BGC-LiteNet: BeiDou grid code embedded lightweight neural architecture for real-time UAV fire detection and...

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Early forest fire detection and localization via UAVs are crucial for timely warnings.
• Existing methods struggle with a trade-off between deep learning model accuracy and computational cost, or lightweight model localization precision.

Purpose of the Study:

• To develop an efficient end-to-end framework, BGC-LiteNet, for simultaneous fire detection and precise geographic localization using UAVs.
• To address the computational limitations of UAV platforms for real-time disaster prevention applications.

Main Methods:

• Integration of the BeiDou Grid Code (BGC) spatial standard into neural network feature learning.
• Development of a learnable geographic embedding module for pixel-grid correspondence.
• Implementation of latency-aware lightweight neural architecture search (L-NAS) for joint optimization of accuracy and hardware latency.

Main Results:

• BGC-LiteNet achieved 88.9% mean average precision (mAP) and 92.4% geolocation accuracy with 0.87M parameters and 38.2 ms latency on embedded platforms.
• Robust performance was maintained under challenging conditions: low illumination (mAP 72.3%), dense smoke (mAP 72.6%), and small fire points (recall 86.7%).

Conclusions:

• BGC-LiteNet offers a new paradigm for spatiotemporal intelligent edge computing in disaster prevention.
• The framework enables efficient, accurate, and real-time fire detection and localization on resource-constrained UAVs.