CEEMD-UNet: Feature-Preserving Deep-Denoising Method for High-rate GNSS coseismic displacement

High‑rate Global Navigation Satellite System (GNSS) observations provide direct measurements of surface coseismic displacements and are a key tool for seismic monitoring. However, the signal‑to‑noise ratio (SNR) of high‑rate GNSS coseismic displacement data is often low in the case of moderate‑magnitude or distant earthquakes, and traditional denoising methods struggle to balance noise suppression with signal preservation. To overcome this problem, this study proposes a new feature-preserving deep-denoising method based on complementary ensemble empirical mode decomposition (CEEMD) and UNet network (CEEMD‑UNet). The proposed method first adaptively decomposes the non‑stationary coseismic displacement via CEEMD, effectively overcoming the limitations of conventional denoising techniques. The resulting intrinsic mode functions (IMFs) are then accurately denoised using a UNet network, achieving an effective balance between noise removal and preservation of coseismic signal features. Tests on synthetic data show that after denoising, the average cross‑correlation coefficients in the E, N, and U directions all exceed 0.85, and the average SNR is improved by factors of 14.72, 12.84, and 38.77, respectively. Validation using real GNSS data from the 2018 Mw 7.0 Anchorage, Alaska earthquake indicates that the average root mean square error across stations is reduced by 41.39%, 48.79%, and 67.90% in the three components, while the SNR is increased by factors of 5.05, 2.89, and 0.96. Compared with CEEMD‑WD and UNet‑only denoising methods, the proposed approach demonstrates significant advantages in waveform consistency, amplitude preservation, and cross‑scenario adaptability, offering a reliable technical solution for processing low‑SNR high‑rate GNSS seismic data.