Scholarlink

<p> The red turpentine beetle (<i>Dendroctonus valens</i>) is the most widespread pine pest in North America. Since its invasion into China in 1998, it has killed over 10 million pine trees, causing severe ecological and economic losses, and has been designated as a high-priority managed invasive species. Nevertheless, characterizing its spatiotemporal spread remains challenging, as fundamental processes of growth, dispersal, and establishment are governed by nonlinear responses to environmental conditions. To enable precise prevention and control, this study first analyzes the historical spatiotemporal patterns of D. valens in China and identifies the key factors influencing its spread. Furthermore, we propose a spatiotemporal graph convolutional network-based risk prediction framework (MG-STGCN) for D. valens, integrating multi-dimensional graph structures with a spatial attention structure (SAS) and a gated recurrent unit (GRU) to generate county-level risk predictions across China. The predicted risk distribution exhibits a clear spatial pattern concentrated east of the Hu Huanyong Line, forming a continuous belt across Northeast and North China. High-risk areas are primarily clustered in southern Northeast China and adjacent regions of North China, while only limited and scattered risk signals appear in western transitional zones. Compared with the current distribution, the overall high-risk clusters remain relatively stable, but a slight northeastward shift can be observed, accompanied by emerging signals of westward penetration. The proposed MG-STGCN achieves the best performance among all compared models, with a recall of 89.57%, precision of 90.68%, and an F1-score of 90.12%. The improvements are mainly attributed to the complementary effects of the temporal modeling module and the spatial attention mechanism, which enhance the detection of outbreak signals while improving the discrimination of non-outbreak regions under highly imbalanced conditions. Attention-based analysis reveals that pest spread is dominated by local spatial continuity, with limited long-distance interactions. Different graph structures capture complementary transmission mechanisms, where spatial and wind-driven graphs enhance sensitivity to spread, while host-related connectivity improves specificity by constraining ecologically feasible pathways, resulting in a balanced predictive performance. This framework provides a reliable tool for large-scale pest risk assessment, and supporting more targeted surveillance and management strategies.</p> Graphical abstract <p></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

Multi-graph spatial–temporal graph convolutional networks for predicting the spread of Dendroctonus valens in China

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月26日

查看全文
  • Hongwei Zhou,
  • Yongzheng Li,
  • Jun Yang,
  • Haochang Hu,
  • Chengzhe Wang,
  • Huixiang Liu,
  • Yun Lin

摘要

The red turpentine beetle (Dendroctonus valens) is the most widespread pine pest in North America. Since its invasion into China in 1998, it has killed over 10 million pine trees, causing severe ecological and economic losses, and has been designated as a high-priority managed invasive species. Nevertheless, characterizing its spatiotemporal spread remains challenging, as fundamental processes of growth, dispersal, and establishment are governed by nonlinear responses to environmental conditions. To enable precise prevention and control, this study first analyzes the historical spatiotemporal patterns of D. valens in China and identifies the key factors influencing its spread. Furthermore, we propose a spatiotemporal graph convolutional network-based risk prediction framework (MG-STGCN) for D. valens, integrating multi-dimensional graph structures with a spatial attention structure (SAS) and a gated recurrent unit (GRU) to generate county-level risk predictions across China. The predicted risk distribution exhibits a clear spatial pattern concentrated east of the Hu Huanyong Line, forming a continuous belt across Northeast and North China. High-risk areas are primarily clustered in southern Northeast China and adjacent regions of North China, while only limited and scattered risk signals appear in western transitional zones. Compared with the current distribution, the overall high-risk clusters remain relatively stable, but a slight northeastward shift can be observed, accompanied by emerging signals of westward penetration. The proposed MG-STGCN achieves the best performance among all compared models, with a recall of 89.57%, precision of 90.68%, and an F1-score of 90.12%. The improvements are mainly attributed to the complementary effects of the temporal modeling module and the spatial attention mechanism, which enhance the detection of outbreak signals while improving the discrimination of non-outbreak regions under highly imbalanced conditions. Attention-based analysis reveals that pest spread is dominated by local spatial continuity, with limited long-distance interactions. Different graph structures capture complementary transmission mechanisms, where spatial and wind-driven graphs enhance sensitivity to spread, while host-related connectivity improves specificity by constraining ecologically feasible pathways, resulting in a balanced predictive performance. This framework provides a reliable tool for large-scale pest risk assessment, and supporting more targeted surveillance and management strategies.

Graphical abstract