Abstract:The sugar content of “Huayou” kiwifruits was estimated by hyperspectral imaging. The mean reflectance spectra of square regions measuring 10 × 10, 20 × 20, and 30 × 30 (pixels × pixels) and mask images were extracted. Four sugar content prediction models, i.e., partial least squares (PLS), least-square support vector machine (LSSVM), extreme learning machine (ELM), and error back-propagation network (BP), were established using the processed full spectra, and the effect of spectral extraction regions on the prediction accuracy of sugar content was analyzed. The mean spectra were preprocessed by Savitzky-Golay smoothing combined with standard normal variate transformation. The results showed that the prediction performance of LSSVM, ELM, and BP improved with increasing size of the region of interest. The LSSVM based on mean spectra extracted by mask images exhibited the highest prediction performance; the correlation coefficient of prediction was 0.97, the root mean square error of prediction was 0.86 °Brix, and the relative prediction error was 4.06. The results of this study indicate that selecting appropriate spectral extraction regions in hyperspectral imaging is helpful to improve the prediction accuracy of the established models.
