Early bruise detection in Chinese pear-leaf crabapple based on hyperspectral imaging and deep learning

Early and accurate detection of fruit damage is essential for preserving the postharvest quality of Chinese pear-leaf crabapples. This study developed a rapid, non-destructive method for early bruise detection using hyperspectral imaging (400 - 1000 nm). Hyperspectral images were acquired from fruits subjected to different impact energies (0, 0.14, 0.27, and 0.41 J) over a 7-day storage period. Spectral analysis revealed that specific wavelength regions were closely associated with bruise-induced physiological changes. Partial least squares discriminant analysis, support vector machine discriminant analysis, and a one-dimensional convolutional neural network (1D-CNN) were constructed and compared. The 1D-CNN achieved the best performance, with an overall validation accuracy of 89.9% using full-spectrum data and 91.3% using selected feature wavelengths (520 - 560, 625 - 695, and 710 - 760 nm) at 2 h post-impact. This simplified feature-wavelength model maintained high accuracy while substantially reducing computational complexity, demonstrating its potential for real-time sorting applications. The results indicate that hyperspectral imaging coupled with 1D-CNN offers a robust and efficient tool for non-destructive, early-stage bruise detection in Chinese pear-leaf crabapples.