Regional Seasonal Crop Yield Forecasts Through Hybrid Crop Modeling and Remote Sensing Data Assimilation

Reliable regional crop yield forecasts are increasingly challenged by climate variability, extreme weather events, and growing pressure on land and water resources. Process-based crop and agro-ecosystem models provide a physically consistent framework to assess these impacts, yet their predictive skills at regional scales remain limited by uncertainties in initial conditions, parameterization, and the representation of in-season stress dynamics. At the same time, Earth observation (EO) data provide spatially explicit information on crop phenology and vegetation status that can help constrain and update model simulations.

This study investigates hybrid modeling and data assimilation strategies to improve seasonal yield predictions by integrating satellite-derived vegetation indicators (e.g., fraction of absorbed photosynthetically active radiation (FAPAR)) with the process-based agro-ecosystem model LPJmL. The focus is on regional-scale applications, using Bavaria, Germany as a case study representative for a heterogeneous and hydrological complex landscape, and on assessing how EO-informed initial states and in-season updates influence yield predictions throughout the growing season.

Time series of FAPAR observations are used to characterize crop phenology and canopy dynamics during the growing season and are integrated with LPJmL simulations through different coupling strategies. As LPJmL does not natively support continuous EO data assimilation, several integration pathways are explored, including parameter forcing, ensemble-based approaches, and hybrid extensions that combine process-based modeling with machine learning components trained on model outputs, EO, and meteorological inputs. These hybrid elements are designed to leverage EO and meteorological information to account for non-linear effects and growth-stage-dependent responses that are difficult to capture in purely process-based algorithms.

Meteorological forcing is derived from ERA5-Land reanalysis and C3S seasonal forecast data, with sensitivity experiments exploring the role of seasonal forecast information. Particular emphasis is placed on the role of climate extremes during critical phenological phases and their implications for seasonal yield variability. Model calibration and evaluation are conducted using historical yield statistics and regionally consistent land-use information, allowing an assessment of uncertainty related to parameter choices, assimilation strategies, and hybrid model components.

The presented framework contributes to ongoing efforts to link regional crop models with EO vegetation dynamics data through scalable and transferable methods. By combining process understanding with data-driven constraints, this work aims to improve the robustness of seasonal yield forecasting and to support future applications in agricultural and food security monitoring, climate impact assessment, and adaptation planning.