Multi temporal multispectral UAV remote sensing allows for yield assessment across European wheat varieties already in tillering stage

Abstract

AbstractHigh throughput field phenotyping techniques employing multispectral cameras allow to extract a variety of variables and features to predict yield and yield related traits, but little is known about which types of multispectral features may allow to forecast yield potential in the early growth phase. In this study, we hypothesized that the best features for predicting yield in an early stage might be different from the best predictors for the late growth stages. Based on a variety testing trial of 19 European wheat varieties in 2021, multispectral images were taken on 19 dates ranging from tillering to harvest by an unmanned aerial vehicle measuring reflectance in five bands, including visible bands, Red-edge and the near-infrared (NIR). Orthomosaic images were created, and then the single band reflectances, vegetation indices (VI) and texture features (TF) based on a gray level correlation matrix (GLCM) were extracted. We evaluated the performance of these three types of features for yield prediction and classification at different growth stages by, i) using features on each of the measurement dates, ii) smoothing features across the 19 dates, and iii) combining features across the directly adjacent dates, in combination with the random forest models. Our results showed that, for most features, measurements at the flowering stage showed the best performance and the Red reflectance was able to predict yield with a RMSE of 47.4 g m-2(R2= 0.63), the best VI was NDRE predicting yield with a RMSE of 47.9 g m-2(R2= 0.63), the best TF was contrast predicting yield with a RMSE of 57.2 g m-2(R2= 0.46) at the booting stage. Combining dates improved yield prediction in all dates and made the prediction errors more stable across dates. Rather than the Red-edge band, visible bands especially the Red band enabled to distinguish between the high- and low-yielding varieties already in the tillering stage, with a total accuracy of 76.7%. The study confirms our hypothesis and further implies that, in the early stages, the visible bands may be more effective than Red-edge bands in assessing the yield potential in a range of testing varieties.