In order to analyze the robustness of the near infrared spectroscopy model, this paper proposes a method of automatically generating the fuzzy membership by introducing the fuzzy membership when building the model. This method constructs a description function in the data domain of spectrum samples, introduces two factors?confi?dent factor and trashy factor, and then obtains the fuzzy membership function of samples from a mapping function. It automatically generates the fuzzy membership of each sample after optimizing parameters. On that basis, the re?gression model of apple sugar content was built based on fuzzy support vector machines ( FSVM) . The experimental results revealed that comparing with regular multivariate linear regression ( MLR) , partial least squares regression (PLSR) and support vector machines (SVM), the FSVM model showed the best performance with the change of training samples, under the influence of five noises, i.e. Gaussian noise, multiplicative noise, baseline shift, base?line slope and wavelength shift. The model shows better performance in robustness, especially generalization ability and anti?noise ability, primarily due to the contribution of fuzzy membership.