To simplify the temperature extraction steps for Brillouin scattering and also improve the extraction precision, we propose a new method for directly obtaining the temperature characteristics of Brillouin gain spectra based on the radial basis function neural network. The Brillouin scattering spectra at various temperatures are used as the training set to establish the temperature model. The temperature can be obtained through directly inputting the Brillouin spectra into the model. The effects of three methods of smooth fitting, back propagation neural network and radial basis function neural network on the temperature measurements are compared. In the experiment, 77 groups of data at sweeping frequency intervals of 0.175, 1, 5, 10, and 20 MHz are selected and also those at different linewidths are expanded. The results show that, the root-mean-square error (RMSE) based on the radial basis function neural network is relatively small. Moreover, the RMSE increases slowly with the increase of step frequency. When the step frequency is 20 MHz, the error of single line width is up to 0.8002 ℃ and that of multiple line width is 1.0814 ℃, 33.04% and 42.88% of that by the smooth fitting method, and 40.25% and 55.89% of that by the back propagation neural network, respectively. The convergence is improved to a certain extent as a result of calculation step reduction in the method based on the radial basis function neural network.