next up previous [pdf]

Next: Application to multicomponent data Up: Liu and Fomel: Local Previous: Example of Time-frequency Characterization

Application to Ground-roll attenuation

Seismic data always consist of signal and noise components. The time-frequency denoising algorithm is an effective method for handling noise problems (Elboth et al., 2010). Ground roll is the main type of coherent noise in land seismic surveys and is characterized by low frequencies and high amplitudes. Current processing techniques for attenuating ground roll include frequency filtering, FK filtering (Yilmaz, 2001), radon transform (Liu and Marfurt, 2004), wavelet transform (Deighan and Watts, 1997), and the curvelet transform (Yarham and Herrmann, 2008). Askari and Siahkoohi (2008) applied the S transform to ground-roll attenuation. Here, we propose a similar strategy, except that we are applying the proposed local time-frequency decomposition instead of the S transform.

We applied our methods to a land shot gather contaminated by nearly radial ground roll (Figure 4a). All time-domain images are obtained after automatic gain control (AGC). We applied the forward LTF decomposition to each trace to generate a time-frequency cube (Figure 5a). Note that the ground roll is distributed at localized time-space (left-down section of Figure 5a) and time-frequency (right-down section of Figure 5a) positions. The LTF decomposition is flexible, due to its adjustable time-frequency resolution. Therefore, we designed a simple muting filter to remove the noise components localized in both frequency and space (Figure 5b). The inverse LTF decomposition brings the separated signal back to the original domain (Figure 4). Figure 4c shows the difference between raw data (Figure 4a) and denoised result using LTF decomposition (Figure 4b). It is possible to design more complicated but more powerful masks. Without a time-space mask, our method of simply muting selected frequencies would reduce to band-pass filtering.

The LTFK and LXFK decompositions generate data in different domains (Figure 6a and 7a), which show the trend of ground-roll noise in the frequency-wavenumber sections. Simple frequency-wavenumber masks (Figure 6b and 7b) can eliminate ground-roll noise in the decomposition domains. The recovered signals using the inverse LTFK and LXFK decompositions produce similar results (Figure 8a and 8b, respectively). Furthermore, different decompositions can be cascaded to improve their denoising abilities. For comparison, we used a simple high-pass filter. Figure 9a shows that the high-pass filter fails in removing noise, a larger filter window can damage the seismic signal. Another choice is FK filtering (Figure 9b), which cannot remove the low-frequency part of ground-roll noise. The result is similar to that of the LXFK decomposition (Figure 8b), but the proposed method tends to remove more noise than the standard FK filter (especially near location of time 2.7s and offset 1.2km in Figure 8b and 9b) because of the decomposition's locality and its more flexible design. Radial trace (RT) transform is another approach to deal with ground-roll noise, which is a simple geometric re-mapping method of a seismic trace gather. Idealized ground roll is transformed to small temporal frequency by the RT transform and can be eliminated by applying the RT transform, followed by high-pass filtering and the inverse RT transform (Claerbout, 1983; Henley, 1999). Figure 10a shows that the RT transform performs better than the high-pass filter or the FK filter. However, it still has trouble separating signal and noise near the source. Figure 10b shows the denoised result after cascading the proposed LXFK and LTF decompositions, which achieved the best result in this case (especially at locations around the bottom left corner).

dat siltft iltft
Figure 4.
Field land data (a), denoised result using LTF decomposition (b), and difference between raw data (Figure 4a) and denoised result using LTF decomposition (Figure 4b) (c).
[pdf] [pdf] [pdf] [png] [png] [png] [scons]

ltft mask
Figure 5.
Local $ T$ -$ X$ -$ F$ spectra (a) and filter mask in $ T$ -$ X$ -$ F$ domain (b).
[pdf] [pdf] [png] [png] [scons]

tfk masktfk
Figure 6.
Local $ F$ -$ K$ -$ T$ spectra (a) and filter mask in $ F$ -$ K$ -$ T$ domain (b).
[pdf] [pdf] [png] [png] [scons]

fxk maskfxk
Figure 7.
Local $ F$ -$ K$ -$ X$ spectra (a) and filter mask in $ F$ -$ K$ -$ X$ domain (b).
[pdf] [pdf] [png] [png] [scons]

sitfk sifxk
Figure 8.
Denoised results using different local decompositions. LTFK decomposition (a) and LXFK decomposition (b).
[pdf] [pdf] [png] [png] [scons]

brshot ifk
Figure 9.
Denoised data using different methods (shown for comparison). High-pass filter (a) and FK filter (b).
[pdf] [pdf] [png] [png] [scons]

resu8 csiltft
Figure 10.
Denoised result by using RT transform with high-pass filter (a) and cascading LXFK and LTF decompositions (b).
[pdf] [pdf] [png] [png] [scons]

next up previous [pdf]

Next: Application to multicomponent data Up: Liu and Fomel: Local Previous: Example of Time-frequency Characterization