Using remote sensing, process-based crop models, and machine learning to evaluate crop rotations across 20 million hectares in Western Australia

Remote sensing has been widely employed to identify crop types and monitor crop yields on farms. Here, we combine successive seasons of these products to identify crop rotations in each field across 20 million hectares of the Western Australian Wheatbelt. We used the APSIM crop model to define the starting soil water, temperature stresses, biomass, and crop yield to characterize the prevailing agro-environment of that field. These remote sensing data and APSIM crop modeling outputs were then combined, with machine learning, to predict the effect of the complex interaction between agro-environment and crop rotation on wheat yield. Predictions from machine learning are employed to evaluate the benefits or otherwise of crop rotation across Western Australia for every field in the study region. In general, if fields subjected to a wheat-cereal rotation were instead subjected to a wheat-canola rotation, then 68% of these fields were predicted to experience a yield increase of between 0 and 1850 kg ha-1. However, only 28% of fields planted to canola were predicted to have a yield benefit of 200 kg ha-1 or more on the following wheat crops. On average, annual pastures generated a slight yield penalty of 47 kg ha-1 to the following wheat crop. The findings from this study, using crop models, remote sensing, and machine learning, indicate that the benefits of break crops and pastures to farmers is less than the 400 to 600 kg ha-1 benefit commonly reported from field experiments. These management insights could underpin the development of future decision aids or agricultural digital twins for crop management decisions such as crop rotation planning. The approach provides farmers with tangible insights about their production using outputs from crop-based remote sensing and crop modeling.