This study aimed to achieve the rapid quantification and visualization of the starch content in sweet potato via near-infrared (NIR) spectral and image data fusion. The hyperspectral images of the sweet potato samples containing 900-1700 nm spectral information within every pixel were collected. The spectra were preprocessed, analyzed and the 18 informative wavelengths were finally extracted to relate to the measured starch content using the multiple linear regression (MLR) algorithm, producing a good quantitative prediction accuracy with a correlation coefficient of prediction (rP) of 0.970 and a root-mean-square error of prediction (RMSEP) of 0.874 g/100 g by an external validation using a set of dependent samples. The MLR model was further verified in terms of soundness and predictive validity via F-test and t-test, and then transferred to each pixel of the original two dimensional images with the help of a developed algorithm, generating color distribution maps to achieve the vivid visualization of the starch distribution. The study demonstrated that the fusion of the NIR spectral and image data provided a good strategy for the rapidly and nondestructively monitoring the starch content of sweet potato. This technique can be applied to industrial use in the future.
Keywords: Modeling; Prediction; Starch; Sweet potato; Visualization.
Copyright © 2023 Elsevier B.V. All rights reserved.