Predicting within-field grain protein content at scale using agronomic and remote sensing variables, and machine learning

Abstract Purpose Grain protein content (GPC) is a key determinant of the prices that grain growers receive, but there is considerable variability within and between fields, farms, and seasons. Despite growing interest in measuring and mapping within-field GPC variability, the uptake of grain protein sensors has been slow, resulting in considerable knowledge gaps. Building a predictive model to map GPC in areas of a farm without a GPC sensor can provide growers with valuable insights for better management decisions. Methods This paper presents a data-driven, machine learning (random forest) approach to predict GPC and yield within agricultural fields using 63 paired yield and protein maps collected over four seasons (2020–2023) in Western Australia and northern New South Wales, Australia. Model performance for yield and GPC predictions using different combinations of yield, on-farm agronomic (e.g. sowing and harvest dates, cropping history, variety) and publicly-available (e.g. digital elevation model, radiometric surveys, remotely-sensed satellite imagery) spatial data layers were tested using two validation approaches: leave one Field-Year out cross validation (LOFYOCV) and two-fold cross validation (2FCV) at either a fine-resolution (30 m) or across management classes. Results The 2FCV method, which simulates interpolating GPC within fields to fill-in unsampled areas, outperformed LOFYOCV, which tested extrapolation across unsampled fields. Combining yield, agronomic, and publicly-available data layers produced the best quality predictions of GPC. Conclusion Providing growers with GPC maps can inform management decisions to optimise both yield and quality, leading to more profitable and environmentally sustainable production systems.