GTCN: Gated Temporal Convolutional Networks for Controlled-Source Electromagnetic Data Denoising

To improve the signal-to-noise ratio (SNR) of controlled-source electromagnetic (CSEM) data observed in strong interference environments, a new deep learning network is proposed and named gated temporal convolutional network (GTCN) to map noisy sequences to high-quality sequences. This network is an improvement of two state-of-the-art (SOTA) networks specifically designed for time series processing, temporal convolutional network (TCN) and gated recurrent units (GRUs). A carefully crafted sample set is created by utilizing shift-invariant sparse coding (SISC) methods and used to train the newly proposed network and six other SOTA deep learning networks. Experimental results of the synthetic data indicate that the new network not only outperforms SISC in accuracy and efficiency but also is significantly superior to the other six SOTA deep learning methods. The proposed GTCN method can improve the 0 dB noisy signals to 32.6749 dB and improve the average SNR from −5 to 23.5999 dB. The effectiveness and reliability of the proposed method are also verified through measured data from Sichuan and Yunnan, China. The time series processed by the new approach exhibits more pronounced periodic characteristics, resulting in smoother and more continuous apparent resistivity curves. All these experiments demonstrate that the new scheme is an effective method to improve the quality of CSEM data and contribute to the reliability of CSEM exploration.

I. Introduction

The controlled-source electromagnetic (CSEM) method is widely applied in resource exploration and engineering geological surveys [1], [2], [3], [4]. However, CSEM signals are inevitably contaminated by various levels of cultural noise when collected in strong interference environments. Strong noise is likely to reduce the resolution, interpretability, and reliability of the exploration results. Therefore, high-precision removal of cultural noise becomes increasingly important [5], [6], [7], [8].