Estimating wheat yields in Australia using climate records, satellite image time series and machine learning methods

Closing the yield gap between actual and potential wheat yields in Australia is important to meet the growing global demand for food. The identification of hotspots of the yield gap, where the potential for improvement is the greatest, is a necessary step towards this goal. While crop growth models are well suited to quantify potential yields, they lack the ability to provide accurate large-scale estimates of actual yields, owing to the sheer quantity of data they require for parameterisation. In this context, we sought to provide accurate estimates of actual wheat yields across the Australian wheat belt based on machine-learning regression methods, climate records and satellite image time series. Out of nine base learners and two ensembles, support vector regression with radial basis function emerged as the single best learner (root mean square error of 0.55 t ha(-1) and R-2 of 0.77 at the pixel level). At national scale, this model explained 73% of the yield variability observed across statistical units. Benchmark approaches based on peak Normalised Difference Vegetation Index (NDVI) and on a harvest index were largely outperformed by the machine-learning regression models (R-2 < 0.46). Climate variables such as maximum temperatures and accumulated rainfall provided additional information to the 16-day NDVI time series as they significantly improved yield predictions. Variables observed up to and around the flowering period had a particularly high predictive power with additional information gained from data during grain filling. We further showed that, while all models were sensitive to a reduction of the training set size, a large majority had not reached saturation with a data set of 125 fields (2000 pixels). This indicates that additional training data are likely to further improve the skill of the models. We estimated that observations from 75 fields (1200 pixels) are required for the best single model to reach an R-2 of 0.7. We contend that machine-learning regression methods applied to climate and satellite image time series can achieve reliable crop yield monitoring across years at both the pixel and the country scale. The resulting yield estimates meet the accuracy requirements for mapping the yield gap and identifying yield gap hotspots which could be targeted for further work by agricultural researchers and advisers.