An Efficient Forest Smoke Detection Approach Using Convolutional Neural Networks and Attention Mechanisms

Area of Science:

• Computer Science
• Environmental Science
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are prevalent in forest fire detection but lack optimization for smoke characteristics.
• Existing models may not effectively capture the nuances of smoke plumes, impacting early detection accuracy.

Purpose of the Study:

• To propose a novel CNN-based forest smoke detection model with an optimized backbone architecture.
• To enhance detection accuracy and reduce computational load for early forest fire detection.

Main Methods:

• Developed a CNN model with a novel backbone architecture utilizing kernels of varying sizes for multi-view smoke plume detection.
• Implemented a depth-wise convolution of a coordinate kernel to better extract vertical smoke features and decrease computational cost.
• Integrated an attention mechanism to enable the model to prioritize relevant information.

Main Results:

• The proposed model achieved a detection accuracy of 52.9 average precision (AP), outperforming popular existing models.
• Demonstrated a significant reduction in model parameters and Giga Floating-point Operations (GFLOPs) compared to other models.
• Successfully enhanced the detection of smoke plumes of various sizes and improved feature extraction.

Conclusions:

• The novel CNN backbone architecture offers superior performance in forest fire smoke detection.
• The model provides a computationally efficient solution for early forest fire detection, balancing accuracy and resource utilization.
• This approach represents a significant advancement in applying AI for environmental monitoring and fire prevention.