The Rapid Non-Destructive Detection of the Protein and Fat Contents of Sorghum Based on Hyperspectral Imaging

The protein and fat contents are important indicators for the quality evaluation of brewing sorghum, and a rapid and non-destructive testing method is urgently required to accurately detect them. Hyperspectral imaging (HSI) technology has been widely used in the assessment of the composition of various foods. In this study, different preprocessing methods were used to process the spectral data and determine the optimal preprocessing method. The characteristic spectra were extracted by three combination algorithms, namely, uninformative variable elimination-successive projections algorithm (UVE-SPA), competitive adaptive reweighted sampling-successive projections algorithm (CARS-SPA), and principal component analysis-successive projections algorithm (PCA-SPA). Four models (cascade forest (CF), the backpropagation-genetic algorithm (BP-GA), support vector regression (SVR), and partial least square regression (PLSR)) were established to predict the protein and fat contents based on the full spectrum, the feature spectrum, and fusion data (the integration of the feature spectrum with its corresponding texture features). A comparative analysis revealed that the BP-GA and CF models based on the visible light characteristic spectra extracted by PCA-SPA and UVE-SPA were the best models for predicting the protein and fat contents, respectively; they had respective RPD values of 5.1716 and 12.9724 and respective AB_RMSE values of 0.0916 and 0.0243 g/100 g. The overall results show that HSI combined with machine learning algorithms can rapidly and non-destructively predict the protein and fat contents of sorghum.