This technique separates the signal and the noise by assuming that both components have different multivariate spectra that PEFs approximate. Therefore, we needed one PEF for the noise and one PEF for the signal. We designed 2D PEFs on common-shot gathers. The noise level is so overwhelming that designing a model for it was relatively easy. We selected nine shot gathers with no water column reflections. This choice was made by inspecting all shot gathers after applying a high-pass filter to them to remove most of the source energy. We then stacked the selected gathers to increase the noise coherency and attenuate any remaining signal.
From this noise model we estimated a stationary 25 5 PEF. The signal PEF was obtained using the Spitz approximation Guitton (2004). The size of the signal PEF is 5 2. Having had estimated the noise and signal PEFs, we proceeded to perform noise removal. Figure 2 displays the estimated signal. A reflection is clearly visible in Figure 2a. Figures 1 and 2 have the same clip for direct comparison. In Figure 2b, the thermohaline structure of the water column is revealed. We followed with a 10-60 Hz bandpass, a gain with the first power of time, and a mute. Since several near offsets needed for prestack wavefield-continuation migration were missing (not acquired or lost to PEFs boundaries), we used an offset continuation tool for regularly sampled data Vlad and Biondi (2001). An image of the denoised data prestack migrated with a constant velocity of 1520 m/s is shown in Figure 3. The reflections are clearly visible throughout the section. The coherency of these events decreases for depths greater than 600m.