Spectral evolution characteristics of mild bruising in Xinjiang prunes based on visible and near-infrared spectroscopy

Xinjiang prunes are one of the most favorite fruits in western China. It is often required to timely identify mild bruising in Xinjiang prunes during postharvest sorting and transportation. Internal damage can also accumulate over time, leading to a decline in fruit quality. This study aims to monitor the temporal spectral pattern evolution of the mild bruise in Xinjiang prunes using visible and near-infrared reflectance spectroscopy (Vis-NIR). (1) Mildly bruised samples were prepared using a self-designed collision device. A comparison was then performed on reflectance and transmittance acquisition modes. Reflectance Vis-NIR was selected as the spectral acquisition. According to a standard white-reference peak control and a one-factor-at-a-time strategy, key parameters were optimized for the spectral repeatability, including integration time, source–sample distance, spectral smoothing, and measurement posture. (2) Reflectance spectrum was collected at 1, 6, 24, and 48 h after bruising, with intact fruits set as the control group. Mean spectral analysis was performed after acquisition. Dimensionality-reduction visualization was conducted using principal component analysis (PCA) and uniform manifold approximation and projection (UMAP). Spectral differences were obtained over time points. Support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) classifiers were then developed, in terms of recognition performance and generalization. The results showed that the reflectance of mildly bruised samples varied significantly over time in the 700~950 nm region, with the most pronounced difference from the control at 24 h. After multiplicative scatter correction (MSC), detrending (DT), and continuous wavelet transform (CWT) preprocessing, the SVM achieved high accuracies above 92.00%, and the CWT–uninformative variable elimination–SVM (CWT-UVE-SVM) model reached an accuracy of 96.67% at 24 h. Once trained on mildly bruised samples, the model maintained excellent generalization in transfer validation on moderately and severely bruised samples, with the 99.17% accuracy for severe bruising. In addition, a time label increased the accuracy from 94.17% to 97.78%. The finding can provide strong references and data support for the postharvest dynamic detection and sorting strategy.