Non-destructive estimation of the bruising time in kiwifruit based on spectral and textural data fusion by machine learning techniques

Detection of kiwifruit bruising time is one of the essential indicators for reducing postharvest losses and assessing internal quality. To investigate if the bruised time of kiwifruit could be non-destructively detected, hyperspectral imaging (HSI) technology was used to acquire hyperspectral images of kiwifruit from two different varieties at various bruising time. 70 kiwifruit samples from each of the two varieties were included in the study, with a total of 490 (7 x 70) hyperspectral images collected for each variety across seven bruising time. The spectral feature of the bruised areas of kiwifruit were extracted using the uninformative variable elimination (UVE), competitive adaptive reweighted sampling (CARS), and principal component analysis (PCA) methods, combined with the successive projection algorithm (SPA), respectively. Besides, the textural feature of the bruised kiwifruit were extracted using the gray-level co-occurrence matrix (GLCM). Finally, Partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), and one-dimensional convolutional neural network (1D-CNN) were built to identify the bruising time of kiwifruit using the spectral feature, textural feature, and fused data (spectra and texture), respectively. The results suggests that the 1D-CNN model built by fused data was the most effective in identifying kiwifruit bruising time, with identification accuracies of 94.55% for 'Hayward', 97.95% for 'Wanhong', and 95.23% for the mixed kiwifruits. The studies suggests that the HSI technique combined with machine learning could effectively identify the bruising time of kiwifruit and provide a reference for kiwifruit quality grading detection.