Distributed optical fiber sensing system based on Brilouin scattering has attracted wide attention for its ability of sensing the measured field by detecting the continuously distributed information in time and space. Considering the trait of the spectral shape variance during the Brillouin scattering process in optical fiber and the requirement of high accuracy, a novel method based on radial basis function neural (RBFN) networks in which the output layer weights are adjusted by Levenberg-Marquardt method is presented. A model of actual Brillouin spectrum is constructed by Gaussian white noise on the theoretical spectrum, the core frequency is 11.213 GHz and the weight is 4∶1. Comparing the proposed algorithm with traditional back propagation (BP) neural networks, polynomial five times curve fitting and piecewise cubic spline interpolation, the relative error of the new method is 0.0015179% and the temperature error is 0.152 ℃. The appraised parameters are better than other three algorithms at the same test system under different pulse widths and temperatures. The numerical and experimental results show that the RBFN networks is suitable for the fitting of Brillouin scattering spectrum, and the forecast accuracy is improved efficiently.