Signal Recovery from Randomly Quantized Data Using Neural Network Approach

Area of Science:

• Geophysics
• Machine Learning
• Signal Processing

Background:

• Seismic data often suffers from under-quantization, impacting data quality and analysis.
• Accurate seismic deconvolution is crucial for extracting valuable subsurface information.
• Existing methods struggle with robustness and computational efficiency in multichannel seismic data processing.

Purpose of the Study:

• To develop an efficient and robust seismic deconvolution scheme using a long short-term memory (LSTM) autoencoder.
• To improve the compression of massive seismic datasets and enhance data robustness to quantization errors.
• To outperform existing seismic deconvolution methods in terms of accuracy, robustness, and computational complexity.

Main Methods:

• Implementation of a long short-term memory (LSTM) autoencoder for multichannel seismic deconvolution.
• Development of a robust estimation technique to recover sparse reflectivity from under-quantized seismic data.
• Adjustment of quantization error to improve data robustness and visual saliency.

Main Results:

• The proposed LSTM autoencoder scheme significantly suppresses erroneously estimated impulses.
• The method demonstrates superior robustness to changes in quantization intervals.
• Validated on field and synthetic datasets, the LSTM autoencoder outperforms steepest decent and basis pursuit methods in quality and computational complexity.

Conclusions:

• The LSTM autoencoder-based seismic deconvolution offers a significant advancement in processing under-quantized multichannel seismic data.
• The proposed method provides a robust and computationally efficient solution for seismic data enhancement.
• This approach boosts visual saliency and improves the overall quality of seismic data analysis.