Home

Quick Links

Search

 
Automatic QC of denoise processing using a machine learning classificationNormal access

Authors: M. Bekara and A. Day
Journal name: First Break
Issue: Vol 37, No 9, September 2019 pp. 51 - 58
DOI: EAGE-EXPORT-FAKE-DOI
Special topic: Machine Learning
Language: English
Info: Article, PDF ( 942.83Kb )
Price: € 30

Summary:
Noise attenuation is an important part of a typical seismic data processing sequence. The general purpose of noise attenuation is to improve the resolution of seismic images. It can also be used to pre-condition the data prior to the application of certain processes to avoid the generation of artefacts. For example, applying a de-bubbling filter to marine seismic data that contains moderate swell noise will create visible low-frequency artefacts in the output. Therefore, adequately removing the swell noise with techniques similar to the one proposed by Bekara and van der Baan (2010) is a prerequisite before applying such a process. The first step in any denoising process is to optimize the key parameters to ensure that noise has been adequately attenuated without causing attenuation of signal. This is usually done by testing the filter using a range of parameters on a small subset of the data known as the test lines. At this stage, the optimality of the denoise process is visually assessed by the processor and there are generally no issues, as the volume of the test lines is small. Once the parameterization is approved, it will then be used to process data from the entire survey. However, there is no guarantee that the chosen parameterization will be optimal for the whole survey because the characteristics of the noise and the signal can change dramatically throughout the data. For example, in a typical narrow azimuth towed streamer marine survey (~ 9000 km2), which can last for up to four months, the swell noise characteristics (peak frequency, amplitude and spatial contamination) can vary signifi-cantly, as they depend on the weather conditions. This will make any fixed parameterization for the swell noise attenuation filter potentially sub-optimal for some parts of the survey. Therefore, Quality Control (QC) is mandatory to give people confidence in the denoise processing and to allow the processor to move to the next step in the processing sequence. In brief, the QC process is an attempt to provide a reliable answer to the following two questions: 1. Is there any residual noise or signal attenuation in the data after the application of the filter? 2. If yes, where is it in the survey? A reliable QC can take a considerable amount of time and resources in a typical seismic processing project. It is therefore advantageous to automate this process to improve project turnaround. This paper expands on previous work by Spanos and Bekara (2013), where an unsupervised outlier detection approach was used to answer the above two questions. To improve the reliability of the automatic QC, we propose using a supervised learning approach to build an automatic classifier to predict the type of filtering as one of three classes (mild, optimal or harsh). The framework was tested on full-scale production to QC the results of noise attenuation prior to wavefield separation for towed streamer dual-sensor marine data.


Back to the article list