[M-GI34] Data-driven geosciences

[MGI34-P03] Simultaneous estimation of continuous and discontinuous variables from observation images using a Markov-random-field (MRF) model

*Tatsu Kuwatani¹, Kenji Nagata², Masato Okada³ (1.Japan Agency for Marine-Earth Science and Technology (JAMSTEC), 2.National Institute for Materials Science (NIMS), 3.Graduate School of Frontier Sciences, The University of Tokyo)

Keywords:Bayesian estimation, Geofluid

Quantitative geological interpretation of geophysical observation data is essential in geosciences. In such inversion problems, geological unknown parameters can be modeled as either continuous variables or discontinuous variables. For example, Iwamori et al. (2021, JGR) assumed geometrical property and amount of geofluids as continuous variables, whereas they assumed the type of fluid (melt or aqueous fluid) and lithology as discontinuous variables. This study develops a general method that simultaneously extracts spatial structures of continuous and discontinuous variables from observed images by extending the Markov random field (MRF) modeling, a Bayesian image-processing method which effectively uses a spatial continuity of physical quantities. Specifically, we use a Potts spin for a discontinuous variable and introducing it into the Markov random field as a region-based latent variable. In this presentation, we will briefly introduce the preliminary synthetic inversion tests that simultaneously estimate the continuous fluid distribution and the discontinuous lithological structure from the seismic velocity structure.