Objective: The sugar content in apple directly affects its taste. The near infrared spectrum data of apples were collected in the wavelength range of 900~1 700 nm to detect the sugar content nondestructively in conjunction with the chemometrics method. Methods: Firstly, the spectral data were corrected by baseline, scattered, smoothed, and scaled in turn, and then the best preprocessing method was selected by minimizing the root mean square error of cross-validation. Secondly, 7 and 52 feature variables were selected by successive projection algorithm (SPA) and competitive adaptive reweighted sampling (CARS) respectively. Finally, the linear PLS model and nonlinear ELM model were established with the feature variables selected by SPA method, CARS method and their combination as input variables respectively. Results: The results showed that the modeling effect of the combined feature variables was better than that of the single method, and the nonlinear models were better than that of the linear models. Conclusion: ELM model established by using combined characteristic variables has the best prediction effect, with RMSEC=0.710 1, R2c=0.883 8, RMSEP=0.637 5, R2p=0.894 5, which can provide theoretical reference for the development of apple hyperspectral detection device.