Abstract: High-resolution marine magnetotelluric inversion has always been an important topic in the field of marine magnetotelluric sounding. To improve the accuracy of marine magnetotelluric inversion, by taking advantage of the high imaging accuracy of marine seismic detection methods, the velocity structure obtained by deep seismic detection as prior information was added into the convolutional neural network magnetotelluric inversion. By constructing two-channel data as input of inversion network, the velocity-constrained convolutional neural network magnetotelluric inversion was realized. Based on the proposed method, data sets established for geological profiles of the Qianliyan region in the South Yellow Sea were trained. Results show that the proposed method could improve the accuracy and the vertical resolution of the inversion. Moreover, the results of anti-noise tests also show that the proposed method could achieve higher resolution inversion for noisy data sets. This study provided a new idea for high-resolution marine magnetotelluric inversion.