Research on magnetotelluric square wave noise suppression based on a combination of long short-term memory and deep neural networks

Abstract Measured magnetotelluric data are frequently subject to interference from square wave noise, which, due to its large temporal scale, can significantly affect the accuracy of impedance estimation across a wide frequency band. Since both the effective signals and square wave noise are non-stationary, with the latter exhibiting variable bandwidth, conventional time-domain and frequency-domain filtering, as well as robust estimation methods, face significant limitations in processing accuracy. To address these challenges, this paper proposes a method for square wave noise identification and removal through the collaborative use of long short-term memory (LSTM) networks and deep neural networks (DNN). The LSTM network, with its capability to model multiscale temporal features, can adapt to square wave noise of varying temporal scales and accurately localize the periods of noise occurrence. On this foundation, the DNN further concentrates on the identified noise regions to precisely extract and suppress noise components, thereby preventing excessive correction of the effective signals. Validation with synthetic and field data demonstrates that this method surpasses traditional approaches in both noise identification accuracy and denoising precision, significantly enhancing the quality of apparent resistivity and phase curves affected by square wave noise, thereby providing strong support for improving the reliability and resolution of magnetotelluric exploration.