DREMnet: An Interpretable Denoising Framework for Semi-Airborne Transient Electromagnetic Signal

TL;DR

This work addresses SATEM signal denoising under complex noise by introducing DREMnet, an interpretable disentangled representation framework that splits data into content $Z_s$ and context $Z_n$. The model uses an RWKV-based encoder–decoder architecture with Contextual-WKV for bidirectional temporal modeling and Cover Embedding to preserve local signal details, along with a mutual-information-based objective (CLUB) and a swapping mechanism to enforce disentanglement. Experiments on large synthetic resistivity-model datasets and field data show DREMnet outperforms CNN-, diffusion-, and transformer-based denoisers in denoising quality (higher SNR, lower MSE, better SSIM) and yields inversion results that align more closely with theoretical signals, revealing subsurface structures more clearly. The proposed approach provides interpretable denoising in noisy SATEM environments and holds promise for improving large-scale, hard-to-access surveys in mineral exploration and groundwater studies.

Abstract

The semi-airborne transient electromagnetic method (SATEM) is capable of conducting rapid surveys over large-scale and hard-to-reach areas. However, the acquired signals are often contaminated by complex noise, which can compromise the accuracy of subsequent inversion interpretations. Traditional denoising techniques primarily rely on parameter selection strategies, which are insufficient for processing field data in noisy environments. With the advent of deep learning, various neural networks have been employed for SATEM signal denoising. However, existing deep learning methods typically use single-mapping learning approaches that struggle to effectively separate signal from noise. These methods capture only partial information and lack interpretability. To overcome these limitations, we propose an interpretable decoupled representation learning framework, termed DREMnet, that disentangles data into content and context factors, enabling robust and interpretable denoising in complex conditions. To address the limitations of CNN and Transformer architectures, we utilize the RWKV architecture for data processing and introduce the Contextual-WKV mechanism, which allows unidirectional WKV to perform bidirectional signal modeling. Our proposed Covering Embedding technique retains the strong local perception of convolutional networks through stacked embedding. Experimental results on test datasets demonstrate that the DREMnet method outperforms existing techniques, with processed field data that more accurately reflects the theoretical signal, offering improved identification of subsurface electrical structures.

Figures (11)

• Figure 1: The overall framework, the data is encoded by the encoder $E$ into content factors and context factors. $G_s$ is used to obtain an accurate clean signal, while $G_n$ ensures the correctness of the disentangled representations.
• Figure 2: The Encoder and Decoder architecture (on the left) and the DR block (on the right) consist of the signal mixing module and the channel mixing module.
• Figure 3: Cover Embedding, as illustrated in the case where the cover length is 3, combines the current signal data along with the next two signal data into a token with a dimension of $C$.
• Figure 4: Model (left), forward modeling data (center), noise-added data (right).
• Figure 5: The clean signal $s$ is disentangled into signal factors $Z_s^1$ and noise factors $Z_n^1$, while the noisy signal $n$ is disentangled into signal factors $Z_s^2$ and noise factors $Z_n^2$. (a) The clean signal is decoded using $(Z_s^2, Z_n^1)$. (b) The noisy signal is decoded using $(Z_s^1, Z_n^2)$. (c) The MSE statistics of the decoded data from $(Z_s^1, Z_n^2)$ on the test set, compared to the noisy signal.