An Efficient Self-supervised Seismic Data Reconstruction Method Based on Self-Consistency Learning

This paper proposes a self-supervised, lightweight deep learning method that leverages self-consistency learning and inter-component correlations to achieve high-quality reconstruction of irregularly acquired seismic data without requiring external training datasets.

The Gist

Here is an explanation of the paper using simple language and creative analogies.

The Big Problem: The Broken Puzzle

Imagine you are trying to solve a massive, 1,000-piece jigsaw puzzle that shows a beautiful landscape. However, someone has gone through and randomly removed 50% of the pieces. The picture is now full of holes, making it impossible to see the mountains, rivers, or trees clearly.

In the world of geophysics, this is exactly what happens during seismic exploration. Scientists send sound waves underground to "see" oil, gas, or rock structures. They place sensors (receivers) along a line to catch these echoes. But the ground isn't a perfect grid; there are rivers, roads, and steep hills. Sometimes, they can't place sensors in certain spots. This leaves the data with "holes," making the final image of the underground blurry and useless.

The Old Ways: Guessing and Heavy Lifting

For years, scientists tried to fix these holes in two main ways:

1. The "Math Wizard" Approach: They used complex math formulas to guess what the missing pieces looked like based on the ones they had. But this was slow, required a human to tweak the settings constantly, and often the guesses were a bit off.
2. The "Big Data" Approach: They used Artificial Intelligence (Deep Learning). But usually, to teach an AI to fix a puzzle, you need to show it thousands of perfect puzzles first. In geology, we rarely have perfect data to use as a teacher. Plus, these AI models are like giant, heavy elephants—they take forever to run and need massive computers.

The New Solution: The "Self-Taught" Detective

This paper introduces a new method called Self-Consistency Learning (SCL). Think of this method as a detective who doesn't need a textbook or a teacher; they just use their own brain to solve the case.

Here is how it works, step-by-step:

1. The "Mirror" Trick (Self-Consistency)

Imagine you have a broken mirror. You look at the reflection of your left eye in the right side of the mirror. You know what your left eye looks like. Now, imagine you can use that knowledge to guess what the missing left side of the mirror would show.

The new method does this with seismic data. It says: "If I take the data I have, fill in the holes, and then pretend those filled-in parts are the 'missing' parts again, my computer should be able to predict the original data I started with."

It creates a loop of logic:

• Step A: The computer looks at the broken data and tries to fill the holes.
• Step B: It takes that new, filled-in version, hides some of it again, and asks, "Can you predict the part I just hid?"
• Step C: If the computer's guess matches what it originally saw, it knows it's doing a good job. If it doesn't match, it learns from the mistake.

This is called Self-Consistency. The data is its own teacher. It doesn't need a library of perfect examples; it just needs to be consistent with itself.

2. The "Tiny Brain" (Lightweight Network)

Most modern AI models are like super-computers in a backpack—they are huge and heavy. This new method uses a Lightweight Network.

Think of it like a Swiss Army Knife versus a Full Kitchen.

• Old methods try to bring a whole kitchen (millions of parameters) to fix a sandwich.
• This new method uses a tiny, efficient Swiss Army Knife with only 188,849 parts (parameters).

Because it is so small and efficient, it can run on a standard computer and process massive amounts of data without getting tired. It doesn't need to chop the big puzzle into tiny, manageable pieces (which often causes jagged edges); it can look at the whole picture at once.

Why This Matters

The authors tested this on two huge, real-world datasets from Alaska (one of the most difficult places to do geology because of the terrain).

• The Result: The new method filled in the holes better than the old math methods and better than the heavy AI models.
• The Speed: It was incredibly fast. While old methods took hours to fix a single line of data, this method did it in minutes.
• The Quality: The reconstructed images were sharper, with fewer "artifacts" (weird vertical lines or blurs that look like static on an old TV).

The Takeaway

This paper presents a clever, efficient way to fix broken seismic data. Instead of relying on massive libraries of perfect data or slow, heavy computers, it uses a smart, self-teaching loop and a tiny, efficient AI to reconstruct the underground world.

It's like giving a detective a mirror and a magnifying glass, allowing them to solve the mystery of the missing pieces using only the clues already present in the room. This makes exploring for oil, gas, and understanding our planet's crust much faster, cheaper, and more accurate.

Technical Summary

Here is a detailed technical summary of the paper "An Efficient Self-supervised Seismic Data Reconstruction Method Based on Self-Consistency Learning."

1. Problem Statement

Seismic exploration is critical for understanding subsurface structures, but complex surface conditions (terrain undulations, obstacles, restricted access) often prevent the uniform deployment of receivers. This results in irregularly sampled seismic data, characterized by missing traces.

