Determination of firmness, soluble tannin content, and storage stage of post-harvest astringent persimmons using VNIR-SWIR hyperspectral imaging and deep learning

Firmness (FI) and soluble tannin (ST) content are key indicators reflecting the degree of astringency removal in persimmons. In this study, a novel nondestructive method was developed using hyperspectral imaging (HSI) technology to measure the FI, ST content, and storage period of stored Mopan persimmons. For quantitative analysis of FI and ST, three spectral datasets were employed: VNIR, SWIR, and the fused VNIR+SWIR spectra. Three deep learning architectures were implemented, including a CNN, BiLSTM, and a self-attention-enhanced CNN-BiLSTM-Self-Attention hybrid model. Results demonstrated that the CNN-BiLSTM-Self-Attention model outperformed PLSR, LS-SVM, and other deep learning models, achieving the best performance for FI using VNIR spectra (residual predictive deviation, RPD=3.93) and ST using SWIR spectra (RPD=3.75) in the prediction set. For qualitative analysis of the storage period, Spearman’s rank correlation analysis indicated a significant correlation (P ≤ 0.01) among storage time, FI, and ST content. Consequently, FI, ST, and the VNIR+SWIR spectral data were integrated to construct a gradient boosting decision tree (GBDT) model, which achieved 100% classification accuracy in both the calibration and prediction sets. Overall, this study demonstrates HSI-deep learning integration's feasibility for predicting persimmon deastringency degree, providing a promising approach to determine optimal edible quality of astringent persimmons. • HSI combined with DL was used to determine the quality of astringent persimmon. • Correlation analysis of quality attributes of astringent persimmons was performed. • The GBDT model was used to achieve accurate discrimination of storage stages. • The integrated architecture of CNN_BiLSTM_Self_Attention achieved the best results.