During the grading of Korla fragrant pears, the packaging must be removed for internal quality measurement. This inefficient process severely reduces detection efficiency. Achieving efficient and non-destructive detection of their quality while packaged is key to solving this problem. This study combined a Fully Connected Neural Network (FCNN) with Gaussian Filtering (GF), Savitzky-Golay (SG), and Simple Moving Average Filter (SMAF) separately to correct the spectral data of packaged pears (PPs). Concurrently, using the corrected spectral data, three prediction models for soluble solids content (SSC) and firmness (FI) of multi-origin Korla fragrant pears were constructed: Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and Lasso. The results indicate that the FCNN-GF correction method performs best best; PLSR exhibited the best prediction performance. The prediction results of PPs-FCNN-GF-PLSR for SSC and FI were as follows: R 2 ₚre (0. 84, 0. 78), RMSE ₚre (0. 449 %, 0. 379 N), RPD (2. 48, 2. 16). These results closely approached the prediction levels of UPs. In external independent validation, the prediction results for SSC and FI were R 2 (0. 83, 0. 75), RMSE (0. 276 %, 0. 252 N), RPD (2. 30, 1. 89), respectively. This study achieves efficient detection of pear quality indicators in their packaged state, holding significant practical application value for simplifying the quality grading process of Korla fragrant pears and enhancing economic benefits. • A near-infrared spectroscopy correction method were developed. • An efficient detection of the quality of packaged Korla fragrant pears were realized. • The impact caused by copying paper on the NIR spectra of pears was analyzed. • Developed a multi-origin prediction model for indicators of Korla fragrant pears.
Xu et al. (2026) studied this question.