Deep learning-based forest fire detection using an improved SSD algorithm with CBAM.

Area of Science:

• Computer Science
• Environmental Science
• Artificial Intelligence

Background:

• Forest fires pose significant ecological threats due to rapid spread and destructive potential.
• Existing forest fire detection systems struggle with variable flame/smoke features and environmental interference, leading to false positives and missed detections.

Purpose of the Study:

• To develop a novel object detection model, CBAM-SSD, for improved forest fire detection.
• To enhance the accuracy and reliability of detecting flames and smoke in complex environmental conditions.

Main Methods:

• Employed data augmentation (geometric, color transformations) to address data limitations.
• Integrated the CBAM module into the SSD backbone for adaptive feature extraction, focusing on critical fire regions.
• Reconstructed the feature extraction structure to improve perception of variable flame and smoke characteristics.

Main Results:

• CBAM-SSD achieved a mAP@0.5 of 97.55% for flames and smoke, a 1.53% improvement over baseline SSD.
• Flame detection AP50 reached 96.61% (3.01% increase) with 96.40% recall.
• Smoke detection AP50 reached 98.49% with 98.80% recall, demonstrating superior accuracy and reduced detection errors.

Conclusions:

• The CBAM-SSD model is lightweight, suitable for real-time forest fire detection.
• The model significantly improves detection accuracy and reduces false and missed detections.
• Offers an efficient, convenient, and accurate solution for real-time forest fire monitoring.