TEM-NLnet: A Deep Denoising Network for Transient Electromagnetic Signal With Noise Learning

Transient electromagnetic (TEM) method is a widely adopted technology in geophysics. TEM signals received by coils will be disturbed by complex noises. Compared with traditional filtering-based methods, deep-learning-based TEM signal denoising methods achieved impressive denoising performance. However, the existing deep-learning-based methods rely heavily on simulated noise with a certain distribution to construct paired datasets for supervised learning. In real scenarios, if the noise distribution of acquired TEM signals has a huge difference (e.g., the type of noise distribution, the level of noise) with that of the simulated datasets, the trained model may not always be valid. To address this issue, a novel noise-learning-inspired deep denoising network (namely, TEM-NLnet) is proposed for TEM signal denoising. Specifically, instead of inserting the simulated noise, we first learn the noise appeared in real-world signals through generative adversarial networks (GANs), such that the generator can produce the learned noise to construct paired datasets for training. Then, a deep-neural-network-based denoiser is imposed to learn mapping from the noise TEM signal to the corresponding noise-free one. Extensive experiments on the simulated and actual geological datasets show that compared with other state-of-the-art TEM denoising methods, our proposed method achieves better performance in terms of quantitative and visual results. Models and code are available at https://github.com/wmyCDUT/TEM-NLnet_demo.