High-resolution Magnetotelluric Data Inversion Constrained with Seismic Texture

Magnetotelluric (MT) data inversion reconstructs the subsurface resistivity structure with the measured natural electromagnetic (EM) field. Due to the attenuation of low frequency EM waves in geological conductors, MT inversion has relatively low sensitivity and resolution to subsurface structures. The reflection seismic section provides abundant and intricate structural information complementary to the MT data, which has the potential to improve the resolution of MT data inversion when utilized appropriately. In this work, we study a high-resolution MT data inversion constrained with the seismic texture. An effective operator for geophysical texture extraction is constructed based on a VGG19 neural network (NN) trained on ImageNet, a general image processing dataset. The texture measurement is constructed by the mean and variance of feature maps obtained from specific convolutional layers. The inversion is implemented based on the TensorFlow framework, and a composite objective function consisting the MT data misfit, texture discrepancy and two Tikhonov regularization penalties is optimized with Adam algorithm. Both numerical and field data inversion show that the proposed inversion yields a resistivity model presenting a fine texture consistent with the seismic section and a decreased or at least unchanged MT data misfit.