Summary. Introduction

Transcription

1 . Tony Martin*, Cristiano Saturni and Peter Ashby, ION Geophysical Summary Modern marine seismic surveys may contain many Terabytes of data. In seismic processing terms, understanding the impact and effectiveness of a single action on these data is cumbersome, time-consuming and generally subjective. In this paper we present a case study using multidimensional attributes to define anomalies in the product of a seismic processing algorithm that attenuates noise from a dataset. In addition to direct statistical measures of the data, we define metrics describing consistency and effectiveness of the denoising algorithm. Collectively these measures form a multi-dimensional space that is interrogated by an unsupervised learning algorithm. The product is a data partitioning that provides the user with a sequence of outliers for further investigation. Introduction Seismic processing projects require a multitude of complex algorithms that use significant and costly compute resources. Before a process is run on an entire dataset it is tested on a sparse subset of that data to confirm consistency across the dataset. Given global subsurface complexities this assumption is questionable. To determine if the process has been successful, some form of quality control (QC) is performed after the action has completed. Unfortunately this QC is also coarsely sampled and subjectively analyzed. In this paper we present an approach to QC that interrogates all of the data. The method defines a number of metrics that describe the effectiveness of the process. The example shown focuses on the process of removing swell noise from a marine seismic survey, but could be modified to apply to any procedure in a processing flow. Method Despite advances in acquisition design, most marine seismic data are contaminated by noise directly associated with sea surface waves. The amplitude and wavelength of the surface waves can result in pressure changes at the receivers (Schonewille et al 2008). This in turn manifests itself as a high amplitude, low frequency response and is anomalous to the desired recording. For numerous reasons this noise should be removed from the record. We call this step swell denoising and implement an iterative F-X prediction filter approach to remove the noise. No a priori information is used to assist the attenuation process, however analysis of the data can help constrain the process to target the noise. Often one set of parameters are used for an entire survey whose acquisition may take several months and encounter numerous environmental conditions. As such it becomes ever more important that QC of the process is performed on all the acquired data. In our experiment we use an unsupervised machine learning algorithm to mine the data. An unsupervised process is one used to find patterns or structures in unlabeled data and is unbiased by constraints or weighting to a predefined output. An example of this is a clustering algorithm. This differs from supervised learning where the process is provided with examples of desired outputs by a user. The objective of supervised learning is to find a rule that produces the desired output from any input data. In machine learning the term feature is used to describe a dimension, metric or attribute. By determining statistical properties and RMS measures of seismic amplitudes in the input, product and difference (input-output) of a denoising process, Spanos et al (2013) presented a 3-D space suitable for partitioning by a clustering algorithm. The partitioned product, based on these measures separates the data into subsets of good or bad filtering. In our experiment we add additional features. The focus of these is twofold: 1) To provide more targeted features for an unsupervised learning algorithm. 2) To provide features that could be used in conjunction with a priori information for future supervised learning systems. In an attempt to improve the separation between desirable and undesirable outcomes we have added measures that define consistency and effectiveness of a process. To achieve this we use a Canny filter. Canny filters are edgedetecting filters that work directly on the seismic sample. Canny (1986) defines the filter as a multi-stage process where: 1) A Gaussian filter is applied to denoise the image. 2) A gradient operator is applied to obtain intensity and direction. 3) Non-maximum suppression determines if the pixel is a good candidate for edge detection. 4) Hysteresis thresholding is applied to find where edges begin and end. SEG New Orleans Annual Meeting Page 4790

