Sweet corn yield prediction using machine learning models and field-level data

Abstract

The advent of modern technologies, acquisition of large amounts of crop management and weather data, and advances in computing are reshaping modern agriculture. These advancements have unlocked the power of data by providing valuable insights and more accurate yield predictions. This study utilizes a historic US sweet corn dataset to: (a) evaluate machine learning model performances on sweet corn yield prediction and (b) identify the most influential variables for crop yield predictions. The sweet corn data comprised field-level data for over a quarter-century period (1992–2018) from two primary commercial sweet corn production regions for processing, namely the Upper Midwest and the Pacific Northwest. Several machine learning models were trained to predict field-level sweet corn yield from 67 variables of crop genetics, management, weather, and soil factors. The random forest model outperformed all trained models with the lowest RMSE (3.29 Mt/ha) and the highest Pearson’s correlation coefficient (0.77) between predicted and observed yields. Variable importance plots revealed the top three most influential predictor variables as year (time), location (space), and seed source (genetics). Season long total precipitation and average minimum temperature during anthesis were the two most important weather variables in yield prediction. This is the first report of using fine-scale (time and space) crop data and advanced data analytics to leverage insights into commercial sweet corn production.