Deep-learning surrogate crop modelling for scalable seasonal-to-climate crop-risk assessment

Anticipating climate-related crop stress requires crop-risk information that is spatially explicit, probabilistic and fast enough for large seasonal forecast and climate-scenario ensembles. We present the Surrogate Engine for Crop Simulations Framework (SECSF), a deep-learning framework that emulates the process-based ECroPS model using only daily minimum and maximum temperature and precipitation. Trained on ERA5-forced ECroPS simulations for grain maize and spring barley, SECSF closely reproduces daily crop-growth dynamics and harvest timing while reducing computational cost by around four orders of magnitude, enabling ensemble-scale inference suitable for research and operational pipelines such as agricultural early warning and adaptation planning under uncertainty across seasonal-to-climate timescales. When forced with seasonal forecast data, SECSF captures spatially coherent crop-risk patterns across Europe in the high-impact year 2022 and is consistent with independent monitoring, supporting its use for probabilistic Areas of Concern. Under CMIP6 scenarios, SECSF identifies the Mediterranean basin as a hotspot of maize-risk signals through mid-century, with more mixed signals in central and northern Europe.