Surrogate impact modelling for crop yield assessment

TL;DR

This paper introduces SECSF, a deep-learning framework that emulates crop growth models using minimal climate data, enabling efficient large-scale crop yield risk assessments across Europe.

Contribution

The study develops a novel deep-learning surrogate model that accurately replicates process-based crop simulations with reduced computational costs.

Findings

SECSF reproduces crop growth and harvest timing with high fidelity.

It captures interannual and spatial crop stress patterns across Europe.

SECSF identifies the Mediterranean as a persistent yield risk hotspot under climate scenarios.

Abstract

This study presents the Surrogate Engine for Crop Simulations Framework (SECSF) a group of deep-learning models that emulate the process-based ECroPS model using only daily maximum and minimum temperature and precipitation. In this study we emulate grain maize and spring barley. Trained on ERA5-forced ECroPS simulations, SECSF reproduces crop growth dynamics and harvest timing with high fidelity. Critically, SECSF extremely reduces computational costs enabling ensemble-scale inference suitable for operational pipelines. When driven by seasonal data, SECSF captures the interannual and spatial patterns of crop stress across Europe and aligns with independent monitoring, supporting its use as a probabilistic Areas of Concern indicator for early warning. Under CMIP6 SSP3-7.0 and SSP5-8.5 scenarios, SECSF consistently identifies the Mediterranean basin as a persistent hotspot of yield risk through mid-century, with central-northern Europe showing mixed signals. These results demonstrate that a streamlined, data-efficient emulator can provide robust seasonal-to-climate risk assessments at continental scale.