Differential optical absorption spectroscopy (DOAS) has become a widely used method to measure trace gases in the atmosphere. It can identify trace gases through narrow-band molecular absorption. Concentration of trace gases is retrieved using method of least-squares fits of reference spectra to the measurement spectra. A novel retrieval method based on radial basis function (RBF) neural network was developed to retrieve the concentration of trace gases in DOAS system. The coefficient of the hidden layer was trained by modified nearest neighbor clustering algorithm, and that of the output layer was trained by gradient descent algorithm. These result in a fast speed of convergence of network. At last, there was comparison between the novel retrieval method and the conventional least-squares fitting. The experimental results show that the reliability and accuracy of DOAS are improved and detection limits are decreased by using the novel retrieval method.