Objective The transient electromagnetic (TEM) method serves as a primary geophysical technique for groundwater detection in coalfields at present, with the detection results directly influencing the implementation of water hazard prevention and control in coal mines. Since it is difficult to avoid electromagnetic interference sources such as power lines during data acquisition, TEM signals are prone to be contaminated by electromagnetic noise. Given that primary denoising techniques such as wavelet transform and empirical mode decomposition (EMD) are yet to be further improved, this study developed a novel Gaussian process regression (GPR)-based denoising method for TEM signals.

Methods First, time alignment was applied to the noisy TEM signals to ensure roughly equivalent signal magnitudes. Next, non-parametric regression-based fitting was performed for the aligned TEM signals using the radial basis function (RBF), aiming to capture the non-linear trend of the signals and separate noise. Finally, temporal alignment was reversed to obtain the final denoising results.

Results For theoretical TEM signals contaminated by four types of individual noise (e.g., sinusoidal, triangular wave, uniform, and Gaussian noise), the proposed method increased their signal-to-noise ratios (SNRs) by 24.61 dB‒36.03 dB and decreased their mean relative errors (MREs) by 5.93%‒9.06%. For TEM signals contaminated by two types of mixed noise, the proposed method increased their SNRs by 28.05 dB and 26.92 dB and decreased their MREs by 5.22% and 8.35%. For field experimental data from a coalfield, the new method increased their SNR by 18.76 dB and decreased their MRE by 175.92%. The oscillatory effects of noise in the induction curves of the survey points were significantly eliminated. Furthermore, the inverted resistivity sections of the survey line exhibited restored longitudinal geoelectric structures and lateral continuity of the strata. The experimental results were roughly consistent with those of noise-free TEM signals. Compared to the wavelet transform, the proposed method exhibited a significant improvement in the denoising results.

Conclusions The GPR-based denoising algorithm yields significant denoising results for TEM signals contaminated by theoretical or field experimental noise. For practical applications, its denoising effect can be further improved by improving its covariance function. The results of this study provide a novel, practical solution for the TEM signal denoising.