Classification of kiwifruit storage day based on hyperspectral imaging and convolutional neural networks

Accurate monitoring of kiwifruit storage time is crucial for reducing postharvest losses. A method combining hyperspectral imaging (HSI) and convolutional neural networks (CNN) was developed for storage day identification. Three cultivars— ‘Xuxiang’, ‘Cuixiang’, and ‘Hongyang’ —were imaged at 0, 3, 6, and 9 days of storage. Principal component analysis extracted the first three components, and clustering plots highlighted spectral differences. Competitive adaptive reweighted sampling (CARS) and sequential forward selection (SFS) were employed to select characteristic bands. Classification models based on partial least squares discriminant analysis, least squares support vector machine, random forest, and CNN were constructed. The CARS–CNN model achieved 100 % accuracy on the calibration set and prediction accuracies of 95.0 %, 97.5 %, and 97.5 % for three cultivars, respectively, outperforming other models. These results validate the reliability of HSI–CNN for classifying kiwifruit storage day and support the development of online fruit quality monitoring systems for storage time management.