An image clustering framework to simplify multi-realization output.

Solene Panhaleux and Thomas Viard and Guillaume Caumon. ( 2010 )
in: Proc. 30th Gocad Meeting, Nancy

Abstract

Data interpretation in geosciences often relies on indirect subsurface observations. Hence geoscientists have to solve the inverse problem of finding which subsurface state resulted in the observed set of measures. As the solution is most often not unique, many realizations that solve the problem are typically generated, e.g. in seismic inversion for rock facies classification. Such realizations can be considered as possible 3D images of the subsurface. Accurate interpretation needs to take into account all images produced with this methodology, which is a challenging task because of the possibly large number of such images. The classification of those images into families can then be valuable for the selection of the most relevant ones, which is likely to make the interpretation easier. The aim of this project is to develop an efficient methodology to classify those images. A workflow in three steps is proposed. First, we perform a pre-processing of the data through image segmentation; we use a k-means algorithm. Then, we compute the similarity between the images using Hausdorff distance; this step is meant to determine global trends in the pretreated images. All the images are then grouped according to their similarity. Selecting typical images within each family is then straightforward and can save a lot of CPU power when running costly algorithms through the images, such as flow simulation or time-to-depth migration for instance.

Download / Links

BibTeX Reference

@INPROCEEDINGS{PanhaleuxGM2010,
    author = { Panhaleux, Solene and Viard, Thomas and Caumon, Guillaume },
     title = { An image clustering framework to simplify multi-realization output. },
 booktitle = { Proc. 30th Gocad Meeting, Nancy },
      year = { 2010 },
  abstract = { Data interpretation in geosciences often relies on indirect subsurface observations. Hence geoscientists have to solve the inverse problem of finding which subsurface state resulted in the observed set of measures. As the solution is most often not unique, many realizations that solve the problem are typically generated, e.g. in seismic inversion for rock facies classification. Such realizations can be considered as possible 3D images of the subsurface. Accurate interpretation needs to take into account all images produced with this methodology, which is a challenging task because of the possibly large number of such images. The classification of those images into families can then be valuable for the selection of the most relevant ones, which is likely to make the interpretation easier.
The aim of this project is to develop an efficient methodology to classify those images. A workflow in three steps is proposed. First, we perform a pre-processing of the data through image segmentation; we use a k-means algorithm. Then, we compute the similarity between the images using Hausdorff distance; this step is meant to determine global trends in the pretreated images. All the images are then grouped according to their similarity. Selecting typical images within each family is then straightforward and can save a lot of CPU power when running costly algorithms through the images, such as flow simulation or time-to-depth migration for instance. }
}