In recent years, high-throughput plant phenotyping (HTPP) platforms in the laboratory or directly in the field have multiplied. The use of optical instruments such as NIR Spectroscopy or hyperspectral imaging tends to increase offering the possibility of generating large quantities of data in an automatic way. This represents a potential application for plant breeding. However, the large amount of data often called massive data leads to difficulties in managing and analyzing them. Processing this massive amount of spectral data is challenging. Partial-Least-Squares (PLS) is the most widely used method for predicting biochemical variables based on a linear relationship with spectra. When dealing with large amounts of spectral data, complex structures and non-linear relationships appear. This can compromise linear regression approaches. Recently, a method called 'parSketch-PLS' has recently been proposed (Metz et al. 2020)to implement a local approach to a large volume of spectral data. This method combines a fast neighborhood search method (parSketch) with PLS. This method is valid to predict categorical variables by adding a Discrimination-Analysis (DA) step like PLS with PLS-DA. In this presentation, we propose to compare parSketch-PLS-DA with the reference method PLS-DA in a context of varietal discrimination. For this purpose, a spectral database was formed by collecting 1,300,000 spectra from hyperspectral images of leaves of four different sunflower genotypes. Results show that the prediction model obtained by PLSDA has a classification error close to 23% on average across all genotypes. ParSketch-PLSDA method outperforms PLS-DA by greatly improving prediction qualities by 10%. These results are encouraging and allow us to anticipate the future bottleneck related to the generation of a large amount of data from phenotyping generating complex structures and non-linear relationships.