mechanism to minimize background noise interference.
• Introduced an adaptive-feature-extraction module for better feature fusion and utilized the CIoU loss function for enhanced positive sample recognition.

Main Results:

• The improved YOLOX network demonstrated superior performance, with mAP@50 and mAP@50_95 increasing by 6.4% and 2.17% respectively, compared to the standard YOLOX.
• Achieved a 96.3% mAP in multi-target and small-target flame scenarios, outperforming FasterRCNN, SSD, and YOLOv5 by significant margins.
• Exhibited a lower omission rate and higher detection accuracy, proving effective for small-target detection in challenging fire conditions.

Conclusions:

• The proposed improved YOLOX network offers enhanced capabilities for forest fire detection from UAVs, addressing key limitations of current technologies.
• This advancement provides crucial support for UAV-based forest fire patrol and rescue operations, improving early warning and response effectiveness.