Hyperspectral Imaging and Optimized Convolutional Neural Network for Quality Assessment of Sweetpotato

The recent integration of hyperspectral imaging (HSI) with deep learning techniques has emerged as an innovative strategy for precise predictive analysis in agricultural and biological domains. However, the effectiveness of these techniques highly depends on their appropriate optimization. This study combines HSI and Convolutional Neural Network (CNN)-based regression for predicting dry matter (DM) in different varieties of sweetpotatoes. Spectral data were extracted from images captured using a visible near-infrared hyperspectral imaging system (400-1000 nm). The hyperparameters of the CNN were optimized utilizing Bayesian Optimization (BO). The optimized CNN showed a 10.71% improvement in R²p and a 46.61% improvement in RPD over the Partial Least Squares Regression (PLSR) model. This achievement highlights the efficiency and growing importance of applying hyperspectral imaging in conjunction with deep learning for advanced predictive analyses.