Image synthesis with graph cuts: a fast model proposal mechanism in probabilistic inversion

Abstract

Geophysical inversion should ideally produce geologically realistic subsurface models that explain the available data. Multiple-point statistics is a geostatistical approach to construct subsurface models that are consistent with site-specific data, but also display the same type of patterns as those found in a training image. The training image can be seen as a conceptual model of the subsurface and is used as a non-parametric model of spatial variability. Inversion based on multiple-point statistics is challenging due to high nonlinearity and time-consuming geostatistical resimulation steps that are needed to create new model proposals. We propose an entirely new model proposal mechanism for geophysical inversion that is inspired by texture synthesis in computer vision. Instead of resimulating pixels based on higher-order patterns in the training image, we identify a suitable patch of the training image that replace a corresponding patch in the current model without breaking the patterns found in the training image, that is, remaining consistent with the given prior. We consider three cross-hole ground-penetrating radar examples in which the new model proposal mechanism is employed within an extended Metropolis Markov chain Monte Carlo (MCMC) inversion. The model proposal step is about 40 times faster than state-of-the-art multiple-point statistics resimulation techniques, the number of necessary MCMC steps is lower and the quality of the final model realizations is of similar quality. The model proposal mechanism is presently limited to 2-D fields, but the method is general and can be applied to a wide range of subsurface settings and geophysical data types.

References in corpus (1)

• Summary statistics from training images as prior information in probabilistic inversion

Cited by in corpus (7)

• Training-image based geostatistical inversion using a spatial generative adversarial neural network
• Inversion using a new low-dimensional representation of complex binary geological media based on a deep neural network
• Integration of adversarial autoencoders with residual dense convolutional networks for estimation of non-Gaussian hydraulic conductivities
• Deep generative models in inversion: a review and development of a new approach based on a variational autoencoder
• Adaptive sequential Monte Carlo for posterior inference and model selection among complex geological priors
• Diffusion models for multivariate subsurface generation and efficient probabilistic inversion
• Geological modeling using a recursive convolutional neural networks approach