Transient electromagnetic data denoising based on cluster analysis and locally weighted linear regression

Abstract

In transient electromagnetic surveys, the collected data inevitably contain noise originating from both natural and cultural sources. This noise has the potential to mask transient electromagnetic responses linked to geological features, thereby posing challenges in accurately interpreting subsurface structures. Hence, the implementation of effective noise reduction techniques is crucial in ensuring the accuracy and reliability of inversion outcomes in transient electromagnetic surveys. This study introduces a novel approach that merges k-means clustering with locally weighted linear regression to denoise transient electromagnetic data. The results from synthetic examples illustrate that the k-means locally weighted linear regression method can predict transient electromagnetic data closely resembling true values, similar to the long short-term memory autoencoder. Occam's inversion results derived from denoised data using both the k-means locally weighted linear regression and long short-term memory–autoencoder methods can well reflect the true model. Notably, a key advantage of the k-means locally weighted linear regression method is its independence from labelled data as the sample set. The k-means locally weighted linear regression method was applied to field data collected at the Narenbaolige coalfield in Inner Mongolia, China. Occam's inversion models generated from the denoised field data delineate the boundary between the basaltic body and sedimentary rocks, aligning with drilling data. The inversion models derived from the noisy field data also can capture this boundary, but deep section views reveal the presence of numerous intricate high-resistivity anomalous bodies. These observations highlight the effectiveness of the k-means locally weighted linear regression method in denoising transient electromagnetic data.