AI-driven pest forecasting and decision-support for precision forestry management

Thesis Abstract

The rapid expansion of forestry operations and increasing incidences of pest outbreaks pose significant threats to sustainable timber production and ecosystem services, necessitating proactive, scalable management solutions. This study addresses the gap in integrated decision-support systems (DSS) that combine AI-driven pest forecasting with precision forestry interventions to optimize pest control, resource allocation, and operational timing under dynamic climate and stand conditions. The aim is to develop and validate an AI-enabled forecasting and decision-support framework that integrates multi-source data to deliver actionable pest risk assessments and targeted management recommendations for contemporary forestry systems. Specific objectives include (1) to synthesize a multi-disciplinary model linking pest phenology, host susceptibility, and environmental drivers within a machine learning forecasting pipeline; (2) to design a real-time DSS that translates predictive outputs into calibrated silvicultural actions and treatment schedules; (3) to quantify the performance of the framework against conventional pest management approaches in terms of accuracy, cost-efficiency, and ecological impact; and (4) to assess user acceptance and operational feasibility among forestry professionals. A convergent mixed-methods design is employed. The study uses a quantitative forecasting component and a qualitative usability assessment. The population comprises managed temperate conifer forests spanning 12 sites across a regional timber district, with a total inventory of approximately 120,000 hectares. A stratified random sample of 48 plots (4,800 ha) is selected to capture variability in pest pressure, stand age, species composition, and microclimate. Data collection integrates (i) historical pest incidence records (10 years, site-level), (ii) remotely sensed stand metrics (multispectral and LiDAR data, 2019–2024), (iii) in situ climatic and soil data from 60 meteorological stations, (iv) trap-and-release pheromone monitoring data for key pests, and (v) management intervention records (chemical, biological, and mechanical control). The AI forecasting component employs ensemble models combining gradient boosting (XGBoost), recurrent neural networks (LSTM), and survival analysis to predict short-term (0–6 months) and medium-term (6–24 months) pest risk, with feature engineering guided by the Theory of Planned Behavior and the Optimal Foraging Theory to capture human-environment interactions and pest dynamics. Model training uses 70% of the data with 10-fold cross-validation, and the remaining 30% is held out for external validation. Model performance is evaluated using RMSE, AUC-ROC, Brier score, and calibration plots. The decision-support module applies a rule-based layer and a reinforcement learning (Q-learning) module to generate site-specific interventions—timing and type of silvicultural treatments, pheromone disruption, and targeted pesticides—while incorporating economic and ecological constraints. Qualitative data are gathered via semi-structured interviews (n=24 forestry professionals) and focus groups (6 groups) to evaluate DSS usability, perceived reliability, and organizational fit. Thematic analysis is conducted to identify barriers and drivers of adoption, with coding aligned to Technology Acceptance Model constructs and organizational readiness for change. Ethical considerations include informed consent, data anonymization, and compliance with forest stewardship regulations. Expected findings indicate that the integrated AI forecasting and DSS will achieve higher predictive accuracy for pest outbreaks (AUC-0.87–0.92) and stronger timeliness in intervention recommendations compared with traditional monitoring-based decision-making. The framework is anticipated to reduce unnecessary chemical inputs by 22–35% and improve net present value of stand management by 8–15% under baseline scenarios, while maintaining or enhancing biodiversity indicators. Sensitivity analyses are expected to reveal robustness to data sparsity and varying pest pressure, with uncertainty quantification embedded in probabilistic forecasts to support risk-informed decisions. The study contributes to knowledge by advancing a scalable, data-driven, end-to-end DSS that bridges ecological forecasting, optimization, and human–machine interaction in forestry contexts, and by offering a transferable methodology for other agricultural and forestry pest systems. The main conclusion is that AI-driven pest forecasting combined with a learning-based decision-support framework can significantly improve pest management efficiency and economic outcomes without compromising ecological integrity. Recommendations include broader adoption across forest types with modular adaptation of models, investment in data infrastructure for continuous updating, the incorporation of citizen-science contributions for trap data, and ongoing trainer-mentor programs to enhance user acceptance and ensure alignment with regulatory standards and ecosystem-based management principles.