Enhanced Seismic Data Denoising: Integrating Polynomial Interpolation and Localized Time-Frequency Information With Deep Neural Network Architectures

The recorded seismic signals are often marred by unwanted noise and artifacts, detracting from the quality and clarity required for accurate analysis. The critical task of removing this undesirable noise is essential for the effective processing of seismic data. In this work, the focus is primarily on the challenge of denoising by using a convolutional neural network operating on a transformation. By employing the polynomial fitting and localized time-frequency analysis, the low-dimensional seismic signals are processed into high-dimensional signals. The combination of time-frequency wavelet domain representation and polynomial interpolation with deep learning techniques results in a notable increase in the signal-to-noise ratio (SNR), which indicates the efficiency of the proposed method. By comparing our method against the traditional transformed and nontransformed-based approaches, including the widely used DnCNN denoiser, we observed that our approach consistently achieves higher SNR values, with the transformation step playing a key role in this improvement. Moreover, the methodology showed reasonably consistent performance across a range of mother wavelets, indicating limited sensitivity to the particular wavelet selected. The practical applications of this enhanced denoising method are vast, with its impact extending to the monitoring of volcanic activity, where data clarity can be lifesaving, and to earthquake detection, where early and accurate recognition is essential.