[S-EM12] Electric, magnetic and electromagnetic survey technologies and scientific achievements

[SEM12-P03] A sparse joint inversion method based on the group lasso regularization

*Mitsuru Utsugi¹ (1.Institute for Geothermal Sciences, Graduate School of Science, Kyoto University)

Keywords:magnetic anomaly, gravity anomaly, multi-data, multi-parameter, joint inversion

In subsurface structure exploration, joint inversion methods using different observation data, such as magnetic and gravity anomalies, have been used in the present studies. Recently, improvements in computing environment have made it possible to compute such large-scale problems, and the joint inversion of multi-data and multi-parameter using such as cross gradient, fuzzy c-means, and correspondence maps has been actively studied. In contrast, this study proposed a joint inversion method based on a regularization method called group lasso. The group lasso is a regularization method in which the model elements can be divided into several groups (clusters), and the elements belonging to each group are constrained to take zero or non-zero values together. In the case of this study, we attempted to apply group lasso to groups of model elements such as density, magnetization, and resistivity assigned to each grid cell divided into the subsurface space. This is expected to derive a highly correlated model with a similar shape, because all model elements are simultaneously take zero or non-zero for each block. The advantages of this method are: 1. ease of implementation, and 2. the ability to additionally impose sparsity on the model. In particular, the optimization for the non-linear and non-derivative group lasso penalty can be solved analytically using the proximity gradient method, and thus can be implemented with a very simple code. Since the group lasso penalty is a vector version of the L1 norm penalty and is a kind of sparse regularization method, it is expected that the sparseness is introduced into the derived model by this inversion. In this presentation, we present the details of the calculation method, the results of the synthetic test, and the result of its application to gravity and magnetic anomaly data obtained in the Hiraniwa plutonic intrusion of the Kitakami Belt, northeastern Japan. We also report the results of three-component magnetic inversion by applying group lasso to the three components of magnetization of each block as a group.