Integrating Crop Simulation and Machine Learning Models to Improve Crop Yield Prediction

Abstract

Accurate crop yield prediction can contribute to the decision-making in farm management. In this paper, we propose a crop yield prediction model which integrates the outputs from a crop simulation model (APSIM). These outputs are included as features in the machine learning (ML) models to improve the transparency and an ensemble model then is designed for prediction. The proposed model is applied to predict the corn yield of 12 US Corn Belt States. RRMSE of 9.34%, 8.99%, and 9.52% was achieved after evaluating the model at the county level for the years 2018, 2019, and 2020 respectively. The proposed model has similar accuracy as the top-performing state-of-art prediction models while providing interpretability. Additional analyses have been conducted to understand the source of model errors. In this analysis, it has been identified that counties and crop reporting districts (CRD) with high cropland ratio have a lower error due to the high data availability in those regions.