Sparse time-frequency representation based on Unet with domain adaptation

Short-time Fourier transform (STFT) is a widely used time-frequency representation (TFR) for seismic processing and interpretation, which is limited by the Heisenberg uncertainty principle, resulting in TF spectrum with low TF resolution. To address this issue, the sparse STFT is proposed, which is formulated and solved as a typical inverse problem. Nevertheless, sparse STFT suffers from the low computational efficiency and the parameter selection. In this paper, we propose the sparse time-frequency representation (STFR) based on Unet with domain adaptation (STFR-UDA) model for solving these issues. First, we utilize the Marmousi model to generate synthetic training data set and corresponding STFR labels by using sparse STFT with fine-tuning parameters. Then, based on synthetic data set, we propose and pre-train the STFR based on Unet (STFR-Unet) model to build the relationship between seismic signal and STFR labels. After validating the well-trained STFR-Unet model, we introduce a domain adaptation scheme to enhance the generalization property of STFR-Unet. In this step, we introduce field data without STFR labels to train the STFR-UDA model. Afterward, we apply it to 3D field data for characterizing fluvial channels and make detailed comparisons with the traditional TFR methods.