Nondestructive and interpretable assessment of pear soluble solids content by hyperspectral imaging coupled with...

Soluble solids content (SSC) is a critical indicator for pear maturity evaluation and quality grading. This study aims to develop a rapid, nondestructive, accurate and interpretable SSC prediction framework by coupling hyperspectral imaging (HSI), broad learning system (BLS), and Shapley additive explanations (SHAP). Raw spectral data of 'Qiuyue' pears were collected and preprocessed. A BLS prediction model was constructed based on the preprocessed full-spectrum data, with its performance compared to partial least squares regression (PLSR), backpropagation neural network (BPNN), and support vector regression (SVR). SHAP was adopted for BLS model interpretability analysis and key wavelength selection. Simplified prediction models were further established by applying BLS, PLSR, BPNN, and SVR to the key wavelengths selected by SHAP and three traditional wavelength selection methods. The results showed that BLS outperformed the other three models in both full-wavelength and key-wavelength modeling, with the full-wavelength BLS model achieving the optimal performance (prediction correlation coefficient R_P = 0.8195, root mean square error of prediction RMSEP = 0.8037). In addition, SHAP effectively screened key wavelengths and quantified the independent contribution of each to SSC prediction, greatly improving the transparency and interpretability of the BLS model. This study confirms the feasibility of the HSI-BLS-SHAP strategy for nondestructive and interpretable pear SSC evaluation, providing a technical reference for nondestructive intelligent fruit quality detection.