Cluster trait prediction using hyperspectral signatures in a population of 221 Riesling clones

Abstract

Context and purpose of the study. Cluster architecture in grapevine plays a critical role in influencing bunch microclimate, thus quality traits, including sugar content, phenolic composition, and disease susceptibility. This study focuses on exploring the use of hyperspectral spectroscopy as a high-throughput phenotyping (HTP) alternative to predict cluster architecture and quality traits by analysing spectral properties of leaves and clusters. Our goal was to develop predictive models based on spectral properties of leaves and clusters, aiming to enhance breeding selection efficiency for cluster traits in 221 clones of white Riesling.

Materials and methods. A subset of 221 Riesling clones displaying diverse cluster architecture were studied. A phenotyping pipeline was established where spectral signatures in three clusters per clone were measured using a spectroradiometer. Additionally, the berries were pressed to obtain juice and analyse its quality by means of density, pH, tartrate, malate, NOPA and the sugar free extract in the juice. In addition to cluster reflectance, leaf reflectance was collected on the same population of study and two methods were used to predict the traits, partial least squares regression (PLSR) and vegetation indices (VI).

Results. PLSR models showed higher accuracy when using cluster reflectance compared to leaf reflectance. Cluster weight prediction with cluster reflectance showed R2 = 0.31, RMSE = 54.58 g, and rachis weight predictions had R2 = 0.19, RMSE = 4.54 g. For cluster weight prediction, the key wavelengths influencing predictions were located at 532-556 nm and 678-700 nm, while those for rachis weight were located at 400-408 nm and 680-713 nm. VI models using cluster reflectance offered comparable accuracy for cluster weight (R2 = 0.31) or better (R2 = 0.28) for rachis weight. For the quality traits, sugar free extract, density and malate content were predicted using vegetation indices related to chlorophyll content and water stress with a very high accuracy (R2 = 0.7, 0.66, 0.52, respectively) using cluster reflectance. However, the prediction accuracy reduced when using leaf reflectance (R2 = 0.21, 0.12, 0.19, respectively). These results indicate the potential uses of spectroscopy to select clones based on cluster architecture and quality traits and the importance of measuring different organs of the canopy, as cluster reflectance derived models performed better than the ones using leaf reflectance.