Vineyard proximal sensing using multispectral imaging to evaluate grape ripening and quality traits using artificial neural networks modeling

Multispectral proximal sensing from vineyards at the fruit zone, combined with machine learning modeling, can serve as a reliable and accurate tool for estimating grape quality traits and ripeness. This study developed machine learning modeling using a multispectral camera with six bands: blue (475 nm), green (560 nm), red (668 nm), red edge (717 nm), near-infrared (842 nm), and thermal (8-14 μm) to estimate important chemometric properties. The physicochemical variables included total soluble solids (TSS), acidity, pH, and phenols in Cabernet, Merlot, and Parellada, along with anthocyanins for the red cultivars (Cabernet and Merlot). The modeling strategies included eleven artificial neural networks (ANN) models: two general models for red cultivars (Cabernet and Merlot) and Parellada, focusing on total soluble solids (TSS), acidity, pH, and important phenolic compounds, with anthocyanins, included only for red cultivars, plus nine individual models for each of these physicochemical compounds. The ANN models demonstrated strong performance, with R 2 values of 0.89, 0.81, and 0.83 for TSS, acidity, and pH in Cabernet and Merlot and 0.93, 0.97, and 0.94 for these parameters in Parellada. These variables are closely linked to grape ripeness. The general models for red cultivars and Parellada that describe all physicochemical parameters achieved R 2 values of 0.79 and 0.70, respectively. Importantly, none of the models exhibited statistical signs of overfitting or underfitting. Implementing these ANN models could enable the use of multispectral imagery to assess grape quality traits, advancing precision viticultural practices.