Development of a two-step cross-cultivar grape quality prediction model using outdoor hyperspectral imaging in three grape cultivars: Rondo, Zweigelt, and Kerner

Abstract

Sensing technologies that are independent of berry-color variation across grape cultivars, particularly under varying environmental stress conditions, are crucial for advancing precision viticulture. This study developed a two-step cross-cultivar quality prediction model for three grape cultivars with red or white berries—Rondo, Zweigelt, and Kerner—using an outdoor hyperspectral camera (HSC). We introduce a seamless soft-clustering approach that groups spectral characteristics without prior cultivar labels, using a coefficient-boosting method integrated with two dimensionality-reduction techniques: principal component analysis (PCA) and uniform manifold approximation and projection (UMAP). For Brix prediction across the three grape cultivars, the R 2 values were 0.72 (normal model), 0.75 (PCA-boosted), and 0.83 (UMAP-boosted); for pH prediction, the values were 0.70, 0.75, and 0.82, respectively. UMAP-boosted models consistently outperformed the normal model (without boosting), and PCA-boosted models also improved performance. While both the normal and PCA-boosted models improved performance, the UMAP-boosted models demonstrated superior accuracy for Brix and pH across both single- and cross-cultivar settings suggesting that UMAP more effectively captures the local and global structures of hyperspectral data, accounting for the complex reflectance curve variations between different cultivars. The primary UMAP-boosted model outperformed six mainstream machine learning and deep learning models—including PLSR, SVR, and NN—achieving the highest R 2 values and lowest RMSE. This framework enables automated clustering and enhances Brix and pH prediction from outdoor hyperspectral data under challenging field conditions. Future work should adapt the sensing platform for varying sunlight intensity and diverse geographical regions.