TEM data denoising based on cluster analysis and locally weighted linear regression

Transient electromagnetic (TEM) field data typically contain noises. When the noises are not properly dealt with, it can lead to unreliable inversion results. TEM data typically have a high signal-to-noise ratio (SNR) in early times and the SNR deteriorate over time. We present a novel denoising algorithm specifically designed for TEM data. The denoising method uses locally weighted linear regression (LWLR) to predict noise-free TEM data for a given time channel using TEM data collected at adjacent time channels. To obtain a better prediction result, we employed the k-means cluster analysis to determine the optimal width for the Gaussian kernel used by the LWLR algorithm. Synthetic tests showed that our algorithm can effectively eliminate the added noise in the TEM data. We also used one-dimensional (1D) inversion to invert the noise-contaminated data and the denoised data. We found that the inversion constructed model agrees better with the true model compared to the constructed model of the noise-contaminated data.