💻 computer science

Large Vision Model-Guided Masked Low-Rank Approximation for Ground-Roll Attenuation

This paper proposes a Large Vision Model-guided Masked Low-Rank Approximation (LVM-LRA) framework that leverages promptable vision models to generate precise masks for targeted ground-roll attenuation, effectively separating noise from reflections while minimizing signal leakage through a novel combination of global and mask-guided local low-rank constraints.

Original authors: Jiacheng Liao, Feng Qian, Ziyin Fan, Yongjian Guo

Published 2026-04-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiacheng Liao, Feng Qian, Ziyin Fan, Yongjian Guo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a beautiful, quiet conversation (the useful seismic reflections) happening in a crowded, noisy room. The problem is that a massive, loud group of people is shouting and stomping right next to you (the ground roll). This noise is so loud and overlaps so much with the conversation that it's nearly impossible to hear the words.

In the world of oil exploration and earthquake study, scientists use "seismic records" to see underground. But just like your noisy room, these records are often ruined by "ground roll"—a type of surface noise that drowns out the important signals.

This paper introduces a new, clever way to clean up this noise without accidentally deleting the conversation you want to hear. Here is how it works, broken down into simple steps:

1. The Old Way: The "Blind Filter" vs. The "Rough Sketch"

Previously, scientists tried to fix this in two main ways, both of which had flaws:

The "Blind Filter" (Global Methods): Imagine trying to quiet the whole room by turning down the volume on everyone, everywhere. This stops the shouting, but it also makes the quiet conversation too soft to hear. You lose the signal.
The "Rough Sketch" (Local Methods): This approach tries to only quiet the people shouting. But to do this, you need a map of exactly where the shouters are. Old methods used simple, hand-drawn maps (manual rules) to guess the location. These maps were often inaccurate, leaving some noise behind or accidentally silencing parts of the conversation.

2. The New Solution: The "Smart Detective" (LVM-LRA)

The authors propose a two-step system called LVM-LRA. Think of it as hiring a highly intelligent detective who doesn't need to be taught from scratch.

Step 1: The Detective Draws the Map (The "Large Vision Model")

Instead of using a rough sketch, they use a Large Vision Model (LVM). Think of this as a super-smart AI that has seen millions of pictures and knows what things look like.

How it works: You don't need to train this AI on seismic data. Instead, you just give it a few clues (prompts), like: "Find the fan-shaped, low-frequency, rumbling noise" and show it one example picture of what that noise looks like.
The Result: The AI instantly draws a perfect, high-definition "mask" (a digital stencil) that highlights exactly where the ground roll is. It's like the detective pointing a laser pointer at the exact spot of the noise, ignoring the rest of the room.

Step 2: The Surgical Removal (The "Mask-Guided Low-Rank Approximation")

Now that we have the perfect map, we don't just delete the noise. We perform a delicate surgery.

The Concept: The scientists treat the seismic data like a complex puzzle. They know that the "good" signals (the underground reflections) have a certain smooth, continuous structure (like a calm river), while the "bad" noise (ground roll) has a different, messy structure (like choppy waves).
The Magic: Using the AI's map, they tell the computer: "Only look at the messy waves inside the laser pointer's circle and try to smooth them out. Leave the calm river outside the circle alone."
The Math: They use a mathematical trick called Low-Rank Approximation. Imagine the noise is a crumpled piece of paper. The computer tries to flatten it out (approximate it) only where the AI said the noise is. This separates the noise from the signal without damaging the signal.

3. Why This is a Big Deal

No Training Required: Usually, AI needs thousands of examples of "clean" vs. "noisy" data to learn. This method is "training-free." It uses the AI's existing knowledge of shapes and physics, making it ready to use immediately in new locations.
Precision: Because the AI draws the map so accurately, it doesn't accidentally delete the good signals (signal leakage). It's like using a scalpel instead of a sledgehammer.
Versatility: It works on fake data (simulations) and real-world data from different oil fields, proving it's robust.

The Bottom Line

This paper is about using a smart, pre-trained AI detective to draw a perfect map of where the noise is, and then using mathematical surgery to remove only that noise. The result is a crystal-clear picture of the underground, revealing secrets that were previously hidden by the "shouting" ground roll.

