Enhanced pine wilt disease outbreak prediction: Integrating deep learning- detected infected trees with species distribution modeling

Abstract

Pine Wilt Disease (PWD) severely damages pine forests worldwide, economically and ecologically. Current PWD control employs reactive “detect and remove” approaches, which fail as the disease spreads before symptoms appear. We integrated U-Net deep learning detection with species distribution modeling to overcome these limitations. For our study area in Jinju, South Korea, we identified 8365 PWD-infected trees during September 2022–April 2023. We detected infected trees using aerial imagery and U-Net, and combined this data with historical infection records for prediction models that reflect PWD transmission mechanisms. We compared three cases: (A) using only environmental predictor variables, (B) adding historical infection records into case A, and (C) integrating historical infection records with deep learning-detected infected trees into case B. Adding infection data improved prediction accuracy compared to that of the environmental-only case A. Case C, integrating both infection data, showed the highest prediction accuracy. Cases B and C predicted more concentrated high-risk outbreak areas (30.5 % and 28.9 % smaller than those in Case A), which enables efficient resource allocation. Drone-based surveys confirmed that our predicted highrisk areas closely matched actual infection patterns. Infection proximity variables contributed most to model performance, outweighing those of environmental variables. This approach transforms PWD management from reactive to proactive, allowing targeted interventions that optimize limited resources. It provides a practical framework for managing forest diseases with similar transmission mechanisms. Our findings underscore the importance of incorporating disease transmission dynamics into ecological modeling and demonstrate how ecological informatics can enhance forest pest management, even with limited data.