Federated learning based fire detection method using local MobileNet.

Area of Science:

• Computer Science
• Artificial Intelligence
• Environmental Monitoring

Background:

• Forest fires pose significant risks, with increasing frequency due to climate change.
• Existing fire detection methods struggle with accuracy in diverse environments.
• Early and precise fire prediction is crucial for mitigating damage.

Purpose of the Study:

• To introduce a novel Federated Learning (FL) based method, Indoor-Outdoor FireNet (IOFireNet), for accurate fire detection and localization.
• To enhance the precision of fire region segmentation and reduce noise in images.
• To improve fire prediction capabilities for early warning systems.

Main Methods:

• Developed Indoor-Outdoor FireNet (IOFireNet) using Federated Learning (FL) for global model aggregation.
• Implemented a Bilateral Filter (BF) for image preprocessing to reduce noise.
• Utilized Super Pixel-based Adaptive Clustering (SPAC) for precise fire region segmentation.
• Employed MobileNet for efficient data processing and prediction of fire spread and severity.

Main Results:

• IOFireNet achieved 98.65% accuracy for fire detection and 97.14% mean IoU for segmentation.
• The SPAC model outperformed graph cut and CRF models in mean IoU by 2.45%.
• The proposed model showed superior accuracy compared to VGG-19, ResNet-50, Inception, and Dense Net.

Conclusions:

• The FL-based IOFireNet offers a robust solution for accurate fire detection and localization in varied environments.
• The integration of BF and SPAC significantly enhances image clarity and segmentation precision.
• This method provides a foundation for more effective early warning systems, crucial for disaster management.