A big data approach to predicting crop yield

Many broadacre farmers have a time series of crop yield monitor data for their paddocks which are often augmented with additional spatial data such as gamma radiometrics surveys or ECa (apparent soil electrical conductivity) from an electro-magnetic induction survey (EMI). In addition there are now readily available national and global datasets which can be used to represent the crop-growing environment. Rather than analysing one paddock at a time, there is an opportunity to explore the value of combining data over multiple paddocks 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 we explored this approach with a particular emphasis on the forecasting ability of the models based on pre- and mid-season information from predictor variables. Several large farms in Western Australia were used as a case study. Yield from wheat, barley and canola crops from 3 different seasons that covered ~15,000 hectares in each year were used. The yield data was processed to a 10 m grid, and for each observation we built an associated space-time cube of predictors. This consisted of grower collected data such as EM and gamma radiometrics surveys, and nationally available data such as MODIS NDVI, and rainfall. Random Forest models were used to predict crop yield of wheat, barley and canola using the space-time cube. Three 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 using cross-validation based on paddocks and years and this was assessed at the spatial resolution of the paddock. The models performed better as the season progressed, largely because more information about within-season data became available (e.g. rainfall). Cross-validated results showed the models predicted yield very accurately, with an RMSE of 0.36 to 0.42 t/ha, and an LCCC of 0.89 to 0.92 at the paddock resolution. The more years of yield data that were available for a paddock, the better the predictions were. The generic nature of this method makes it possible to apply to other agricultural systems where yield monitor data is available. A data-driven approach to predicting crop yield as an alternative to using mechanistic models has several advantages. Future work should explore integration of more data sources into the models and focus on using the models to inform management decisions such as fertiliser applications