[Paper Review] Deep Autoassociative Neural Networks for Noise Reduction in Seismic data

Debjani Bhowmick, Deepak Gupta+2 more|arXiv (Cornell University)|May 1, 2018

Seismic Imaging and Inversion Techniques1 references4 citations

TL;DR

This paper proposes deep autoassociative neural networks (autoNNs) for noise reduction in seismic data, leveraging stacked autoencoders to learn nonlinear, sparse representations that reconstruct clean signals from noisy inputs. The method achieves over 90% noise reduction in synthetic tests and significantly suppresses time- and frequency-localized noise in real seismic data, though with some trade-off in spatial resolution.

ABSTRACT

Machine learning is currently a trending topic in various science and engineering disciplines, and the field of geophysics is no exception. With the advent of powerful computers, it is now possible to train the machine to learn complex patterns in the data, which may not be easily realized using the traditional methods. Among the various machine learning methods, the artificial neural networks (ANNs) have received enormous attention. A variant of ANNs, autoassociative neural network (autoNN) tries to learn the reconstruction of input itself using backpropagation. In an autoNN, the input and output are the same, and an approximation to the identity mapping is obtained in a nonlinear setting. AutoNNs have primarily been used to extract sparse internal representations of any input and reduce its dimensionality. In this paper, we explore the potential of autoNNs in reducing random noise in geophysical data. In this paper, the first results of this study are presented. The synthetic mathematical example demonstrates the concept of autoNN. For the test seismic data, it is observed that autoNN can significantly remove the vertical time- and frequency-local noise, however, the resolution of the output signal is compromised to a certain extent. Future work includes testing larger examples with several different types of noise, and using deep-stacked-autoNNs to further reduce the noise, ensuring minimal compromise with the resolution of the signal.

Motivation & Objective

• To develop a deep learning approach for denoising geophysical data, particularly seismic signals corrupted by non-coherent and localized noise.
• To extend traditional autoassociative networks into deep architectures for improved noise suppression beyond shallow networks.
• To evaluate the performance of deep autoNNs on both synthetic mathematical signals and real seismic data.
• To investigate the trade-off between noise reduction and signal resolution in seismic data restoration.

Proposed method

• A deep autoencoder architecture with three hidden layers (300-400-300 neurons) is used to learn a nonlinear mapping from noisy input to clean output.
• The encoder compresses the input into a lower-dimensional latent representation using nonlinear activation functions (e.g., sigmoid), while the decoder reconstructs the original signal.
• The network is trained using backpropagation to minimize the mean squared error between the original and reconstructed signals.
• A stacked autoencoder approach is employed, pre-training each layer greedily before fine-tuning the full network.
• For training, 1.1 million synthetic seismic patches (9×42 samples) are generated, including both clean and noisy versions with monofrequency sinusoids.
• The model is trained for 50,000 epochs with a 95:5 training-validation split, using gradient descent optimization.

Experimental results

Research questions

• RQ1 Can deep autoassociative neural networks effectively reduce non-coherent and time/frequency-localized noise in seismic data?
• RQ2 How does the performance of deep autoNNs compare to shallow autoencoders in noise reduction tasks?
• RQ3 To what extent does the autoencoder preserve signal resolution while removing noise?
• RQ4 Can the autoencoder generalize to unseen noisy seismic traces not present in the training set?

Key findings

• The deep autoNN achieved a mean noise reduction efficiency of 90.27% on 100 test samples of synthetic signals, demonstrating strong denoising capability.
• In seismic data, the autoNN successfully suppressed vertical, time- and frequency-localized noise, significantly improving data interpretability.
• The model exhibited regularization effects, leading to a slight reduction in signal resolution, particularly in high-amplitude events.
• The network performed best on high-frequency noise (110–220 Hz), but struggled with low-frequency and multifrequency noise due to limited training coverage.
• The stacked autoencoder architecture outperformed shallow networks, indicating the benefit of deeper representations in capturing complex noise patterns.
• The method shows promise for future adaptation to diverse noise types, especially with deeper and more specialized architectures.