Hierarchical variable selection for predicting chemical constituents in lamb meats using hyperspectral imaging

Abstract

Hierarchical variable selection method (UVE–SPA–CSA) based on uninformative variable elimination (UVE), successive projections algorithm (SPA) and clonal selection algorithm (CSA) was proposed and applied for the first time to near infrared (NIR) hyperspectral data of lamb meats for predicting chemical constituents of fat, protein and water contents. Instead of selecting different sets of optimum wavelengths for these chemical constituents, only a set of optimum wavelengths were selected with the proposed technique for fat, protein and water. At first, sensitive wavelengths were identified using combinations of UVE and SPA for fat (910, 961, 988, 1011, 1064, 1084, 1192 and 1212nm), protein (921, 944, 947, 971, 1021, 1091, 1269 and 1396nm) and water (910, 961, 994, 1058, 1131, 1195, 1198 and 1312nm), and merged into instrumental optimal wavelengths (IOW). Then, on the basis of the built optimization formulations and CSA, only seven wavelengths (1021, 1084, 1091, 1192, 1212, 1269 and 1396nm) were selected as the prediction optimum wavelengths (POW) for predicting fat, protein and water contents in lamb meats. The multiple linear regression (MLR) models were developed to relate absorbance spectra of lamb samples and their chemical constituents (i.e. fat, protein and water contents) using POW. The fat, protein and water contents were predicted with correlation coefficient of calibration (Rc) of 0.95, 0.80 and 0.91, and residual prediction deviation (RPD) of 4.13, 1.31 and 2.53, respectively. Based on the obtained MLR equations, the distribution maps of these chemical constituents within lamb meats were generated to help knowing and understanding the heterogeneity of lamb meats. The results indicated that the proposed UVE–SPA–CSA is useful for variable selection of hyperspectral imaging and prediction of chemical constituents of lamb meats.