Abstract

As farming practices become increasingly automated, the quantity of high resolution on-farm production information grows exponentially and so does the need for high-throughput computing solutions to aid management. High resolution (5 m) wheat yield forecasting is presented here using two machine learning approaches: (a) Bootstrapped Regression Trees (BRR) where predictions are pixel-wise and (b) Convolutional Neural Networks (CNN) where predictions use neighbouring pixels. This study focused on three aims. First, to compare the two approaches in a yield forecasting task that included publicly available data and on-farm gathered yield data. Second, to study any benefit of adding more layers of information in the modelling process, e.g. proximal soil sensing surveys. Third, to evaluate the value of including information from contiguous neighbouring fields in order to forecast within-field wheat yield at harvest. Results showed that BRR modelling using publicly available Sentinel data with the addition of local electromagnetic induction surveys or gamma radiometric surveys produced the best forecasts as determined by the classical performance metrics. The results from the CNN models improved with the addition of publicly available data from neighbouring fields and produced a spatial distribution pattern that most closely resembled the actual yield data. Within-field yield forecasting using machine learning techniques and publicly available data shows good potential, and this work suggests that the choice of yield forecasting methodology may depend on the type and extent of spatial data that is available for use in forecasting.