For a large number of experimental data, the BRDF surface fitting method based on B-spline function and least squares theory, the ill-conditioned normal equations, the low accuracy of the results and long CPU time may be appeared. Thereby, in this paper by using the BP learning method, combined with the training process of L-M algorithm, an improved method is presented. And the method is applied to the BDRF data processing, the result shows that this method is effective, and has greatly improved in accuracy and reduced running time.