Noise learning of audio magnetotellurics data with a combination of inception multi-scale residual and dual-decoder U-Net

Abstract

Abstract Audio magnetotellurics (AMT), as a commonly used reliable geophysical technique, provides outstanding metal ore exploration capabilities based on the resistivity structure of the earth. However, the accuracy of data acquisition is significantly impacted by external electromagnetic (EM) noise caused by heavy industrial activities and high-speed railway construction. Therefore, accurately learning the characteristics of noise and acquiring high-quality data becomes particularly crucial. In order to accurately extract noise features, such as triangular waves and sharp impulses commonly found in the time domain, we propose an AMT data noise learning method based on inception multi-scale residual dual-decoder U-Net (IMRDDU-Net). First, we input both the noisy dataset and the noise dataset into the IMRDDU-Net, aiming to train the network. Second, the encoder in the IMRDDU-Net, composed of inception and downsampling modules, learns and extracts feature from the noisy dataset. Then, the main decoder aggregates features through skip connections and multi-scale residual dense block (MRDB), while the auxiliary decoder independently reconstructs features. Finally, the outputs from the two decoders are fused to form a nonlinear mapping between the noisy data and the noise data, achieving the separation of high-quality data and noise data. In the simulation experiments, the similarity between the denoised data and the original high-quality data reached 98%. The proposed method was applied to AMT measured data from the Tongling ore cluster area. The resulting apparent resistivity curves are smoother and more continuous. These results demonstrate improved learning of AMT noise characteristics and a significant enhancement in denoising effectiveness and reliability.