2 The Canny filter product is then fed into a transform where the detected edges are partioned into azimuth sections relative to vertical. An additional weighting to the edge intensity and spatial extent is also determined. In the difference data we attempt to detect edges that are consistent with coherent signal in an orientation where we might expect this energy to reside. We use vertical edge detection to search the output data for remnant noise, which will tend to appear as vertical features in these images. These two features tell us two things: whether the process has removed energy we do not want it to and if there is remnant noise in the output (Figure 1). A B pairs depends on the characteristics of the swell in the data. The noise estimate is determined on the input and difference data. The new features are then used alongside the statistical measures, such that there are up to seven features that can be used by the partitioning process. We implement a normalization scheme such that all features are equally weighted in the partitioning process. In our initial experiment we use a top down hierarchical clustering algorithm, although the process is not limited to this form of clustering. Top down hierarchical algorithms produce several partitions organized in a hierarchy. The set of points is initially partitioned into classes which are recursively partitioned in other classes using the same algorithm with different parameters. Typically a graph is produced from the original dataset using a distance function and a threshold; the connected components of the graph are computed, thus obtaining a first partition; then the points of each connected component are recursively partitioned using the same distance function and a smaller threshold. C D For this initial trial we attempt to understand the value that can be added by supplementing the statistical measures introduced by Spanos et al (2013) with the data-derived features described in this section. Example Figure 1: A and C) Output shot gather after denoising. B) Same shot with edge-detecting Canny Filter D) After transform sectioning into azimuth components. Highlighted are energy with vertical edges (blue) indicating remnant noise in the outpu datat. Measures of data consistency can help constrain the data mining. To determine a measure of consistency we perform a 2-D FFT on a restricted time-offset window in shot gather data. From the transformed data we generate the ratio of RMS amplitudes in low frequency-low wavenumbers against low frequency-high wavenumbers. This is considered to be a quantitative estimate of the noise present in the data. The exact choice of frequency-wavenumber In the following example we have processed data with a swell denoising algorithm. The data is contaminated with a moderate amount of swell noise. There are approximately 2500 shot gathers used in this exercise. In the data distribution the first half of gathers are less affected by swell noise than the second half. At discrete locations the parameterization of the module to remove swell noise was artificially altered to under and over filter the data. In no instance was any of the denoise processing outside the bounds of realistic acceptability; hence a possible margin of error by the user is encompassed. Both denoise and QC processes were run on shot gathered data. Using the input, output and difference data we create a multi-feature space and mine the data using a hierarchical clustering algorithm. A number of partitioning tests have been run on different combinations of features, but in this paper we concentrate on the benefit of adding all our new features. These are indicators of noise in the output, coherent primary energy in the difference and noise analysis measures in the input and difference. The comparison is therefore between a three-feature space where only the original statistical measures from Spanos et al (2013) are used, and a seven-feature space containing the new metrics. SEG New Orleans Annual Meeting Page 4791

3 The process generates a number of displays for the user. The first is a decay curve of the maximum distance of any point between clusters against the number of iterations of partitioning. This is shown in Figure 2. The second plot is shown in Figure 3 and displays the number of points in each new cluster per iteration of partitioning. A cross-plot of a pair of features after partitioning is also available. Finally a list of shot gathers in each cluster is also generated. Based on these the user can evaluate the clusters and interrogate the outliers. Minimum distance between clusters Statistical features Statistical + new features Number of iterations Figure 2: QC deliverable showing a plot of maximum distance (between any point in a clusters) against the number of iterations. Note that the green curve shows a greater distance between points for any iteration than the red. The simplest way to determine the benefits of the new features is to understand differences or similarities in the partitioned data when these features are included. From the images in Figure 2 and 3 we see for the first 60 iterations: 1) Using the statistical measures we see that no cluster is partitioned with more than five shot gathers in it. 2) The statistical measures form clusters with much smaller distances for these initial 60 iterations. 3) Using the new features alongside the statistical ones, clustering of significant numbers of shot gathers occurs at iteration 10. 4) The separation at iteration 10 forms a new cluster containing 1174 points. This partitioning reflects the different noise content present in the second half of the data distribution. 5) The partitioning formed by the new features shows a greater distance between clusters, such that these clusters are better defined and easier to separate by the algorithm. Number of points in smallest clusters Statistical features Statistical + new features Number of iterations Figure 3: QC deliverable showing plot of number of points in the smallest partition against the number of iterations. Note the green curve shows a large data population split at iteration 10, whereas the red curve never achieves more than 5 points per iteration. It can be seen in Figure 2 that the separation that occurs in the initial five iterations forms clusters with the largest maximum distance between clusters. Therefore the initial iterations should identify the data with the most significant anomalous behaviour. So, it is instructive to understand which data are being partitioned in the first five iterations. To do this we look at the seismic data clustered by the three and seven feature analysis. Figures 4 and 5 show the seismic data partitioned from the full dataset by the first five iterations of this experiment. In both figures the data delineated by a red bars are from the clustering with three statistical features, whilst the green bars represent those separated when the four new features are also used. Of the twelve gathers separated by either approach, seven are common to both products. These are displayed with both green and red bars. There are two shot gathers with only green bars in figures 4 and 5. These are only isolated when the new features are included. Analysis of these two gathers shows a consistency with the remaining seven green barred shots. The noise on input and output is equivalent and there is evidence of unwanted primary energy in the difference of all nine shot gathers. The inclusion of the feature that uses a Canny filter to search for coherent energy in the difference data has influenced the SEG New Orleans Annual Meeting Page 4792

