DP-TTA: Test-time Adaptation for Transient Electromagnetic Signal Denoising via Dictionary-driven Prior Regularization

TL;DR

This work tackles the domain-shift problem in transient electromagnetic (TEM) signal denoising by introducing DP-TTA, a test-time adaptation framework that leverages a dictionary-driven prior of intrinsic TEM characteristics such as exponential decay and smoothness. A CNN-based encoder-decoder (DTEMDNet) is augmented with a dictionary-learning branch to enforce a prior through sparse codes, while test-time adaptation uses three self-supervised losses—denoising consistency, sparse-code consistency, and one-order variation regularization—to adapt to new domains without labeled target data. The approach demonstrates superior performance over traditional TEM denoisers and existing TTA methods on both simulation and real-world TEM datasets, with notable gains in SNR and improved robustness to domain shift. Overall, DP-TTA provides a principled and effective mechanism for robust TEM signal denoising in diverse field conditions, with potential applicability to other signals possessing intrinsic, domain-invariant properties.

Abstract

Transient Electromagnetic (TEM) method is widely used in various geophysical applications, providing valuable insights into subsurface properties. However, time-domain TEM signals are often submerged in various types of noise. While recent deep learning-based denoising models have shown strong performance, these models are mostly trained on simulated or single real-world scenario data, overlooking the significant differences in noise characteristics from different geographical regions. Intuitively, models trained in one environment often struggle to perform well in new settings due to differences in geological conditions, equipment, and external interference, leading to reduced denoising performance. To this end, we propose the Dictionary-driven Prior Regularization Test-time Adaptation (DP-TTA). Our key insight is that TEM signals possess intrinsic physical characteristics, such as exponential decay and smoothness, which remain consistent across different regions regardless of external conditions. These intrinsic characteristics serve as ideal prior knowledge for guiding the TTA strategy, which helps the pre-trained model dynamically adjust parameters by utilizing self-supervised losses, improving denoising performance in new scenarios. To implement this, we customized a network, named DTEMDNet. Specifically, we first use dictionary learning to encode these intrinsic characteristics as a dictionary-driven prior, which is integrated into the model during training. At the testing stage, this prior guides the model to adapt dynamically to new environments by minimizing self-supervised losses derived from the dictionary-driven consistency and the signal one-order variation. Extensive experimental results demonstrate that the proposed method achieves much better performance than existing TEM denoising methods and TTA methods.

Figures (11)

• Figure 1: Illustration of the domain shift problem. In TEM applications, a domain represents a specific data acquisition environment. For example, high-altitude plateaus, hilly mineral exploration zones, and urban survey areas often exhibit distinct noise characteristics due to differences in geological conditions, equipment, and external interference, etc. Existing models trained on a specific simulation environment (a.k.a., source domain) are difficult to generalize to other environments (a.k.a., target domain), as the discrepancy (i.e., domain shift) of noise characteristics between two domains causes performance degradation. As illustrated in the bottom of the figure, test-time adaptation strategy aims to mitigate this issue by dynamically adjusting the pre-trained model during the testing phase, resulting in better adaptation to new environment with improved denoising performance.
• Figure 2: Overview of the proposed framework. (a) illustrates the complete DTEMDNet architecture integrating the TEM signal dictionary. In the source-domain training, DTEMDNet is pre-trained using two supervised losses. (b) demonstrates DP-TTA, where the DTEMDNet performs dynamic self-adaptation at test time by applying three self-supervised losses (Denoising result, sparse code, one-order variation). Note: Input 2D-Signal denotes the 2D image obtained using the 1D to 2D conversion method in9258400. The detailed definitions of the above losses are provided in later sections.
• Figure 3: Simulation samples. (a) Source dataset sample: $Q_1 = 1300$, $Q_2 = 2.5$, $B = 4.0$, $SNR = 22$. (b) AGN dataset sample: $Q_1 = 600$, $Q_2 =3.6$, $B = 3.5$, $SNR = 10$. (c) LFI dataset sample: $f = 1.5$, $A = 30$. (d) HFI dataset sample: $f = 50$, $A = 30$. (e) IMP dataset sample: $k = 30$, $A_i = 70$. (f) CMP dataset sample.
• Figure 4: Specific location of TEM signal acquisition. Dazhou data acquisition region, with AB length of 0.97 km and the blue dotted line spanning 500 m. The bottom and left panels provide the geographic coordinates of the sampling area.
• Figure 5: Training loss curves of different denoising models on the source domain dataset. DTEMDNet exhibits faster and more stable convergence compared with other baselines, which can be attributed to the guidance of the dictionary-driven prior.