Scholarlink

<p>Real-time fire detection and precise geographic localization using unmanned aerial vehicles (UAVs) are critical for early forest-fire warning. However, existing approaches face a fundamental dilemma: achieving high detection accuracy requires complex deep learning models with prohibitive computational costs, while lightweight models sacrifice localization precision due to the lack of spatial priors. This study proposes BGC-LiteNet, an end-to-end framework that integrates the BeiDou Grid Code (BGC)—China’s national spatial reference standard—directly into neural network feature learning. A learnable geographic embedding module encodes pixel-grid correspondences at the input stage, enabling simultaneous detection and localization without external GIS post-processing. To achieve millisecond-level inference on resource-constrained UAV platforms, we develop a latency-aware lightweight neural architecture search (L-NAS) that jointly optimizes detection accuracy and hardware latency. Experimental results on a multi-scenario UAV dataset demonstrate that BGC-LiteNet achieves 88.9% mean average precision (mAP) and 92.4% geolocation accuracy with only 0.87&#xa0;M parameters and 38.2 ms latency on embedded platforms. The model maintains robust performance under challenging conditions including low illumination (mAP 72.3%), dense smoke (mAP 72.6%), and small fire points (recall 86.7%). By embedding structured geographic priors and optimizing model latency simultaneously, BGC-LiteNet establishes a new paradigm for spatiotemporal intelligent edge computing in disaster prevention applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月21日

查看全文
  • Haiwen Yin,
  • Yong Yu,
  • Andong Hong,
  • Mengyang Hu,
  • Shunhang Wang,
  • Zhifan Zhang

摘要

Real-time fire detection and precise geographic localization using unmanned aerial vehicles (UAVs) are critical for early forest-fire warning. However, existing approaches face a fundamental dilemma: achieving high detection accuracy requires complex deep learning models with prohibitive computational costs, while lightweight models sacrifice localization precision due to the lack of spatial priors. This study proposes BGC-LiteNet, an end-to-end framework that integrates the BeiDou Grid Code (BGC)—China’s national spatial reference standard—directly into neural network feature learning. A learnable geographic embedding module encodes pixel-grid correspondences at the input stage, enabling simultaneous detection and localization without external GIS post-processing. To achieve millisecond-level inference on resource-constrained UAV platforms, we develop a latency-aware lightweight neural architecture search (L-NAS) that jointly optimizes detection accuracy and hardware latency. Experimental results on a multi-scenario UAV dataset demonstrate that BGC-LiteNet achieves 88.9% mean average precision (mAP) and 92.4% geolocation accuracy with only 0.87 M parameters and 38.2 ms latency on embedded platforms. The model maintains robust performance under challenging conditions including low illumination (mAP 72.3%), dense smoke (mAP 72.6%), and small fire points (recall 86.7%). By embedding structured geographic priors and optimizing model latency simultaneously, BGC-LiteNet establishes a new paradigm for spatiotemporal intelligent edge computing in disaster prevention applications.