Random noise attenuation in seismic data by non-local mean filtering

The reliability of seismic mapping is strongly dependent upon the quality of the records. Seismic records are usually affected by various types of noise such as ground rolls, multiples, random noise, reflection and reflected refraction from near surface structures, and so on. Random noise resulted from random oscillation during acquiring data is one of the most important and harmful noises that exist in seismic data over entire time and frequency. Random noise attenuation is an important step in seismic data processing affecting the data interpretation. Many efforts have been made to remove this type of noise from the seismic data. The predictive filter is an ordinary method commonly used for random noise attenuation from seismic data. This filter can be used in various domains, such as the f-x domain (Haris and White, 1997) and the discrete cosine domain (Lu and Liu, 2007). Jones and Levy (1987) removed events which were not coherent trace-to-trace events by means of the Karhunen-Loeve transform. Karsli et al. (1996) applied complex trace analysis to seismic data to suppress random noise and improve the temporal resolution of the data.Terickett (2008) attenuated the random noise from the seismic data by Cadzow filtering of a constant frequency slice in an f-xy domain. Bekara and Baan (2008) attacked the random noise problem using the empirical mode decomposition technique. In this study, we used non-local means filtering (Buades et al., 2005) developed for image denoising for random noise suppression in seismic data. In this method, a seismic record can be considered as an image. The non-local means method is based on the assumption that the image content is likely to repeat itself within some neighborhood. The neighborhood of a pixel is generally chosen to be a square with a dimension size of three to nine centered upon the pixel of interest. However, the size and shape of the neighborhood can vary. In this method, for each pixel i with a neighborhood N i, the pixels j with the neighborhoods N j similar to that of the interest pixel are found. The denoised value of a pixel i is determined by a weighted average of all the pixels in the image. The weight of each pixel j can be calculated by the similarity between the two pixels i and j computed using the Gaussian weighted Euclidean distance between the neighborhood around the pixel i and the neighborhood around the pixel j. To investigate the efficiency of the proposed method, we tested the non-local means algorithm on both synthetic and real seismic data. We also compared the obtained results by that of the traditional mean and median algorithm for seismic data denoising. To investigate further, we applied the three denoising methods to synthetic seismic data with different amounts of signal-to-noise ratios. After analyzing all of the results, the non-local means algorithm proved to be a better algorithm for seismic data denoising and had the best performance among the three methods.