Light-Weight Student LSTM for Real-Time Wildfire Smoke Detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Environmental Monitoring

Background:

• Wildfire detection is crucial, with camera-based systems offering cost-effectiveness but struggling to differentiate smoke from fog/clouds.
• Existing deep learning methods for smoke detection often require heavy models, hindering real-time application.
• Accurate and efficient wildfire smoke detection is essential for timely response and mitigation.

Purpose of the Study:

• To develop an efficient deep learning model for wildfire smoke detection that overcomes the limitations of existing methods.
• To improve the accuracy and speed of camera-based wildfire smoke detection systems.
• To reduce the computational load of deep learning models for real-time wildfire smoke detection.

Main Methods:

• A You-Only-Look-Once (YOLO) detector was combined with a long short-term memory (LSTM) classifier to analyze spatial and temporal smoke characteristics.
• A teacher-student framework was applied to a deep LSTM model to create a shallower, more efficient student model.
• The student LSTM model was designed with reduced layers and cells while aiming to preserve the performance of the original deep LSTM.

Main Results:

• The proposed teacher-student framework significantly reduced the number of parameters by up to 8.4-fold compared to the deep LSTM.
• The student model demonstrated a faster processing time, crucial for real-time detection applications.
• The lightweight student LSTM model maintained comparable detection performance to the deep LSTM on a wildfire benchmark dataset.

Conclusions:

• The teacher-student framework effectively creates a lightweight yet high-performing LSTM model for wildfire smoke detection.
• This approach enhances the feasibility of real-time, low-cost camera-based wildfire detection systems.
• The method offers a promising solution for improving early wildfire detection and response.