Scholarlink

<p>This paper proposes LSRO-YOLO (Lightweight Scale-Shared Rotated Object Detector), a lightweight oriented object detection architecture for Unmanned Aerial Vehicle (UAV) forest fire monitoring under resource-constrained onboard computing and complex canopy backgrounds. It addresses two major challenges in UAV wildfire detection: scale variations caused by changes in flight altitudes, and the irregular shapes and arbitrary orientations of early-stage wildland fire spots from top-down views. Specifically, an Efficient Up-Convolution Block (EUCB) is introduced to improve feature reconstruction under strict computational constraints. A C2f with Dilated-wise Reparam Block (C2f-DRB) enlarges the effective receptive field for diffuse smoke, and a Lightweight Scale-Shared Oriented Bounding Box head (LSS-OBB) enables orientation-aware localization with parameter sharing and model compression. In-domain evaluations on three complementary benchmarks demonstrate that LSRO-YOLO achieves mAP@0.5 scores of 90.8% on UAV-simulated Wildfire dataset, 85.7% on fixed-surveillance D-Fire dataset, and 96.5% on multi-scene Fire-8 dataset. The proposed method also shows competitive performance compared with recent advanced methods such as AsymmetricFPN, YOLOGX, and SlimNeck-YOLOv8. In addition, cross-dataset evaluation further demonstrates that the model, when trained on D-Fire, retains partial detection capability on unseen domains (mAP@0.5 scores: 53.1% on Fire-8, 35.0% on Wildfire). Relative to the YOLOv8n-OBB baseline, LSRO-YOLO reduces the parameter count by 25.3% to 2.30&#xa0;M and lowers computational complexity by 18.1% to 6.8 GFLOPs. Further comparisons with recent lightweight detectors, including YOLO11n, YOLO26n, and D-FINE-n, show that LSRO-YOLO achieves the lowest parameter count while maintaining competitive detection accuracy. These results indicate LSRO-YOLO provides an effective accuracy–efficiency trade-off and represents a promising candidate architecture for memory-constrained UAV forest fire monitoring, while hardware-level validation on embedded platforms remains an essential next step toward practical deployment.</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

A scale-shared feature-reconstruction network for rotation-aware forest fire detection under constrained UAV computing resources

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Chunmei Yang,
  • Chuanyang Cao,
  • Jie Yan,
  • Guoge Wang,
  • Xinghao Zhang,
  • Yucheng Ding,
  • Chaoqun Ma,
  • Yanwen Liu

摘要

This paper proposes LSRO-YOLO (Lightweight Scale-Shared Rotated Object Detector), a lightweight oriented object detection architecture for Unmanned Aerial Vehicle (UAV) forest fire monitoring under resource-constrained onboard computing and complex canopy backgrounds. It addresses two major challenges in UAV wildfire detection: scale variations caused by changes in flight altitudes, and the irregular shapes and arbitrary orientations of early-stage wildland fire spots from top-down views. Specifically, an Efficient Up-Convolution Block (EUCB) is introduced to improve feature reconstruction under strict computational constraints. A C2f with Dilated-wise Reparam Block (C2f-DRB) enlarges the effective receptive field for diffuse smoke, and a Lightweight Scale-Shared Oriented Bounding Box head (LSS-OBB) enables orientation-aware localization with parameter sharing and model compression. In-domain evaluations on three complementary benchmarks demonstrate that LSRO-YOLO achieves mAP@0.5 scores of 90.8% on UAV-simulated Wildfire dataset, 85.7% on fixed-surveillance D-Fire dataset, and 96.5% on multi-scene Fire-8 dataset. The proposed method also shows competitive performance compared with recent advanced methods such as AsymmetricFPN, YOLOGX, and SlimNeck-YOLOv8. In addition, cross-dataset evaluation further demonstrates that the model, when trained on D-Fire, retains partial detection capability on unseen domains (mAP@0.5 scores: 53.1% on Fire-8, 35.0% on Wildfire). Relative to the YOLOv8n-OBB baseline, LSRO-YOLO reduces the parameter count by 25.3% to 2.30 M and lowers computational complexity by 18.1% to 6.8 GFLOPs. Further comparisons with recent lightweight detectors, including YOLO11n, YOLO26n, and D-FINE-n, show that LSRO-YOLO achieves the lowest parameter count while maintaining competitive detection accuracy. These results indicate LSRO-YOLO provides an effective accuracy–efficiency trade-off and represents a promising candidate architecture for memory-constrained UAV forest fire monitoring, while hardware-level validation on embedded platforms remains an essential next step toward practical deployment.