Block-level macadamia yield forecasting using spatio-temporal datasets
James Brinkhoff, Andrew Robson
Abstract
Early crop yield forecasts provide valuable information for growers and industry to base decisions on. This work considers early forecasting of macadamia nut yield at the individual orchard block level with input variables derived from spatio-temporal datasets including remote sensing, weather and elevation. Yield data from 2012–2019, for 101 blocks belonging to 10 orchards, was obtained. We forecast yield on each test year from 2014–2019 using models trained on data from years prior to the test year. Forecasts are generated in January, for the coming harvest in March–September. A linear model using ridge regularized regression produced consistently good predictions compared with other machine learning algorithms including lasso, support vector regression and random forest. Adding meteorological variables offered little improvement over using only remote sensing variables. The 2019 forecast root mean square error at the block level was 0.8 t/ha, and mean absolute percentage error was 20.9%. When block level predictions were aggregated across the multiple orchards per region, production prediction errors were between 0–15% from 2016–2019. The ridge regression model can be easily implemented in GIS platforms to deliver block-level yield forecast maps to end users.