Random Forests and Support Vector Machine comparison for Digital Model Outcrop classification tasks .

@inproceedings{Leprince2017,
 abstract = { Digital Outcrop Models (DOMs) have a potential to significantly affect how geologists do field work. DOMs can produce very good outcrop coverage and quantification at a relatively modest cost. DOMs also enable a shift from qualitative field observations to quantitative static and dy- namic models deemed analog to subsurface reservoirs. However, moving from a high resolution DOM to an interpreted set of fractures or facies raises technical challenges as it still relies on signif- icant manual interpretation and is subject to uncertainties when it comes to shape extrapolations in three dimensions. For this reason, we report on a methodology where we started to visualize large-scale DOMs and analyze them using machine learning techniques in a plugin of the 3D ge- omodeling software SKUA-GOCAD. In this project we mainly focus on a part of a 10GB DOM acquired by orthophotogrammetry from drone pictures in Utah, USA. More specifically, we want to classify different geological features from DOM attributes such as RGB values, dip, azimuth, etc. For this, we interfaced the OpenCV C++ implementation of Support Vector Machine (SVM) and Random Forests (RF) algorithms in our SKUA-GOCAD plugin. We compare results obtained with these two methods for the basic problem of screening the DOM into rock, scree and vegetation, using training sets of varying sizes. Results suggest that RF provide a more robust classification than SVM and also gives some insights about how to select appropriate training data. We discuss some important perspectives about feature selection and about further avenues for DOM analysis in geomodeling software. 1 },
 author = { Leprince, Yrieix AND Caumon, Guillaume AND Cregu, U L Cnrs },
 booktitle = { 2017 RING Meeting },
 pages = { 1--11 },
 publisher = { ASGA },
 title = { Random Forests and Support Vector Machine comparison for Digital Model Outcrop classification tasks . },
 year = { 2017 }
}