Predicting tilling and seeding operation times in grain production: a comparison of machine learning and mechanistic models

Abstract

Field operations management in grain production requires accurate and timely predictions of operation times for machine tasks. While machine learning (ML) is being adopted more widely in operations management, little is known about its ability to predict tilling and seeding operation times. The aim of this study was to evaluate the prediction performance of ML models for these operation times by using readily available tractor and operations data rather than dynamic environmental data. We collected data between March 2022 and August 2023 from 70 grain fields in the southwest of Germany, including variables such as tractor speed, engine speed, fuel consumption, and field geometry. Operation times exhibited high variability (coefficient of variation [CV] = 0.88). Nine ML algorithms and two conventional mechanistic models proposed by the American Society of Agricultural and Biological Engineers (ASAE EP496.3) were evaluated in a temporal external validation. Random forest (RF) models outperformed all other models, achieving a normalized root mean square error (NRMSE) of 0.215 and a coefficient of determination (R2) of 0.910. Compared to a conventional mechanistic model, the RF model reduced the mean absolute error (MAE) by 37.8 %, and enhanced the R2 by 0.107. The study results highlight the potential of our approach to predict tilling and seeding operation times in grain production without increasing the effort for data collection, offering an accessible and cost-effective solution for resource-constrained grain farming systems that experience data shortages.