A Method for Seismic Data Reconstruction Based on Texture Constrained u-net Network

Abstract

The invention discloses a method for reconstructing seismic data based on a texture-constrained U-net network, comprising: S1, using K-means algorithm to divide a training data set to obtain texture labels of the training data; S2, passing the training data set Train the U‑net network with the texture labels as the texture extractor, and obtain the optimized texture extractor parameters; S3. Connect the trained texture extractor to the reconstruction network in series, and use the optimized texture extractor parameters to extract the reconstruction labels and reconstruct the seismic data The texture information is obtained, and the texture loss is obtained; S4, the seismic data reconstruction is performed after the reconstruction network is optimized by the texture loss. Combined with the feature that the seismic data has rich texture information, the invention can improve the reconstruction accuracy and continuity of the seismic event axis under the condition of limited samples by strengthening the learning of the texture.

Description

technical field [0001] The invention belongs to the technical field of using artificial intelligence for seismic data processing, and relates to a seismic data reconstruction method, in particular to a texture-constrained U-net network-based seismic data reconstruction method. Background technique [0002] Seismic exploration is an important method for studying underground geological structures. However, due to surface obstacles, terrain constraints such as mountains and rivers, and bad and waste roads during the acquisition process, the acquired seismic data is usually under-sampled along the spatial direction, which directly affects subsequent migration imaging, inversion and Interpretation and description of geological formations. Therefore, it is of great practical significance to interpolate and reconstruct the missing seismic data to obtain data with high signal-to-noise ratio, high resolution and high fidelity. [0003] At present, seismic data reconstruction method...

AI Technical Summary

Problems solved by technology

The research has achieved better results than the f-x method, but at the same time the paper also pointed out the limitations of this method: data that is not highly similar to the training data, the reconstruction effect is not good

[0005] These research results demonstrate the great potential of DL in the field of seismic data interpolation, but also expose an important problem: the requirement of huge training samples and the difficulty of network generalization

That is to say, if the network test effect is to be good, a large amount of labeled training data is required. If the training data is not large enough, the traditional DL network cannot achieve good generalization performance.

Images