As you probably know by know, I've been at the EAGE conference in Vienna this week. I skipped out yesterday and flew over to the UK for a few days. I have already written plenty about the open source workshop, and I will write more soon about the hackathon. But I thought I'd pass on my highlights from the the Automated Interpretation session on Tuesday. In light of Monday's discussion, I made a little bit of a nuisance of myself by asking the same post-paper question every time I got the chance:
Can I use your code, either commercially or for free?
I'll tell you what the authors responded.
The universal character of salt
I especially enjoyed the presentation by Anders Waldeland and Anne Solberg (University of Oslo) on automatically detecting salt in 3D seismic. (We've reported on Anne Solberg's work before.) Anders described training eight different classifiers, from a simple nearest mean to a neural network, a supprt vector model, and a mixture of Gaussians classifier. Interestingly, but not surprisingly, the simplest model turned out to be the most effective at discrimination. He also tried a great many seismic attributes, letting the model choose the best ones. Three attributes consistently proved most useful: coherency, Haralick energy (a GLCM-based texture attribute), and the variance of the kurtosis of the amplitude distribution (how's that for meta?). What was especially interesting about his approach was that he was training the models on one dataset, and predicting on an entirely different 3D. The idea is that this might reveal the universal seismic characteristics of salt. When I asked the golden question, he said "Come and talk to me", which isn't a "yes", but it isn't a "no" either.
Secret horizon tracker
The most substantial piece of machine learning I saw was Eduardo Figueiredo from Pontifical Catholic University in Rio, in the same session as Waldeland. He's using a neural net called Growing Neural Gas to classify (aka or 'label') the input data in a number of different ways. This training step takes a little time. The label sets can now be compared to decide on the similarity between two samples, based on the number of labels the samples have in common but also including a comparison to the original seed, which essentially acts as a sort of brake to stop things running away. This progresses the pick. If a decision can't be reached, a new global seed is selected randomly. If that doesn't work, picking stops. Unfortunately he did not show a comparison to an ordinary autotracker, either in terms of time or quality, but the results did look quite good. The work was done 'in cooperation with Petrobras', so it's not surprising the code is not available. I was a bit surprised that Figueiredo was even unable to share any details of the implementation.
More on interpretation
The other two interesting talks in the session were two from Paul de Groot (dGB Earth Sciences) and Gaynor Paton (GeoTeric). Paul introduced the new Thalweg Tracker in OpendTect — the only piece of software from the session that you can actually run, albeit for a fee — which is a sort of conservative voxel tracker. Unsurprisingly, Paul was also very thorough with his examples, and his talk served as a tutorial in how to make use of, and give attribution to, open data. (I'm nearly done with the grumbling about openness for now, I promise, but I can't help mentioning that I find it a bit ironic that those scientists unwilling to share their work are also often a bit lax with giving credit to others whose work they depend on.)
Gaynor's talk was about colour, which you may know we enjoy thinking about. She had gathered 24 seismic interpreters, five of whom had some form of colour deficiency. She gave the group some interpretation tasks, and tried to gauge their performance in the tasks. It seemed interesting enough, and I immediately wondered if we could help out with Pick This, especially to help grow the sample size, and by blinding the study. But the conclusion seemed to be that, if there are ways in which colour blind interpreters are less capable at image interpretation, for example where hue is important, they compensate for it by interpreting other aspects, such as contrast and shape.
That's it for now. I have a few other highlights to share; I'll try to get to them next week. There was a bit of buzz around the Seismic Apparition talks from ETHZ and Statoil, for example. If you were at the conference, I'd love to hear your highlights too, please share them in the comments.
A.U. Waldeland* (University of Oslo) & A.H.S. Solberg (University of Oslo). 3D Attributes and Classification of Salt Bodies on Unlabelled Datasets. 78th EAGE Conference & Exhibition 2016. DOI 10.3997/2214-4609.201600880. Available in EarthDoc.
M. Pelissier (Dagang Zhaodong Oil Company), C. Yu (Dagang Zhaodong Oil Company), R. Singh (dGB Earth Sciences), F. Qayyum (dGB Earth Sciences), P. de Groot* (dGB Earth Sciences) & V. Romanova (dGB Earth Sciences). Thalweg Tracker - A Voxel-based Auto-tracker to Map Channels and Associated Margins. 78th EAGE Conference & Exhibition 2016. DOI 10.3997/2214-4609.201600879. Available in EarthDoc.
G. Paton* (GeoTeric). The Effect of Colour Blindness on Seismic Interpretation. 78th EAGE Conference & Exhibition 2016. DOI 10.3997/2214-4609.201600883. Available in EarthDoc.
A.M. Figueiredo* (Tecgraf / PUC-Rio), J.P. Peçanha (Tecgraf / PUC-Rio), G.M. Faustino (Tecgraf / PUC-Rio), P.M. Silva (Tecgraf / PUC-Rio) & M. Gattass (Tecgraf / PUC-Rio). High Quality Horizon Mapping Using Clustering Algorithms. 78th EAGE Conference & Exhibition 2016. DOI 10.3997/2214-4609.201600878. Available in EarthDoc.