Coupling agricultural system models with machine learning to facilitate regional predictions of management practices and crop production

Process-based agricultural system models are a major tool for assessing climate-agriculture-management interactions. However, their application across large scales is limited by computational cost, model uncertainty, and data availability, hindering policy-making for sustainable agricultural production at the scale meaningful for land management by farmers. Using the Agricultural Production System sIMulator (APSIM) as an example model, the APSIM model was run for 101 years from 1980 to 2080 in a typical cropping region (i.e., the Huang-Huai-Hai plain) of China. Then, machine learning (ML)-based models were trained to emulate the performance of the APSIM model and used to map crop production and soil carbon (which is a key indicator of soil health and quality) dynamics under a great number of nitrogen and water management scenarios. We found that ML-based emulators can accurately and quickly reproduce APSIM predictions of crop yield and soil carbon dynamics across the region under different spatial resolutions, and capture main processes driving APSIM predictions with much less input data. In addition, the emulators can be easily and quickly applied to identify optimal nitrogen management to achieve yield potential and sequester soil carbon across the region. The approach can be used for modelling other complex systems and amplifying the usage of agricultural system models for guiding agricultural management strategies and policy-making to address global environmental challenges from agriculture intensification.

Automated summary

This article introduces a method that uses machine learning to emulate process-based agricultural system models and discusses its application in agricultural management and policy-making. By simulating crop production and soil carbon dynamics in a typical cropping region, the researchers found that ML-based emulators can accurately and quickly reproduce the predictions of the original model, capturing main processes with less input data. This approach can guide agricultural management strategies and address global environmental challenges arising from agricultural intensification.