A Magnetotelluric Data Denoising Method Based on Lightweight Ensemble Learning

Traditional magnetotelluric (MT) denoising methods often encounter limitations in various scenarios. However, with its robust adaptability and high precision, deep learning has exhibited outstanding denoising performance when applied to MT exploration time series data. Recent researches have mainly focused on developing advanced single deep learning models to enhance MT denoising effectiveness. This article introduces a lightweight ensemble learning approach for MT denoising, aiming to enhance denoising performance via a single deep convolutional network. Our ensemble learning strategy uses a sliding window technique to generate overlapping MT time series segments, thereby providing multiple inputs for a specialized noise-fitting network. This variety of inputs enables a comprehensive understanding of MT data, thereby increasing the probability of identifying complex noise patterns. Then, the outputs from these inputs are integrated using a method that combines shifting averages and adaptive thresholding to obtain more accurate fitted noise contours. Furthermore, we apply a three-layer density-based spatial clustering of applications with noise (DBSCAN) methodology to identify the real noise contours among the fitted noise contours and then to get the residual signal by subtracting those real noise contours. Subsequently, the residual signal is further processed by the pretrained denoising network to eliminate noise artifacts. The efficacy of our approach is validated through experiments conducted with both synthetic and field data, demonstrating substantial improvements in denoising, particularly within mid- and low-frequency ranges. Several interrelated parameters exhibit notable improvements, including apparent resistivity and phase curves, time-frequency domain curves, and so on.