Integrating Efficient1DNet and hyperspectral imaging to reveal chilling injury and quality deterioration in mangoes.

This research utilized hyperspectral imaging in conjunction with deep learning techniques to facilitate the rapid and non-destructive detection of Chilling injury in mangoes. Chilling injury in mangoes is characterized by the inability to ripen properly and an abnormal increase in firmness, which results from the suppression of the conversion of insoluble pectin into water-soluble pectin under low-temperature stress. In the chilling-injured fruits, abnormal cell wall structures that cannot be properly decomposed were observed. Feature band selection methods focused on wavelengths around 960, 1150, 1720, and 1960 nm highlighted the roles of water, carbohydrates, and cell wall components in the detection of Chilling injury. Efficient1DNet model with Standard Normal Variate Transformation preprocessing achieved the highest accuracy. The averaged test-set metrics (mean ± standard deviation) from repeated random splits were 0.925 ± 0.047, 0.918 ± 0.091, 0.917 ± 0.079, 0.886 ± 0.097, and 0.905 ± 0.083 for accuracy, precision, recall, specificity, and F1-score, respectively. Additionally, it demonstrated a smaller memory footprint (1.6 MB) and faster runtime (10.33 s) than traditional CNN models.