TE/TM Pattern Recognition Based on Convolutional Neural Network

An important task in interpreting 2-D magnetotelluric (MT) sounding is to correctly identify the transverse electric (TE) and transverse magnetic (TM) polarization apparent resistivity curves. Traditional pattern recognition methods typically rely on curve morphology for identification. However, traditional methods often involve subjective intervention and require adjustments to deal with different situations, thus presenting limitations. Aiming to solve such problem, we applied neural networks in the pattern recognition process and developed a new TE/TM identification method based on convolutional neural network (CNN), which does not require prior information. We trained the network using data obtained from 2-D MT forward modeling of a series of representative geological models. To validate the stability and accuracy of the algorithm, we conducted synthetic experiments and tested it by field data. We find that in model experiments, the CNN method can better identify TE/TM, with significantly higher accuracy compared to traditional methods. Finally, we applied the method to the North American COPROD2 surveyed line data. By randomly shuffling the collected TE/TM curves and using our method, we were able to correctly recover the original arrangement order before shuffling, demonstrating the stability and accuracy of the new method.