Automatic Reconstruction of Velocity and Displacement from Seismic Acceleration Measurements by a Cycle-Consistent Generative Adversarial Network

The reconstruction of velocity and displacement from acceleration measurements is essential for seismic or microseismic data interpretation. Many existing approaches produce excellent results in both velocity and displacement reconstruction. However, these reconstruction approaches involve a large number of parameters, which must be tuned for different datasets. There is room for improvement in high-precision and immediate reconstruction, particularly for large and noisy datasets. Therefore, we propose a deep learning approach that is parameter-free during inference to automatically reconstruct velocity and displacement from seismic acceleration measurements via a cycle-consistent generative adversarial network (CycleGAN). To obtain paired training data, we use a semi-synthetic method to prepare the dataset. Based on the dataset, we use CycleGAN to train two mapping functions for reconstructing the velocity and displacement from acceleration measurements. We then compare this method with two traditional methods on the test set. For the reconstruction of seismic velocity measurements, the mean average peak error (Erp), average deviation error (Err), and root-mean-square error (Ers) on the test set for the method proposed in this paper are 0.1478, 1.2269, and 0.0097, respectively. Moreover, for the reconstruction of seismic displacement measurements, the mean Erp, Err, and Ers values on the test set for the method proposed in this paper are 0.2004, 0.9783, and 5.6 × 10⁻⁴, respectively. Additionally, under noisy input conditions, the records reconstructed by our approach exhibit the smallest degradation in signal-to-noise ratio (SNR) and preserve more records in the higher SNR range than those reconstructed by traditional methods. The results show that the proposed method achieves comparable accuracy in velocity reconstruction and better accuracy in displacement reconstruction than traditional methods. More importantly, it can automatically reconstruct records and remain robust to input noise without manual parameter tuning, demonstrating good prospects for practical application.