Data Requirements for Forecasting Tree Crop Yield - A Macadamia Case Study

Early tree crop yield forecasts are valuable to industry and to growers, as they inform improved harvest logistics, forward selling, insurance and marketing strategies. Previous work has demonstrated the utility of weather and particularly remote sensing data to forecast tree crop yield at the orchard block scale. In this work, such data were aggregated spatially to block boundaries, and temporally at quarterly intervals. Yield prediction models were trained with a large set of grower-supplied yield data (more than 10 years, 20 orchards, 200 blocks across the Australian growing regions, for a total of 1156 yield records). Yields were forecast three months before harvest begins, and were compared to actual yields. Errors were typically around 10% and 23% at the regional and block levels respectively. Errors in 2020 were higher in non-irrigated regions due to an extreme drought in east Australia. Models were able to describe much of the variability of yields even for orchards not included in the training data, but block-level prediction errors increased by 4.1% in this case. Bootstrap sampling was used to investigate data requirements. At least 400-500 training data points was needed to minimize prediction errors. Weather data alone did not produce satisfactory accuracy, fusing weather and remote sensing data produced the best results. Including predictor data from all 8 quarterly periods from the 2 years before harvest proved a good strategy. These results demonstrate the potential of tree crop forecasting using public spatio-temporal datasets, give guidance on data requirements and identify areas for further work.