Combining Feature Extraction Methods and Categorical Boosting to Discriminate the Lettuce Storage Time Using Near-Infrared Spectroscopy

Lettuce is a kind of nutritious leafy vegetable. The lettuce storage time has a significant impact on its nutrition and taste. Therefore, to classify lettuce samples with different storage times accurately and non-destructively, this study built classification models by combining several feature extraction methods and categorical boosting (CatBoost). Firstly, the near-infrared (NIR) spectral data of lettuce samples were collected using a NIR spectrometer, and then they were preprocessed using six preprocessing methods. Next, feature extraction was carried out on the spectral data using approximate linear discriminant analysis (ALDA), common-vector linear discriminant analysis (CLDA), maximum-uncertainty linear discriminant analysis (MLDA), and null-space linear discriminant analysis (NLDA). These four feature extraction methods can solve the problem of small sample sizes. Finally, the classification was achieved using classification and regression trees (CARTs) and CatBoost, respectively. The experimental results showed that the classification accuracy of NLDA combined with CatBoost could reach 97.67%. Therefore, the combination of feature extraction methods (NLDA) and CatBoost using NIR spectroscopy is an effective way to classify lettuce storage time.