Details of the Abstract
| Title of paper | Magnetotelluric 3D inversion solution using Huber loss function |
| List of authors | Smirnov, M.Yu. , Rydman, O. |
| Affiliation(s) | Luleå University of Technology, Sweden |
| Summary | Solving the MT inverse problem remains a challenge, especially in 3D. One of the problems that needs to be addressed is dealing with noisy and therefore inconsistent data. Here we refer to bias error, which is not reflected in random error estimation. Since the first works of Huber(1964), who introduced the M-estimators in robust statistics, attempts has been made to use robust norms of data term (residuals) when solving inverse problems (Scales, 1988). Robust M-estimators can be implemented using iteratively reweighted least squares (IRLS) scheme. The implementation is straight forward but requires special attention to evaluate weights at every IRLS iteration. In our implementation weight W are period dependent and estimated after the solution of inverse problem is obtained (gauss-newton type iterations converged). Typically, 2-3 IRLS iterations are sufficient to significantly reduce the influence of “bad” data. |
| Session Keyword | 2.0 EM theory, modelling and Inversion |
| File upload |
2.0_magnetotelluric_3d_invers_smirnov.pdf
|