Sparse Time-Frequency Representation Based on Self-Supervised Learning

Abstract

Summary The sparsity-based time-frequency (TF) transforms are widely used to obtain a high localized TF representation in seismic signal processing. These transforms formulate the sparse TF representation as an inverse optimization problem based on hand-crafted prior knowledge. Unlike the traditional sparsity-based TF transforms, the deep learning (DL)-based sparse TF representations don’t need this prior knowledge and make use of a large amount of labeled data. However, the labeled data set is difficult to acquire for field data. To bridge the gap between the traditional sparsity-based transforms and the DL-based transforms, we propose a sparse TF representation based on the short Fourier transform (STFT) and self-supervised learning (STFR-SSL) in the absence of the “ground-truth” TF representation in this study. After training and validating the model, a 3D field data example is used to demonstrate the effectiveness of the proposed STFR-SSL model.