No abstract

Fusion of NIR spectra and weather data by means of canonical powered partial least-squares (CPPLS) was used to correct yield prediction models of spring wheat grain yields. The data comprised seven unique site years and 976 individual plots. In Part I of the study, in which models based on NIR spectra alone were used, we achieved a high degree of explained yield variance (up to 94%), but encountered large problems with bias and skewness in the computed regression models. We hypothesised that weather variation between seasons is the governing process behind this behaviour. In the present part of the study (Part II), the bias and skewness problems were significantly reduced by the inclusion of aggregated weather variables as additional predictors. Average bias and skewness of the final predictions corresponded to reductions of 38% and 23%, respectively. The weather data used are available about four weeks prior to harvest, which implies that the approach would allow early yield prediction. In conclusion, this study shows the potential for improving NIR spectrabased prediction models by including season-specific information, such as weather data, which is becoming more readily available through improvements in meteorological services.