Enhancing the fine-scale prediction of pine wilt disease through the integration of a multihead attention mechanism.

backgroundPine wilt disease (PWD) is recognized as the most severe forestry pest in China, with highly complex spatial transmission dynamics. Accurate prediction of its potential risk areas is of great significance for effective prevention and control. In this study, Chongqing, China, was selected as the case study region. Based on stand-level PWD occurrence data from 2017 to 2020, we constructed a transmission topology and proposed a graph convolutional network (GCN) model enhanced with a multihead attention mechanism to enhance risk prediction in unaffected areas.

resultsThe results indicate that PWD in Chongqing exhibits a dynamic pattern of 'continuous spread-localized containment' and a spatial distribution characterized by a 'core-diffusion' structure. Compared with traditional GCN, support vector machine and graph transformer models, the improved model achieved slightly better performance in terms of accuracy, recall and F1-score, indicating improved predictive performance, although the overall improvement was moderate. The predicted high-risk areas suggested a spatial distribution pattern of 'local clustering with hierarchical interleaving'.

conclusionThese findings suggest that integrating attention mechanisms into GCN can contribute to improving the identification of potential PWD risk zones. Stand-level prediction enables fine-scale localization to specific forest compartments, which may facilitate the early identification of potential high-risk areas. This can support the prioritization of control resources and the design of hierarchical prevention strategies, serving as a decision-support tool. Overall, this study provides methodological insights and methodological insights for fine-scale monitoring, early-warning system development, and the formulation of evidence-based prevention strategies for PWD. © 2026 Society of Chemical Industry.