Edge-Based Autonomous Fire and Smoke Detection Using MobileNetV2

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Forest fires present significant ecological, human, and climatic threats, demanding advanced detection methods.
• Existing detection systems (sensors, satellites, centralized analysis) face limitations like delays, false alarms, and restricted deployment.
• Current deep learning models, while accurate, are often too resource-intensive for real-time edge applications.

Purpose of the Study:

• To develop an autonomous, edge-based forest fire and smoke detection system.
• To optimize a lightweight deep learning model for resource-constrained edge devices.
• To enable real-time, localized wildfire detection and early mitigation.

Main Methods:

• Utilized a lightweight MobileNetV2 convolutional neural network for image analysis.
• Trained the model on a balanced dataset of fire, smoke, and non-fire images.
• Deployed and optimized the system on a Raspberry Pi 5 edge device for autonomous operation.

Main Results:

• Achieved a test accuracy of 97.98% for fire and smoke detection.
• Demonstrated near real-time inference with an average latency of 0.77 seconds per frame (1.3 FPS).
• Generated predictions locally, ensuring security and robustness without cloud dependency.

Conclusions:

• The proposed edge-based system offers a cost-effective and scalable solution for remote wildfire monitoring.
• Combines high accuracy, speed, and autonomous edge processing for timely fire detection.
• Enhances environmental monitoring and early wildfire mitigation capabilities.