Enhancing Magnetotelluric Data Quality Using Deep Learning-Based Denoising Models: A Study of CNN and LSTM Academic Article uri icon

abstract

  • Noise interference in Magnetotelluric (MT) signals significantly undermines the accuracy of subsurface resistivity analysis, leading to potential errors in geophysical interpretation and challenges in resource exploration. To address this critical issue, this study develops denoising models based on Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) to enhance the quality of MT signals while preserving their original structure. The proposed models utilize MT data, incorporating electric field (E) and magnetic field (H) components, to effectively reduce noise. Evaluation results demonstrate that CNN outperforms LSTM, increasing the Signal-to-Noise Ratio (SNR) by up to 58.8% (or 1.6 times) in the Hx channel. CNN also records lower Normalized Mean Square Error (NMSE) values across all channels, ranging from 0.006 to 0.033, while maintaining a high correlation coefficient of 0.999 in the Hz channel. Moreover, CNN is significantly faster, with processing times of 24.83 to 29.16 seconds—up to three times faster than LSTM, which requires 67.38 to 70.69 seconds. The superior performance of CNN in mitigating noise in MT data is attributed to its architecture, which focuses on local patterns. This makes it particularly effective for handling localized and sporadic noise, as observed in the Hx and Hy channels with recurring amplitude patterns. In contrast, LSTM is less effective for MT data with unstructured noise, such as in the Hz channel, due to its sequential approach, which is better suited for capturing long-term temporal relationships. This study aligns with the Sustainable Development Goals (SDGs), specifically SDG 9, which promotes innovative applications of deep learning technology, and SDG 7, which emphasizes improving the accuracy of renewable energy exploration. The findings provide a foundation for developing more adaptive and efficient denoising models, contributing to environmental sustainability and the advancement of clean energy exploration in the future.

publication date

  • 2025

number of pages

  • 26

start page

  • 399

end page

  • 425

volume

  • 6

issue

  • 2