Geomodeling Using Generative Adversarial Networks and a Database of Satellite Imagery of Modern River Deltas
Summary Several studies on deep generative models for use in geomodeling show encouraging results with binary training data. An important question is what type of training data to use, since realistic 3D geology with natural variability is difficult to create. The advent of multiple types of remote sensing data of subaerial and subaqueous sedimentary patterns provides new possibilities in this context. Here, we train a Wasserstein GAN using 20,000 multispectral satellite images of subsections of 40 modern river deltas. The generated output has three facies and a background facies, and all quantitative evaluation methods of the unconditional output show a close overlap between the model and training data distributions. Standard MCMC sampling conditional on soft and hard data works well as long as the likelihood model is balanced against the prior model. Transfer learning, i.e. fine-training a small subset of the network parameters on smaller dataset of interest, such as highly non-stationary images of river deltas with similar characteristics, also shows promising results.
Paper
Full text
Geomodeling Using Generative Adversarial Networks and a Database of Satellite Imagery of Modern River Deltas
Semantic Scholar · Computer Science · 2019
Abstract
Summary Several studies on deep generative models for use in geomodeling show encouraging results with binary training data. An important question is what type of training data to use, since realistic 3D geology with natural variability is difficult to create. The advent of multiple types of remote sensing data of subaerial and subaqueous sedimentary patterns provides new possibilities in this context. Here, we train a Wasserstein GAN using 20,000 multispectral satellite images of subsections of 40 modern river deltas. The generated output has three facies and a background facies, and all quantitative evaluation methods of the unconditional output show a close overlap between the model and training data distributions. Standard MCMC sampling conditional on soft and hard data works well as long as the likelihood model is balanced against the prior model. Transfer learning, i.e. fine-training a small subset of the network parameters on smaller dataset of interest, such as highly non-stationary images of river deltas with similar characteristics, also shows promising results.