Fat content is an important indicator of the quality of walnuts. In order to achieve the rapid nondestructive detection of walnut fat content, the near infrared spectrum of walnut kernel was collected in the spectral range of 1040~2560nm. Multivariate scatter correction and standard normalized variate were used to preprocessing the original spectral information. And abnormal samples were eliminated by the Mahalanobis distance method. Then the feature bands were screened by the method, which combined competitive adaptive reweighting sampling (CARS) and correlation coefficient method (CCM) algorithm. Finally, the partial least squares regression and the support vector machine regression algorithm were used to establish prediction model for the fat content of walnut kernels. The results showed that with the six feature bands selected as input, the validation set coefficient of the walnut kernel fat content prediction model established by partial least squares regression algorithm was 0.86, and the root mean square error was 1.5849%. The validation set coefficient of model established by the support vector machine regression algorithm was 0.88 and the root mean square error was 1.3716%. It was showed that the modeling quality of the support vector machine regression algorithm was better than the partial least squares regression algorithm. The support vector machine regression prediction model established by the feature bands could sharply reduce the modeling complexity and realize the rapid nondestructive detection of the fat content of walnut kernel.