4 clustering. All nine gathers indicated by green bars are from the over filtered data. The gathers shown with only red bars in figure 4 and 5 are isolated by the statistical measures only. These are inconsistent with the rest of the distribution of 12 shot gathers. There is no indication of increased noise content in the input and they do not have any remarkable features in the difference data. These three gathers are actually from an under filtered portion of the data. the statistical distribution of the data and filtering process, as well as indicators of process effectiveness and consistency, we can use unsupervised machine learning algorithms to partition the data to indicate outliers and potential problems. Figure 5: Difference (input-output) shot gather data from both the three and seven feature analysis. Those with red bars underneath them are separated by the three feature analayis, whilst those with green are separated by the seven feature analysis. Figure 4: Input shot gather data from both the three and seven feature analysis. Those with red bars underneath them are separated by the three feature analayis, whilst those with green are separated by the seven feature analysis. In this instance the separated data using three statistical measures isolates a collection of both good and bad filtered data. This is undesirable. The nine shot gathers separated by using all seven features appear to contain only badly filtered data and all would require attention by the geophysicist. Conclusions Quality control of seismic processing algorithms on large volumes of data is time-consuming, unrepresentative and essentially qualitative. In this paper we have presented a data mining approach to quality control in an attempt to overcome these issues. By defining metrics or features that provide information about We have shown that the inclusion of the features that use Canny filters and noise estimates create more consistent clusters of data. Distribution plots are used to identify the data that requires investigation by the processing geophysicist. By utilizing intelligent metrics it has been shown that this type of approach to quality control will enable a timely interrogation of an entire dataset. The new features presented here have been specifically designed to facilitate the development of a supervised learning algorithm. For this they must be used with additional external information such as weather conditions, currents and acquisition configurations. Acknowledgements The authors wish to thank John Brittan, Jacques Leveille and Nick Bernitsas for helpful comments and ION Geophysical for permission to publish this work. SEG New Orleans Annual Meeting Page 4793

5 EDITED REFERENCES Note: This reference list is a copyedited version of the reference list submitted by the author. Reference lists for the 2015 SEG Technical Program Expanded Abstracts have been copyedited so that references provided with the online metadata for each paper will achieve a high degree of linking to cited sources that appear on the Web. REFERENCES Canny, J., 1986, A computational approach to edge detection: IEEE Transactions on Pattern Analysis and Machine Intelligence, 8, no. 6, , PMID: Schonewille, M., A. Vigner, and A. Ryder, 2008, Advances in swell-noise attenuation: First Break, 26, no. 12, Spanos, A., and M. Bekara, 2012, Using statistical techniques to improve the QC process of swell noise filtering: 75th Conference & Exhibition, EAGE, Extended Abstracts, doi: / SEG New Orleans Annual Meeting Page 4794