The Importance of Parameterization in Machine Learning

It is well known that machine learning is based on complex statistics and programming, which for most of us, are not straightforward. When it comes to choosing the best technique, either supervised or unsupervised, the decision might not be easy and sometimes may even be overwhelming. Where do you even begin?

Luckily, much research has been done recently to help geologists find an ideal starting algorithm, or the best machine-learning package. Here we discuss how just focusing on your choice of parameterization in a given machine-learning algorithm is one easy way to obtain better results without having to switch from technique to technique.

And by “parameterization,” we simply mean the user-chosen parameters that are entirely controlled by the geoscientist.

Questions on parameterization might seem simple but can have a high impact on the classification results: Will cropping the seismic volume improve the results? Will increasing the analysis window show better classifications? What if we use more or less training data? How about just picking the training polygons in the less geologically complex areas?

All these questions are valid and do impact the results.

Please log in to read the full article

It is well known that machine learning is based on complex statistics and programming, which for most of us, are not straightforward. When it comes to choosing the best technique, either supervised or unsupervised, the decision might not be easy and sometimes may even be overwhelming. Where do you even begin?

Luckily, much research has been done recently to help geologists find an ideal starting algorithm, or the best machine-learning package. Here we discuss how just focusing on your choice of parameterization in a given machine-learning algorithm is one easy way to obtain better results without having to switch from technique to technique.

And by “parameterization,” we simply mean the user-chosen parameters that are entirely controlled by the geoscientist.

Questions on parameterization might seem simple but can have a high impact on the classification results: Will cropping the seismic volume improve the results? Will increasing the analysis window show better classifications? What if we use more or less training data? How about just picking the training polygons in the less geologically complex areas?

All these questions are valid and do impact the results.

Probabilistic Neural Networks

In this analysis, we chose probabilistic neural networks for the supervised classification of seismic facies related to the main, larger-scale architectural elements found in the Gulf of Mexico: salt, mass-transport deposits and hemipelagic sediments (figure 1). In the PNN technique, the parameters the geoscientist can control are the type and number of seismic attributes chosen, the facies polygons picked, the amount of training data, the size of the seismic volume and the analysis window size used for the calculation of seismic attributes or any filtering of them.

As input seismic attributes we chose, based on our interpreter knowledge: coherence, k1 most-positive curvature, k2 most-negative curvature, envelope and grey level co-occurrence matrix contrast and dissimilarity. However, the PNN algorithm used was coupled with an exhaustive searching algorithm that defines the best combination of seismic attributes. Therefore, not all the attributes were used to perform each classification.

In the pre-processing stage, we applied an edge-preserving smoothing filter, known as Kuwahara filter, to all the seismic attributes. This filter has been well documented to improve seismic facies classification by smoothing the interior structures while sharpening the edges, therefore facilitating the delimitation of facies.

To demonstrate how to obtain better results just by varying initial parameters, we ran a series of different tests in which we changed a single parameter at a time:

  • Parameter 1: Increasing the amount of training data
  • Parameter 2: Increasing the size of the analysis window
  • Parameter 3: Cropping the seismic volume into two smaller volumes
  • Parameter 4: Picking the facies polygons in areas more geologically complex

Impact of Parameterization

The facies classification results are shown in figure 2. The initial results (figure 2a) were obtained using one line and one slice for training, an analysis window for the Kuwahara filter equal to the default bin size, and the facies polygons picked in the salt body No. 1, which was less faulted and easier to map the continuity of the conformable reflectors.

The initial results show the hemipelagic sediments of the deeper areas being incorrectly classified as salt facies, as well as the faults located over the salt body No. 1. The time slice below shows an MTD being completely misclassified as salt.

When increasing the amount of training data to double (figure 2b), it means, two lines and two slices for training, the classification shows improved results by doing this. The misclassifications within the hemipelagic sediments of the deeper parts and the MTDs decreased considerably and the borders of both salt bodies were better delimited.

Figure 3c shows the results after increasing three times the analysis window of the Kuwahara filter applied to the seismic attributes. Notice that all the facies were better classified – the salt body, the MTDs, and the hemipelagic sediments. The biggest challenge continues to be the faults in the hemipelagic sediments over the salt body, which continue to be misclassified.

Figure 3d shows the classification results after cropping the volume into two smaller volumes. However, the results do not show any further improvement. On the contrary, an increased number of vertical artifacts being misclassified as salt facies within the MTDs are seen, as well as a stronger response of other previous misclassifications.

Finally, after performing the facies classification using training polygons picked in areas that are more complex, such as those that are highly faulted (figure 3e), we were able to eliminate the misclassifications in the hemipelagic sediments strongly faulted over both salt domes.

Notice that for almost all the classifications, the attributes k1-most positive curvature, envelope and coherence rank within the best combination.

Discussion

The initial results, in which the sedimentary strata of the deeper part were classified as salt, can be explained by a possible effect of amplitude attenuation with depth, which makes the reflectors look dimmer, therefore, a lower response in the envelope attribute, similar to that seen for the salt facies. This was overcome by adding more training data as well as using a larger analysis window of the Kuwahara filter. On the other hand, the vertical artifacts within the MTDs, and the fault areas that were incorrectly classified as salt facies can be related to a high k1-most positive curvature response, like that related to the salt domes. The poor results obtained after cropping the seismic volume may be related to an increase in the confidence of the algorithm by having the same amount of training data but a smaller volume, therefore, a stronger response of incorrect classifications. With the last parameter, the way in which the geoscientist selects the area for training, we could demonstrate the importance of helping the machine by training in areas with more complex geology, so that the algorithm can finish the classification more accurately in the less challenging areas.

Conclusions

PNNs were shown to be a great tool for seismic facies classification in seismic data. However, applying correct initial parameters and the geoscientist’s insight and understanding of the geology, and any acquisition and processing artifact, is key to obtaining accurate results. It is advisable to use at least two lines and two slices for training, and a larger analysis window when working with larger scale features, such as those classified here. It is also recommended to select the polygons for training within areas more geologically complex. Cropping the seismic volume did not show any improvement, and the vertical artifacts within MTDs were not correctly classified. We can help the machine-learning performance by varying different parameters, but if the artifacts persist, we could take them to be the associated limitations that may not be further overcome, unless there is a better-quality seismic survey.

Acknowledgements

We want to thank U.S. Geological Survey and the Bureau of Energy Management for providing the dataset. We would also like to thank the sponsors of Attribute Assisted Seismic Processing and Interpretation consortium software at the University of Oklahoma, which was used for the seismic attributes and machine learning. We also thank to Schlumberger for the software licenses provided.

Comments (1)

A couple of editorial comments and then a suggestion for improvements
Some of the references to figures in the text do not match the figures. "Parameters" as used in this article are called "features" in machine learning parlance. "Parameters" refers to the operating parameters of the ML algorithm. In this case it would be number of layers, number of "neurons" per layer, training method, whether or not simulated annealing was employed, etc. None of this sort of info is found in the article. Inputs to this neural net should have included time, and inline number and crossline number for each trace. That would essentially encode a 3D component to the training set which will (greatly?) assist in obtaining more successful classification.
8/9/2022 11:11:52 AM

You may also be interested in ...