Leaf and cluster spectral signatures reveal trait-dependent prediction performance for grapevine cluster architecture and juice quality

Grapevine cluster architecture is a key selection target in breeding programs because it influences disease susceptibility, yield stability and juice quality. High-throughput phenotyping offers a rapid and non-destructive approach to capture biochemical and structural variation in these traits, yet the influence of plant organ reflectance and data partitioning strategies on trait prediction remains poorly understood. In this study, we evaluated how hyperspectral reflectance from different grapevine organs contributes to the prediction of cluster architecture and juice quality traits in two clonal populations of Riesling and Pinot. Using partial least squares regression (PLSR), we assessed the prediction accuracy of eight cluster architecture and six juice quality traits under two data partitioning strategies. Models based on cluster reflectance outperformed those using dry leaf reflectance for most traits, except for pH. Partitioning the dataset by cluster type increased trait variance and improved predictions for number of berries (R2 = 0.53), berry diameter (R2 = 0.79), and total acidity (R2 = 0.48). Visible, red-edge and NIR spectra were most informative regions to predict the traits studied. Together, our results highlight the importance of organ-specific data and appropriate calibration strategies to improve phenomic models for the development of scalable proxies for grapevine improvement.