Hyperspectral Imaging Combined With Image Fusion Features and Machine Learning to Discriminate Different Origins, Grades, and Shelf‐Life of Oranges

ABSTRACT This study reports the first application of hyperspectral feature fusion technology combined with machine learning algorithm to qualitatively distinguish oranges from different growing regions, grades, and shelf‐lives. For origin traceability, a partial least squares discriminant analysis (PLS‐DA) model outperformed competing methods, including deep learning models. Based on a hyperspectral dataset of 146 oranges from four different producing regions in China, with the training and prediction accuracy reaching 100% and 97.1%, respectively. Furthermore, 224 oranges were assigned as either high or lower grade based on their chemical quality data (soluble solid content, vitamin C, and sugar–acid ratios). PLS‐DA and random forest (RFT) models, incorporating spectral processing and image fusion, established a grading mechanism with 94.4% prediction accuracy. Finally, for shelf‐life determination (0, 10, and 20 days), a PLS‐DA model utilizing spectral, image, and spectrum—image feature fusion data also achieved 94.4% accuracy. In addition, the fused characteristic bands combined with spectral analysis indicate that chlorophyll, sugar, and moisture content are the key chemical substances that distinguish different types of oranges. The results of this study provide strong technical support and theoretical evidence for applying hyperspectral fusion technology to other similar fruit evaluations.