Ensembling biogeochemical models to assess uncertainty in the estimation of agri-environmental indicators at the spatial scale

Agricultural practices can cause substantial nutrient losses, greenhouse gas (GHG) emissions, and soil degradation, highlighting the importance of agri-environment-climate measures (AECMs) under the EU Common Agricultural Policy (CAP). Evaluating AECM performance is challenging due to spatial and temporal variability, diverse pedoclimatic conditions, and limited field monitoring. To address this, this PhD research developed an integrated framework for the Veneto Region, combining mechanistic improvements to the EPIC model, a multi-model ensemble including EPIC, DNDC, and DayCent, and regional-scale GIS analysis. The ensemble enabled model comparison, revealing structural uncertainty and identifying areas of higher or lower reliability, supporting nuanced assessment of AECM effectiveness. The first objective focused on improving EPIC’s representation of ammonia (NH3) volatilization through an hourly-step submodel capturing ammonium adsorption, pH-dependent NH3–NH4+ partitioning, and mass transfer. Calibration and validation across two contrasting pedoclimatic conditions and fertilizer types demonstrated enhanced prediction of cumulative and daily NH3 losses compared to the original model (R2 = 0.79 vs. 0.50; RMSE = 10.6 vs. 17.8 kg N ha-1), enabling more accurate evaluation and management strategies to reduce volatilization. The second objective evaluated EPIC, DNDC, and DayCent against a four-year maize lysimeter experiment under free-drainage and shallow water table conditions. Continuous measurements of N2O and CO2 fluxes, SOC, yield, nitrogen uptake, water percolation, and NO3- leaching were used for calibration and validation. EPIC most accurately reproduced SOC, yield, and NO3- leaching, DNDC captured N uptake reasonably, while DayCent overestimated N2O under shallow water tables. All models struggled to capture observed variability. The ensemble reduced structural uncertainty and allowed systematic inter-model comparison, though EPIC often outperformed it for specific variables, highlighting areas of higher or lower predictive confidence. In the third chapter, the calibrated ensemble was applied across 4,176 polygonal units, representing 17 crops, multiple fertilization regimes, 58 soils, and 15 meteorological areas, totaling ~543,000 simulations to evaluate AECM effectiveness under the 2023–2027 Rural Development Complement. Inputs and outputs were integrated with GIS data on soil, topography, climate, and management to assess key C and N fluxes. Results showed that AECM effectiveness depended on location: conservation agriculture increased SOC, while pastures and meadows reduced nutrient losses but increased NH3 volatilization. The ensemble highlighted inter-model variability and areas of higher uncertainty, guiding prioritization of interventions and field validation. Overall, this research demonstrates that combining mechanistic model improvements, multi-model ensembles, and spatially explicit analysis enhances the reliability of environmental assessments, informs evidence-based AECM design, and supports context-specific, results-oriented policy. Further improvements could include expanding ensemble size, incorporating long-term observations, and linking biophysical outputs with socio-economic data to optimize efficient, equitable, and climate-smart agricultural management.