Prediction models for grain yield based on remote sensing data are commonly shown to perform reasonably well for one single cropping season. The model performances often drop, however, when data from more years is included. This may be caused by biased data, resulting from diverging growth conditions from year to year, which affects the canopy reflectance spectra. In this study we tested if data fusion of reflectance data and meteorological variables could be used for bias and skewness correction, in order to achieve more robust models for predicting cereal yields. To do so, three similar field trials with spring wheat were established during the years 2007-2010. The total dataset comprises 976 plots, covering four cropping seasons. Using a FieldSpec3 spectroradiometer, reflectance spectra was obtained from each plot. At harvest, grain yield was measured. Models for yield prediction were then created by using reflectance data and Powered Partial Least Squares (PPLS). To correct for bias, we used meteorological data obtained from the Norwegian cereal yield prognoses program, comprising 24 aggregated variables from a network of weather stations. When using reflectance data alone, model bias was in the range of -14 gm^-2 to 160 gm^-2, and it depended largely on which year the yield model was validated against. Subsequently, we augmented the analysis with the 24 meteorological variables. Using these variables together with the reflectance data, the model biases and skewnesses were almost completely removed. The current study shows that data fusion between reflectance data and meteorological data may be a feasible way of correcting bias problems in wheat yield predictions.