UAV-FlameNet: A Lightweight and High-Precision Flame Detection Model for UAV Aerial Fire Monitoring

Area of Science:

• Computer Vision
• Artificial Intelligence
• Wildfire Management

Background:

• Early flame detection using Unmanned Aerial Vehicles (UAVs) is critical for wildfire prevention.
• Challenges include small object scales, morphological distortions, background interference, and limited edge-computing power in aerial imagery.

Purpose of the Study:

• To propose UAV-FlameNet, a lightweight deep learning algorithm for efficient and accurate flame detection from UAVs.
• To address the challenges of small flame detection, background interference, and computational constraints in edge devices.

Main Methods:

• Developed UAV-FlameNet, a lightweight algorithm based on YOLOv11.
• Introduced ADown lossless downsampling to preserve faint flame features.
• Implemented Dynamic Cascaded Flame Aggregation (DCFA) with deformable convolutions and variance-aware attention to model thermal radiation and suppress glares.
• Utilized MPDIoU loss for accelerated bounding box convergence.

Main Results:

• UAV-FlameNet achieved 83.2% mAP@0.5 on the LAWD dataset.
• The model has 2.72 M parameters and 5.66 GFLOPs, outperforming YOLOv11 by 2.2% in accuracy while reducing computational load by 12.1%.
• Achieved real-time performance with 121 FPS on GPU and 11.7 FPS on CPU.

Conclusions:

• UAV-FlameNet offers a Pareto-optimal balance of accuracy, robustness, and deployment efficiency.
• Provides a reliable solution for UAV-borne real-time fire early warning systems.
• Demonstrates significant improvements over existing methods for aerial flame detection.