An approach to forecast grain crop yield using multi-layered, multi-farm data sets and machine learning

Abstract

Many broadacre farmers have a time series of crop yield monitor data for their fields which are often augmented with additional data, such as soil apparent electrical conductivity surveys and soil test results. In addition there are now readily available national and global datasets, such as rainfall and MODIS, which can be used to represent the crop-growing environment. Rather than analysing one field at a time as is typical in precision agriculture research, there is an opportunity to explore the value of combining data over multiple fields/farms and years into one dataset. Using these datasets in conjunction with machine learning approaches allows predictive models of crop yield to be built. In this study, several large farms in Western Australia were used as a case study, and yield monitor data from wheat, barley and canola crops from three different seasons (2013, 2014 and 2015) that covered ~ 11 000 to ~ 17 000 hectares in each year were used. The yield data were processed to a 10 m grid, and for each observation point associated predictor variables in space and time were collated. The data were then aggregated to a 100 m spatial resolution for modelling yield. Random forest models were used to predict crop yield of wheat, barley and canola using this dataset. Three separate models were created based on pre-sowing, mid-season and late-season conditions to explore the changes in the predictive ability of the model as more within-season information became available. These time points also coincide with points in the season when a management decision is made, such as the application of fertiliser. The models were evaluated with cross-validation using both fields and years for data splitting, and this was assessed at the field spatial resolution. Cross-validated results showed the models predicted yield relatively accurately, with a root mean square error of 0.36 to 0.42 t ha⁻¹, and a Lin’s concordance correlation coefficient of 0.89 to 0.92 at the field resolution. The models performed better as the season progressed, largely because more information about within-season data became available (e.g. rainfall). The more years of yield data that were available for a field, the better the predictions were, and future work should use a longer time-series of yield data. The generic nature of this method makes it possible to apply to other agricultural systems where yield monitor data is available. Future work should also explore the integration of more data sources into the models, focus on predicting at finer spatial resolutions within fields, and the possibility of using the yield forecasts to guide management decisions.