Non-destructive and on-site estimation of grape total soluble solids by field spectroscopy and stack ensemble learning

Accurately estimating the total soluble solids (TSS) of berries with a non-destructive method is crucial for wine grape growers if wine quality improvements are to be made. At present, the methods employed with the best statistical results are implemented under stable lab conditions, using spectroscopic analysis in the visible-near infrared (VNIR) region. This study explores using field spectroscopy to estimate the TSS of berries directly in the vineyard. A portable visible-near infrared-shortwave infrared (VNIR-SWIR) spectroradiometer measured the reflectance data of grape berries in the 350–2500 nm spectral region. A large in-field multi-season spectral database ( n = 1830) over two years (2023–2024) from three ‘Pinot Noir’ commercial vineyards were selected to develop spectral-region specific (VNIR, SWIR or VNIR-SWIR) machine learning models. Different machine learning modeling pipelines were built using data collected from 2023 and validated using data from 2024 to predict grape TSS based on in-field spectral databases. Subsequently, the performance of using stack ensemble learning (ES) to predict grape TSS was evaluated and compared with three commonly used methods: K-nearest neighbors (KNN), random forest regression (RFR), and support vector regression (SVR). The result on the independent test set showed that, the ES model based on MSC+SG+ 1D spectral data, in the VNIR-SWIR region provided the highest prediction accuracy for grape TSS value, with a coefficient of determinations ( R 2 ) of 0.815, root mean square error (RMSE) of 1.131 °Brix, and a ratio of performance to deviation (RPD) of 2.236, with a Lin’s concordance correlation coefficient (CCC) of 0.897. This study demonstrated the potential of using an ES model to assess the grape TSS rapidly and non-destructively from field spectroscopy data. • Total soluble solids in grape berries were quantified using a large field spectroscopy database. • The performance of stack ensemble model was better than other machine learning models. • Model generalization was test on new harvest season successfully.