[Paper Review] Generating Realistic Geology Conditioned on Physical Measurements with Generative Adversarial Networks

Emilien Dupont, Tuanfeng Zhang+3 more|arXiv (Cornell University)|Feb 8, 2018

Generative Adversarial Networks and Image Synthesis15 references58 citations

TL;DR

The paper uses semantic inpainting with GANs to generate realistic geological realizations that honor sparse physical measurements and capture pattern variability.

ABSTRACT

An important problem in geostatistics is to build models of the subsurface of the Earth given physical measurements at sparse spatial locations. Typically, this is done using spatial interpolation methods or by reproducing patterns from a reference image. However, these algorithms fail to produce realistic patterns and do not exhibit the wide range of uncertainty inherent in the prediction of geology. In this paper, we show how semantic inpainting with Generative Adversarial Networks can be used to generate varied realizations of geology which honor physical measurements while matching the expected geological patterns. In contrast to other algorithms, our method scales well with the number of data points and mimics a distribution of patterns as opposed to a single pattern or image. The generated conditional samples are state of the art.

Motivation & Objective

• Motivate the need for realistic subsurface models given sparse physical measurements.
• Propose a GAN-based semantic inpainting framework to generate multiple realizations of geology conditioned on measurements.
• Demonstrate that the method scales with data and captures a distribution of patterns rather than a single realization.
• Show applicability beyond geology, e.g., to satellite imagery and other geospatial data.

Proposed method

• Train a DC-GAN on a distribution of binary fluvial pattern images to learn p_fluvial(z).
• Represent conditioning as a prior loss Lp(z)=log(1−D(G(z))).
• Represent conditioning on known pixels with a context loss Lc, expanded to propagate constraints spatially via an expanded mask and values.
• Minimize the total loss L(z)=L̃c(z|y,M)+λLp(z) to obtain realizations that honor measurements and look realistic.
• Generate multiple realizations by initializing z randomly and optimizing each to different local minima of L(z).
• Optionally apply the framework to satellite images by training on Landsat-derived data and conditioning on land/water labels.

Experimental results

Research questions

• RQ1 Can a GAN-based semantic inpainting approach generate realistic geological realizations that honor sparse measurements?
• RQ2 Does conditioning on physical measurements scale to many data points and still produce coherent, varied realizations?
• RQ3 How does the proposed method compare to existing training-image-based methods like snesim in terms of realism and coherence?
• RQ4 Can the approach generalize to other geospatial data such as satellite imagery?

Key findings

• Conditional samples are realistic and coherent, honoring measurements across various conditioning densities.
• The method yields samples that are competitive with state-of-the-art while producing realistic channel connectivity and diversity absent in some baselines.
• Increasing the number of measurements does not necessarily degrade realism; the method can handle large conditioning sets.
• Latent-space traversals show smooth changes and coverage of the data manifold, indicating learned pattern variability.
• Applying the approach to Landsat satellite data yields convincing conditioned realizations of land/water patterns.