Abstract: The nonlocal means method has good denoising performance, but its application is newly developing in seismic data processing. The method, using the structural redundancy of data, taking the small window with local structure and neighborhood as unit, conducts weighted arithmetic by using local structural similarity to enhance effective signals and to depress random noises. Aiming at huge amount of pre-stack seismic data, strong background noise and simple local structure, the original nonlocal means method filters each point, conducts weighted calculation after calculating the weight coefficient of all points within data. Because of short points such as huge computation volume and poor adaptability to strong noise background, the original nonlocal means method has been improved. Three modifications have been proposed for the nonlocal means algorithm. Firstly, the scan windows are divided with velocity spectrum; then, pre-selection of similar set is based on singular value decomposition in gradient domain; lastly, selection of self-adaptive filtering parameter is based on the scale of similar set. De-noising results for the test data demonstrate that the method can effectively depress the random noise of seismic data.