Abstract: Objectives: High-rate global navigation satellite system (GNSS) technology enables the direct acquisition of surface coseismic displacements and serves as a vital tool in the field of earthquake monitoring. However, in scenarios involving moderate-to-small earthquakes or distant epicentral distances, the signal-tonoise ratio (SNR) of recorded data is typically low, which severely restricts the accuracy of subsequent studies such as source parameter inversion. Methods: To address this, a novel denoising method fusing complementary ensemble empirical mode decomposition (CEEMD) with a UNet network (CEEMD-UNet) is proposed. This method first adaptively decomposes non-stationary coseismic displacements through CEEMD, effectively overcoming the limitations of traditional denoising methods; then, the UNet network is utilized to precisely denoise the decomposed intrinsic mode functions (IMFs), achieving a balance between efficient noise suppression and the preservation of coseismic signal characteristics. Results: Tests on synthetic data demonstrate that the average cross-correlation coefficient (CC) between the denoised signals and the true signals in the east (E), north (N), and up (U) components exceeds 0.85, with the average SNR improved by factors of 14.72, 12.84, and 38.77, respectively. Further validation based on actual GNSS observation data from the 2018 Anchorage, Alaska, M_W7.0 earthquake shows that the average standard deviation (STD) in the E, N, and U components across all stations after denoising were reduced by 41.39%, 48.79%, and 67.90% respectively, and the SNR were improved by factors of 5.05, 2.89, and 0.96, respectively. Most importantly, compared with the CEEMD-Wavelet denoising (CEEMD-WD) method and UNet denoising method, the CEEMD-UNet method exhibits significant advantages in terms of waveform consistency, amplitude preservation characteristics, and cross-scenario robustness. It can provide technical support for the processing of low SNR high-rate GNSS seismic data. Conclusions: The effectiveness of hybrid signal processing strategies in geophysical data cleaning should be confirmed. The proposed CEEMD-UNet framework provides a reliable solution for improving the quality of high-rate GNSS data and holds significant value for high-precision applications such as earthquake early warning and source mechanism inversion.