Abstract: This study systematically quantified and analyzed the correlation between different regions of interest (ROI) characteristics (part, shape, and size) and the prediction accuracy of soluble solids content (SSC) in kiwifruits to build a kiwifruit SSC prediction model by integrating hyperspectral imaging with ROI selection. The ROI spectral data of whole fruits were preprocessed using multiplicative scatter correction (MSC), Savitzky-Golay (SG) smoothing, standard normal variate (SNV) transformation, or SNV-SG smoothing. A partial least squares regression model was established to predict the SSC of kiwifruits, and performance analysis was conducted to determine the optimal preprocessing strategy. Furthermore, we extracted the ROI spectral information of different shape and size combinations at the equator, calyx, and peduncle of kiwifruits to compare the accuracy of the prediction model. The results revealed that SNV preprocessing yielded the best performance, with a coefficient of determination (RP2) of 0.832 7 and a root mean square error of prediction (RMSEP) of 0.387 1 for whole-fruit ROI prediction set. The ROI characteristics significantly impacted the accuracy of SSC prediction, and the effects of fruit part, shape, and size followed the decreasing order: equator > calyx > pedicel; circular > square; and small > large. Notably, the small circular ROI at the equator yielded the optimal prediction, with RP2 = 0.917 3 and RMSEP = 0.221 7. This study demonstrates the crucial role of ROI optimization in hyperspectral image modeling, clarifies the advantages of the “equator-circular-small” combination, and provides an effective approach for improving the prediction accuracy of SSC in kiwifruits using hyperspectral technology.

Key words: hyperspectral imaging; kiwifruit; soluble solids content; region of interest; feature optimization