Scholarlink

<p>Pine Wilt Disease (PWD) is a severe forest quarantine pest, and its effective control heavily relies on early and accurate detection. Unmanned Aerial Vehicle (UAV) technology provides a platform for large-scale monitoring, but&#xa0;its onboard computing resources are severely limited, imposing stringent requirements on algorithm lightweightness and efficiency. Existing methods based on UAV RGB imagery face bottlenecks in early identification and practical deployment due to&#xa0;limited spectral information and the difficulty in balancing computational efficiency with detection performance. To address these issues, this paper proposes a lightweight multi-spectral fusion detection model (PWD-MFS) for UAV edge computing scenarios. First, to overcome the limitations of RGB information, a multi-spectral fusion dataset is constructed. Using an improved SIFT-RANSAC-based registration method, high-precision alignment between RGB and near-infrared (840&#xa0;nm) bands is achieved, resulting in spatially consistent multi-channel inputs. Second,&#xa0;to meet the real-time requirements of edge deployment, the model architecture integrates the hierarchical semantic design of HGNetv2 and the efficiency of Ghost convolution, significantly reducing the parameter count while maintaining a strong capacity for extracting multi-scale PWD features. Finally, a detection architecture incorporating a Lightweight Shared Convolutional Detection (LSCD) head and a Multi-Path Coordinate Attention (MPCA) module is designed. Through parameter sharing and an adaptive feature enhancement mechanism, the detection accuracy for weakly visible targets is further improved. The model contains only 3.8&#xa0;M parameters and 9.5 GFLOPs, achieving a precision of 0.88, recall of 0.818, F1-score of 0.9, and mAP of 0.848 in testing, outperforming mainstream models and&#xa0;striking a notable balance between accuracy and efficiency. Ablation studies show that introducing the near-infrared band improves key metrics including precision, recall, F1-score, mAP by 4.0%, 3.5%, 2.8%, and 3.8%, respectively, validating the effectiveness of multi-spectral fusion for early PWD detection. This study provides a feasible technical solution for&#xa0;resource-constrained UAV platforms&#xa0;to achieve early, accurate, and real-time monitoring of Pine Wilt Disease. The code is available at <a href="https://github.com/123456WPF/PWD-MFS/tree/master">https://github.com/123456WPF/PWD-MFS/tree/master</a></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

PWD-MFS: a lightweight multispectral fusion network for detection of pine wilt disease with UAV imagery

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月17日

查看全文
  • Pengfei Wang,
  • Baohua Yang,
  • Zhenya Tao,
  • Longwa Zhang,
  • Songyan Lu,
  • Ziwei Liu

摘要

Pine Wilt Disease (PWD) is a severe forest quarantine pest, and its effective control heavily relies on early and accurate detection. Unmanned Aerial Vehicle (UAV) technology provides a platform for large-scale monitoring, but its onboard computing resources are severely limited, imposing stringent requirements on algorithm lightweightness and efficiency. Existing methods based on UAV RGB imagery face bottlenecks in early identification and practical deployment due to limited spectral information and the difficulty in balancing computational efficiency with detection performance. To address these issues, this paper proposes a lightweight multi-spectral fusion detection model (PWD-MFS) for UAV edge computing scenarios. First, to overcome the limitations of RGB information, a multi-spectral fusion dataset is constructed. Using an improved SIFT-RANSAC-based registration method, high-precision alignment between RGB and near-infrared (840 nm) bands is achieved, resulting in spatially consistent multi-channel inputs. Second, to meet the real-time requirements of edge deployment, the model architecture integrates the hierarchical semantic design of HGNetv2 and the efficiency of Ghost convolution, significantly reducing the parameter count while maintaining a strong capacity for extracting multi-scale PWD features. Finally, a detection architecture incorporating a Lightweight Shared Convolutional Detection (LSCD) head and a Multi-Path Coordinate Attention (MPCA) module is designed. Through parameter sharing and an adaptive feature enhancement mechanism, the detection accuracy for weakly visible targets is further improved. The model contains only 3.8 M parameters and 9.5 GFLOPs, achieving a precision of 0.88, recall of 0.818, F1-score of 0.9, and mAP of 0.848 in testing, outperforming mainstream models and striking a notable balance between accuracy and efficiency. Ablation studies show that introducing the near-infrared band improves key metrics including precision, recall, F1-score, mAP by 4.0%, 3.5%, 2.8%, and 3.8%, respectively, validating the effectiveness of multi-spectral fusion for early PWD detection. This study provides a feasible technical solution for resource-constrained UAV platforms to achieve early, accurate, and real-time monitoring of Pine Wilt Disease. The code is available at https://github.com/123456WPF/PWD-MFS/tree/master