Ground roll attenuation using SVD and time-frequency-wavenumber filters

Abstract

Ground roll (GR) is a coherent noise phenomenon with high-amplitude and low-frequency patterns that conceal reflections. To create a correct velocity model, especially from medium to great depth, attenuation of the GR is unavoidable. To carry out attenuation and preserve a virgin data set, out of noise-cone singular value decomposition (SVD) and time-frequency-wavenumber (TFK) filters are used. Here four steps are introduced: (1) SVD, (2) FK filtering, (3) the Gabor transform (GT) and (4) a time-variant band-pass filter. SVD decomposes data into singular vectors and gives us the ability to have sub-images that relate to different singular values. The data set is reconstructed using first singular-vectors that relate to first singular-values, and include more signal and less noise. To preserve the reflection signal, the GR event is flattened using a forward normal move out correction (NMO) followed by the application of SVD to preserve 50% of the introduced events (to remove a part of the GR) and then finally reverse NMO is conducted. Although the FK filter is able to remove all of the GR energy based on the low velocity, the choice of low-dip for filtering causes high-amplitude smearing and distorts waveforms. Consequently, FK is applied to remove events with dips of more than 40 . The results of SVD plus high-dip FK have some issues that are created by merging high-amplitude and low-frequency events at the apex of the noise cone with reflections. Since at shallow depth and near offsets primary reflections have a higher frequency than the GR, the GT was used to deduce the dominant frequency of the GR for different times. Based on the estimated noise frequency, a cone area was chosen to filter the remaining noise and detect reflection events. These procedures were applied on real shot gathers to show the results of GR attenuation.