Semi-supervised wildfire smoke detection based on smoke-aware consistency.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Monitoring

Background:

• Smoke's semi-transparency poses challenges for image analysis due to background integration.
• Limited annotated real-world forest smoke data hinders effective model training.
• Distinguishing smoke from visually similar objects requires robust detection methods.

Purpose of the Study:

• To develop a semi-supervised learning strategy for improved wildfire smoke detection.
• To enhance pixel and context perceptual consistency across diverse backgrounds.
• To refine fire-smoke detection networks for better performance and efficiency.

Main Methods:

• Proposed a novel smoke-aware consistency (SAC) semi-supervised learning strategy.
• Introduced a triple classification assistance strategy for smoke and smoke-like object discrimination.
• Simplified the LFNet fire-smoke detection network to LFNet-v2, integrating the new strategies.

Main Results:

• The proposed SAC strategy maintains pixel and context perceptual consistency.
• Triple classification effectively discriminates smoke from smoke-like objects.
• LFNet-v2 demonstrated superior performance compared to state-of-the-art object detection algorithms on wildfire smoke datasets.
• Achieved satisfactory results under challenging weather conditions.

Conclusions:

• The developed semi-supervised learning strategy and triple classification significantly advance wildfire smoke detection.
• The simplified LFNet-v2 offers an efficient and effective solution for real-world applications.
• The method shows strong potential for wildfire monitoring and early warning systems.