Hybrid DAE-GAN Model with U-Net Architecture for Seismic Signal Denoising Eko Priyatno(12*), Ahmad Kadarisman (12), Santoso Soekirno(2), and Martarizal (2)
1) Directorate of Instrumentation and Calibration
BMKG, Indonesia
*eko.priyatno[at]ui.ac.id
2) Department of Physics Faculty of Mathematics and Natural Sciences, Universitas Indonesia
Abstract
Seismic information is crucial in geophysical research, yet its quality is often affected by various types of disturbances that complicate the analysis of subsurface structures. This study introduces a novel solution using deep learning to reduce noise in three-component seismic data. The proposed architecture is a combination of a Denoising Autoencoder (DAE) and a Generative Adversarial Network (GAN). A U-Net model is used as the Generator to reconstruct a noise-free signal from noise-affected data. On the other hand, a CNN-based Discriminator model serves to distinguish between the reconstructed signals and the original clean signals. The loss function for the Generator is a combination of Mean Squared Error (MSE) to ensure accurate reconstruction and an Adversarial Loss to maintain realistic statistical characteristics. Thus, the resulting signal is not only free from disturbances but also retains the original characteristics of seismic data. This model was trained and tested using data from the STEAD (STanford EArthquake Dataset). The model^s quality was evaluated using quantitative metrics such as Signal-to-Noise Ratio (SNR), RMSE, and PRD on a separate test set. The test results show that this model can significantly increase the SNR and produce a clean signal that is visually and spectrally (using STFT) very similar to the original signal. This method holds great potential for enhancing automation and efficiency in the seismic data pre-processing workflow.