• Consequences: Irregular sampling disrupts the sparsity and low-rank structure of seismic data, severely hindering subsequent processing, inversion, and geological interpretation.
• Limitations of Existing Methods:
  • Traditional Methods (Sparse Transformations, Rank-Reduction): Rely on manually tuned parameters and iterative filtering, leading to low computational efficiency and difficulty in handling large-scale data.
  • Supervised Deep Learning: Requires large external datasets for training, which are often unavailable or lack interpretability.
  • Existing Self-Supervised Methods: Often lack effective constraints on the reconstructed data, leading to unstable performance and artifacts, especially in large-scale scenarios.

2. Methodology

The authors propose a Self-Supervised Self-Consistency Learning (SCL) strategy combined with a Lightweight Convolutional Autoencoder (CAE).

A. Self-Consistency Learning Strategy

Instead of relying on external ground truth, the method leverages the internal correlations within the seismic data itself. The core hypothesis is that a well-trained network should be able to reconstruct missing data from observed data, and conversely, the reconstructed missing data should be consistent with the observed data when re-sampled.

The objective function is derived through a bidirectional prediction logic:

1. Standard Reconstruction: $N(d) \to \text{Reconstructed } m$.
2. Self-Consistency Constraint: The network is trained to ensure that if the reconstructed data is masked again (using a random sampling mask $R'$), the network can reconstruct the original observed data $d$ from it.
3. Loss Function (Equation 9): The total loss consists of three terms:
   • Term 1: Reconstruct observed data from the full input (Standard MSE).
   • Term 2: Reconstruct observed data from the "re-masked" reconstruction (ensuring the reconstruction contains the necessary information to recover the original).
   • Term 3: Consistency Loss: Minimizes the difference between the initial reconstruction and the secondary reconstruction derived from the masked version. This enforces that the network learns the intrinsic data manifold rather than just memorizing the input.

$$\theta = \arg \min_{\theta} \left( \|d - N(d) * R\|^2 + \|d - N(N(d) * R') * R\|^2 + \|N(d) - N(N(d) * R')\|^2 \right)$$

B. Lightweight Network Architecture

• Structure: A Convolutional Autoencoder (CAE) with an Encoder-Decoder structure.
• Efficiency: The network is extremely lightweight, containing only 188,849 learnable parameters.
• Advantage: This low parameter count allows the model to process large-scale seismic data directly without the need to divide data into small patches (a common requirement for larger networks), thereby avoiding patch-boundary artifacts.

3. Key Contributions

1. Self-Supervised Paradigm: The method requires no external datasets for training, making it applicable to any seismic survey where ground truth is unavailable.
2. Self-Consistency Constraint: Introduces a novel loss function that enforces bidirectional consistency between observed and reconstructed data, significantly stabilizing the reconstruction process compared to standard self-supervised approaches.
3. Lightweight Design: The CAE architecture is optimized for efficiency, enabling the processing of massive datasets (hundreds of kilometers of survey lines) in a single pass without patching.
4. Robustness: The method demonstrates strong resistance to noise interference and maintains high fidelity in complex geological settings.

4. Experimental Results

The method was validated on two large-scale public datasets:

• Dataset 1: USGS Beaufort Sea–Arctic Alaska (9 survey lines, up to 206 km long).
• Dataset 2: USGS National Petroleum Reserve–Alaska (NPRA) (7 survey lines, up to 183 km long, total ~815 km).

Comparative Baselines:

• Traditional Self-Supervised Learning (using only MSE loss).
• SCRN (Swin Transformer-based Convolutional Residual Network).
• Traditional methods: Damped Rank-Reduction and Fast Dictionary Learning.

Performance Metrics:

• Quantitative: SCL achieved the highest SNR, SSIM, and R² scores across all survey lines. For example, on the NPRA dataset, SCL averaged an SNR of 12.60, compared to 11.20 for the traditional self-supervised method and 12.01 for SCRN.
• Qualitative:
  • SCL produced results with higher lateral continuity and fewer interpolation artifacts (e.g., vertical streaks) compared to SCRN and traditional methods.
  • In complex areas with strong noise and large trace gaps, SCL successfully reconstructed reflection events where other methods failed or produced discontinuous signals.
• Efficiency:
  • SCL processed a line with ~800 million sampling points in <580 seconds on an NVIDIA RTX 4060.
  • Traditional methods required several hours for the same task.
  • SCL showed linear time growth with data size, whereas traditional methods struggled with scalability.

5. Significance

This paper presents a significant advancement in seismic data processing by addressing the "data scarcity" and "computational efficiency" bottlenecks simultaneously.

• Scalability: It enables the reconstruction of massive, irregularly sampled datasets that were previously too computationally expensive or data-intensive to process effectively.
• Reliability: By removing the dependency on external training data, the method avoids the "black box" risks associated with supervised deep learning, providing a more trustworthy solution for critical geological exploration.
• Practical Application: The high efficiency and stability make it highly suitable for large-scale oil, gas, and mineral exploration projects where terrain limitations frequently cause data gaps.