Evaluation of plot level aggregation statistics for spectral grain yield prediction

Spectral imaging has become popular through the advancement of UAV platforms and is available at increased spatial resolution, especially from RGB and multispectral (MS) sensors. While collecting high resolution spectral data on the plot level is important for accounting for within-plot heterogeneities, reference data is commonly available only on the plot level, thus aggregated to a single value as in the case of grain yield (GY). Therefore, spectral data within plots are generally aggregated (downsampled) to the plot level with the aim of creating a common level of spatial resolution. This holds the advantage of data reduction while potentially losing valuable information. Therefore, this study compares the performance of different aggregation outputs (percentiles) of multispectral and RGB UAV-based data for GY prediction of winter wheat. Correlations varied depending on spectral data, measurement dates and trials. Overall, correlations were found to decrease for percentiles below 20 % and above 85 %, respectively, but to be relatively stable for percentiles in between. In addition, the correlations obtained from mean and median values did not reveal a substantial difference. Results confirm the common choice of median or mean values while lowest and highest percentiles should be avoided irrespective of spectral parameter and growth stage.