A method for two-dimensional imaging of magnetotelluric based on visual self-attention mechanism

Abstract

This invention belongs to the field of geophysical exploration technology, and specifically relates to a magnetotelluric two-dimensional imaging method based on a visual self-attention mechanism. The method of this invention is an end-to-end deep learning imaging approach, which has a significant advantage in computational time compared to traditional inversion methods. By constructing a geophysical theoretical model generation program, diverse geoelectric theoretical models are generated in batches, ensuring the scale and diversity of training samples. By introducing a pre-training process, the self-attention mechanism, with its stronger modeling ability and better ability to capture relative relationships in the data compared to convolutional structures, improves the model's generalization ability. Addressing the problem of missing information about deep anomalies when using only a single mode response data, a magnetotelluric two-dimensional imaging network suitable for TE / TM joint modes is proposed. By optimizing and improving the relative position encoding in the network, the model can correctly acquire and fuse input information from both TE and TM modes.

Images