Prediction of Wheat Yield and Protein Using Remote Sensors on Plots—Part I: Assessing near Infrared Model Robustness for Year and Site Variations

Abstract

Validation of reflectance-based prediction models for plant properties is often performed on just one or two years of data. Hence, we aimed to perform a more comprehensive study regarding the validation of prediction models for grain yield and protein concentration. A FieldSpec3 portable field spectroradiometer was used to measure canopy reflectance in spring wheat. Spectral reflectance data were collected from three different experimental locations in up to four different years during the period 2007–2010, so that seven unique site years were included, comprising, altogether, 976 individual plots. Several datasets had moderate to severe lodging, which had a markedly negative influence on the prediction results. To correct for this problem, a classification model for the classes “lodging” and “standing crop” was calibrated from the spectral data. The model gave a total classification accuracy of 98.3%. Prediction models for grain yield and grain protein concentration were computed by means of the recent statistical method powered partial least squares (PPLS). Models were calibrated and validated on several combinations of the spectral datasets in order to reveal spatial and temporal effects on the prediction performance. The model performance generally increased with increasing variation in the calibration data, both in time (i.e. more years included) and space (i.e. more sites included). The best model for grain yield explained 94% [root mean square error of prediction ( RMSEP) = 156g m−2] of the variance and the predictions of grain protein concentration explained 67% [ RMSEP=1.51 g dry matter (DM)100g−1] of the variance. The performance of the grain yield PPLS models was compared with that of models based on some widely used vegetation indices [normalised difference vegetation index (NDVI), modified soil adjusted vegetation index (MSAVI), red edge inflection point (REIP) and d-chl-ab]. The explained variance of the models based on vegetation indices did not exceed 55%, indicating that these models were inferior to full spectrum models. This study shows that one or two years of spectral measurement are insufficient for building fully operational models for cereal property predictions.