Deep learning for irregularly and regularly missing data reconstruction.

Area of Science:

• Geophysics
• Data Science
• Machine Learning

Background:

• Conventional interpolation methods often rely on assumptions that limit their effectiveness.
• Deep learning (DL) offers a powerful alternative for feature mining and data reconstruction without such constraints.

Purpose of the Study:

• To apply DL for reconstructing irregularly and regularly missing seismic data.
• To develop a DL model capable of transforming incomplete data into complete data.

Main Methods:

• An encoder-decoder U-Net convolutional neural network architecture was employed.
• The model was trained using synthetic and field seismic data with a mean-squared-error loss function and Adam optimizer.
• Feature maps were visualized to illustrate the reconstruction process.

Main Results:

• The DL method demonstrated superior performance compared to a Fourier transform interpolation method for irregularly missing data.
• The model successfully reconstructed regularly missing data, even though it was trained exclusively on irregularly sampled data.
• Effectiveness, superiority, and generalization capabilities were verified.

Conclusions:

• Deep learning provides a robust and effective solution for seismic data reconstruction.
• The developed U-Net based model shows significant potential for handling various types of missing data in geophysics.