Comparison of machine learning and deep learning models for the assessment of rondo wine grape quality with a hyperspectral camera

Abstract

Hyperspectral images provide rich spectral/spatial data and have shown remarkable performance in precision viticulture. Standardized data-processing methods are necessary to reduce the dimensionality and to identify powerful wavelengths. Toward this goal, we evaluated and compared the performance and novel-wavelength-identification ability of five renowned machine learning models: the linear models Ridge and LASSO, a 1D + 2D convolutional neural network (1D +2D CNN) non-linear model, a gradient boosting decision tree (GBDT) using XGBoost as an ensemble model, and an explainable boosting machine (EBM) followed by support vector regression (SVR) as a hybrid model. The model evaluations were conducted using leave-one-out cross-validation (LOOCV) as we sought to clarify the best-fitted machine learning model. Our results demonstrated that Ridge, LASSO, showed better performance with relatively high accuracy but were weak as a wavelength identifier. GDBT-XGBoost showed considerable prediction power and wavelength identification. EBM-SVM emerged as the most powerful model, demonstrating exceptional stability and clear wavelength classification even for destructive measurements under varying environmental stresses across Rondo's growth stages. The combined approach of 1D + 2D CNN algorithms was advantageous to handle the dynamic shapes of spectral curve and horizontal shift of the wavelengths obtained from outdoor data acquisition, and notably, it showed the highest accuracy to predict the brix and pH of wine grapes for both indoor and outdoor sensings. But the combined effects of 1D and 2D CNN algorithms were difficult to clarify the importance of spectral features for the brix and pH prediction. The integrated machine learning models with dimensionality reduction, and proper image acquisition can increase the model's accuracy. The common absorption peaks were observed in the near-infrared region of 700 nm and 900 nm. Those wavelengths should be considered for the development of low-cost sensing platforms with fewer bands. Wavelengths over 900 nm are also important to develop outdoor sensing platforms.