Enhancing magnetotelluric data denoising using a layer-wise progressive extraction network

Abstract High-quality acquisition and analysis of magnetotelluric (MT) data are critical in geophysics and environmental monitoring. However, anthropogenic noise often degrades the reliability and accuracy of such data. To address this challenge, we propose a Progressive Extraction Network (PEN), a novel neural network model employing progressive extraction strategy. PEN integrates two key components: the Feature Extraction Network (FEN) and the Layer-wise Denoising Network (LDN). FEN utilizes sparse layers to extract robust features from noisy data, while LDN employs hierarchical addition layers to progressively extract noise, minimizing information loss and enhancing noise suppression. The PEN model is trained on synthetically generated datasets with significant nonlinear mappings, learning the relationship between noisy signals and noise components. Once trained, PEN isolates and subtracts noise from MT data, enabling effective signal restoration and noise suppression. Experimental evaluations under diverse noise conditions demonstrate the capability of PEN to significantly improve both denoising and data recovery, outperforming conventional approaches. This study highlights potential of PEN for advancing MT signal processing and improving the accuracy of geophysical investigations. Graphical abstract