Automatic Stack Velocity Picking Using an Unsupervised Ensemble Learning Method

TL;DR

This work tackles automatic stack (NMO) velocity picking from velocity spectra without relying on labeled data. It introduces Unsupervised Ensemble Learning (UEL), which fuses three information streams—a spectrum gain method based on Local Normalization, near-reference velocity information via ALWLR, and an Attention Scale-Space Filter (ASSF) clustering—into a confidence-area–driven ensemble, with a Dix-based interval-velocity constraint. The approach delivers high accuracy and robustness across synthetic and field data, outperforming clustering-based methods and a CNN baseline while eliminating the need for extensive training. Practically, UEL reduces labeling costs, improves interpretability, and speeds up seismic processing, especially under moderate to low SNR conditions.

Abstract

Seismic velocity picking algorithms that are both accurate and efficient can greatly speed up seismic data processing, with the primary approach being the use of velocity spectra. Despite the development of some supervised deep learning-based approaches to automatically pick the velocity, they often come with costly manual labeling expenses or lack interpretability. In comparison, using physical knowledge to drive unsupervised learning techniques has the potential to solve this problem in an efficient manner. We suggest an Unsupervised Ensemble Learning (UEL) approach to achieving a balance between reliance on labeled data and picking accuracy, with the aim of determining the stack velocity. UEL makes use of the data from nearby velocity spectra and other known sources to help pick efficient and reasonable velocity points, which are acquired through a clustering technique. Testing on both the synthetic and field data sets shows that UEL is more reliable and precise in auto-picking than traditional clustering-based techniques and the widely used Convolutional Neural Network (CNN) method.

Figures (7)

• Figure 1: The workflow of proposed UEL. We first perform the information gain method on the current velocity spectrum and utilize the ASSF method to cluster scatters of the gained spectrum to extract the sample information. Then, we define the near velocity spectra according to the location information and obtain local velocity information by sequentially implementing spectral gain, low-frequency feature filtering, and feature stacking. Next, the prior velocity information is extracted by referring to the manual picking of the nearest seed spectrum. Finally, we fuse three parts of velocity information to estimate the stack velocity of the current spectrum using an ensemble learning method.
• Figure 2: The gain processing of the velocity spectrum: (a) original CMP gather; (b) original semblance velocity spectrum; (c) the gain result of (b).
• Figure 3: (a) The gained spectrum images of near spectra. (b) The stacked low-frequency map and the visual result of ALWLR. In (b), the red stars are the true stack velocity points, the black dotted curves are the confidence boundaries, and the orange curve is the estimated near reference velocity curve.
• Figure 4: The picking results of UEL. (a) and (c) are the gained spectrum with the velocity curves of UEL (red) and label (green) of S5 and A, respectively. (b), (c) and (e), (f) are the NMO correction gather based on UEL picking and label picking of S5 and A, respectively.
• Figure 5: The stack sections of NMO CMP gathers on dataset A are shown in (g)-(j). (g) and (h) are automatic and manual results of line-2560, respectively. (i) and (j) are automatic and manual results of line-2840, respectively.