Accuracy of carrot yield forecasting using proximal hyperspectral and satellite multispectral data
Suarez, LA and Robson, A and McPhee, J and O'Halloran, J and van Sprang, C, Accuracy of carrot yield forecasting using proximal hyperspectral and satellite multispectral data, Precision Agriculture pp. 1-23. ISSN 1385-2256 (2020) [Refereed Article]
Proximal and remote sensors have proved their effectiveness for the estimation of several biophysical and biochemical variables, including yield, in many different crops. Evaluation of their accuracy in vegetable crops is limited. This study explored the accuracy of proximal hyperspectral and satellite multispectral sensors (Sentinel-2 and WorldView-3) for the prediction of carrot root yield across three growing regions featuring different cropping configurations, seasons and soil conditions. Above ground biomass (AGB), canopy reflectance measurements and corresponding yield measures were collected from 414 sample sites in 24 fields in Western Australia (WA), Queensland (Qld) and Tasmania (Tas), Australia. The optimal sensor (hyperspectral or multispectral) was identified by the highest overall coefficient of determination between yield and different vegetation indices (VIs) whilst linear and non-linear models were tested to determine the best VIs and the impact of the spatial resolution. The optimal regression fit per region was used to extrapolate the point source measurements to all pixels in each sampled crop to produce a forecasted yield map and estimate average carrot root yield (t/ha) at the crop level. The latter were compared to commercial carrot root yield (t/ha) obtained from the growers to determine the accuracy of prediction. The measured yield varied from 17 to 113 t/ha across all crops, with forecasts of average yield achieving overall accuracies (% error) of 9.2% in WA, 10.2% in Qld and 12.7% in Tas. VIs derived from hyperspectral sensors produced poorer yield correlation coefficients (R2 < 0.1) than similar measures from the multispectral sensors (R2 < 0.57, p < 0.05). Increasing the spatial resolution from 10 to 1.2 m improved the regression performance by 69%. It is impossible to non-destructively estimate the pre-harvest spatial yield variability of root vegetables such as carrots. Hence, this method of yield forecasting offers great benefit for managing harvest logistics and forward selling decisions.