Hyperspectral imaging has large data volume and high dimensionality, and original spectra data includes a lot of noises and severe scattering. And, quality of acquired hyperspectral data can be influenced by non-monochromatic light, external stray light and temperature, which resulted in having some non-linear relationship between the acquired hyperspectral data and the predicted quality index. Therefore, the present study proposed that competitive adaptive reweighted sampling (CARS) algorithm is used to select the key variables from visible and near infrared hyperspectral data. The performance of CARS was compared with full spectra, successive projections algorithm (SPA), Monte Carlo-uninformative variable elimination (MC-UVE), genetic algorithm (GA) and GA-SPA (genetic algorithm-successive projections algorithm). Two hundred Korla fragrant pears were used as research object. SPXY algorithm was used to divided sample set to correction set with 150 samples and prediction set with 50 samples, respectively. Based on variables selected by different methods, linear PLS and nonlinear LS-SVM models were developed, respectively, and the performance of models was assessed using parameters r2, RMSEP and RPD. A comprehensive comparison found that GA, GA-SPA and CARS can effectively select the variables with strong and useful information. These methods can be used for selection of Vis-NIR hyperspectral data variables, particularly for CARS. LS-SVM model can obtain the best results for SSC prediction of Korla fragrant pear based on variables obtained from CARS method. r2, RMSEP and RPD were 0.851 2, 0.291 3 and 2.592 4, respectively. The study showed that CARS is an effectively hyperspectral variable selection method, and nonlinear LS-SVM model is more suitable than linear PLS model for quantitatively determining the quality of fragrant pear based on hyperspectral information.