Towards a comprehensive decision support system for agroforestry systems

Abstract Agroforestry systems (AFSs) enhance biodiversity, productivity, and climate resilience, yet their inherent complexity challenges traditional modeling approaches. This study presents an integrated framework combining process-based modeling (PBM), machine learning, and life cycle assessment (LCA) into a user-friendly decision support system (DSS) for AFS analysis and optimization. Multi-scenario simulations using Hi-sAFe (16 configurations varying in latitude and orientation) for maize–poplar systems produced detailed outputs, including land equivalent ratios (LER up to 1.29), crop yield (ranging from 4.3 to 8.1 t ha −1 ), and N 2 O emissions (1.6–9.4 kg N ha −1 ). We trained Random Forest models on PBM diagnostics, microclimate, and design covariates; post-hoc SHapley Additive exPlanations (SHAP) identified maximum air temperature (Tmax), distance to tree rows, and photosynthetically active radiation (PAR) as the dominant predictors of yield and LER. To address data gaps, long short-term memory (LSTM) and TrAdaBoost-LSTM models were used for microclimate forecasting, achieving high fidelity ( R 2 = 0.88) in capturing hourly temperature and humidity trends. Outputs have been integrated into an LCA workflow, producing CO 2 -equivalent and nitrogen footprint metrics that feed into a functional web-based DSS. This platform enables robust, transparent, and scalable AFS evaluation, supporting stakeholders with real-time scenario exploration and scientifically grounded decision-making tools. Furthermore, the framework enables greenhouse gas (GHG) accounting in AFSs by integrating modeled soil organic carbon dynamics and N 2 O emissions. A Germany-wide assessment illustrates the significant GHG reduction potential achievable through large-scale implementation of AFSs.