1. Problem Statement

Ground roll is a dominant form of coherent noise in land seismic acquisition, characterized by low frequency, low apparent velocity, high amplitude, and strong dispersion. Its attenuation is challenging because:

Signal Overlap: Ground roll energy often overlaps substantially with useful reflection signals in both the time-space and frequency-wavenumber domains.
Limitations of Existing Methods:
- Global Methods: (e.g., f-k filtering, SVD, global deep learning) apply uniform suppression across the entire shot gather. This inevitably damages useful signals in reflection-dominated regions, causing signal leakage or distortion.
- Local Methods: (e.g., local SVD, adaptive filtering) restrict attenuation to specific regions but rely on manually designed or simplistic model-based strategies to estimate "masks" (identifying where ground roll exists). These strategies often fail to provide accurate masks, leading to incomplete noise removal or residual signal damage.

The core problem is to achieve effective noise suppression while preserving weak reflection structures, requiring a method that can accurately localize ground-roll-dominant regions without manual annotation.

2. Methodology: The LVM-LRA Framework

The authors propose a two-stage framework called LVM-LRA (Large Vision Model-guided Masked Low-Rank Approximation). This is a training-free approach that does not require paired clean/noisy data.

Stage I: LVM-Prompted Ground-Roll Recognition

Instead of using simple statistical rules to find noise, the method employs a Large Vision Model (LVM), specifically CLIPSeg, to generate a binary mask ( $M$ ) identifying ground-roll-dominant regions.

Multimodal Prompting: The LVM is guided by a combination of prompts to perform zero-shot segmentation:
1. Morphological Prompts: Text describing visual features (e.g., "fan-shaped," "coherent," "low-slope").
2. Physical Property Prompts: Text describing physical attributes (e.g., "Rayleigh-type," "low velocity," "dispersive").
3. Visual Exemplar: A single patch of ground roll provided as a few-shot visual cue to reduce ambiguity.
Output: A dense response map is binarized using a threshold to produce a precise spatial mask ( $M$ ) where $M_{ij}=1$ indicates ground roll and $0$ indicates clean data.

Stage II: Mask-Guided Low-Rank Approximation (LRA)

Once the mask is generated, the seismic gather ( $Y$ ) is decomposed into a useful reflection component ( $X$ ) and a ground-roll component ( $G$ ) using an optimization model.

Mathematical Formulation: The problem is formulated as minimizing the data fidelity term plus two regularization terms:
$\min_{X,G,Z} \frac{1}{2}\|Y - X - G\|_F^2 + \lambda_s \|F(X)\|_* + \lambda_g \|F(Z)\|_*$
Subject to $Z = M \circ G$ (where $\circ$ is the Hadamard product and $F$ is the Fourier Transform).
Dual Low-Rank Constraints:
1. Global Low-Rank on Reflections ( $X$ ): Imposed in the frequency domain to preserve event continuity and global structure.
2. Local Low-Rank on Ground Roll ( $Z$ ): Imposed only within the masked regions ( $Z = M \circ G$ ). This ensures noise separation is confined to the identified noise zones, preventing signal leakage in clean areas.
Optimization: The problem is solved using the Alternating Direction Method of Multipliers (ADMM). The algorithm iteratively updates variables using Singular Value Thresholding (SVT) and Fast Fourier Transforms (FFT).

3. Key Contributions

LVM for Seismic Masking: This is the first study to introduce a Large Vision Model (CLIPSeg) for ground-roll attenuation. It replaces error-prone manual mask estimation with a promptable, zero-shot segmentation mechanism that leverages semantic and visual priors.
Mask-Guided Low-Rank Decomposition: The authors propose a novel model that combines a global low-rank constraint for reflections with a mask-constrained local low-rank constraint for ground roll. This dual-regularization strategy effectively separates overlapping signals while minimizing signal leakage.
Training-Free and Interpretable: The method requires no labeled training data. The LRA component is derived from physical models (low-rank structure in frequency domain), making the parameters physically meaningful and the process interpretable.
Scalability: The prompt-based design allows the model to be flexibly adapted to different survey areas without retraining.

4. Experimental Results

The method was validated on synthetic data and two distinct real-world field datasets, comparing against three baselines: Local SVD (LSVD), f-k Filtering, and Implicit Neural Representations (INR).

Quantitative Performance:
- Synthetic Data: LVM-LRA achieved the highest Signal-to-Noise Ratio (SNR) of 14.50 dB (vs. 7.54 dB for f-k) and the lowest local similarity (0.0285), indicating superior noise removal with minimal signal leakage.
- Real Data: Consistently achieved the lowest local similarity values across both datasets, confirming better separation of noise from reflections compared to baselines.
Qualitative Performance:
- Visual Quality: The method preserved weak reflection events and structural continuity better than baselines, which often showed distortion or residual noise.
- Spectral Analysis: In the frequency-wavenumber (f-k) domain, LVM-LRA results showed a more compact useful-signal band with significantly reduced dispersed ground-roll energy compared to other methods.
- Mask Accuracy: The LVM-generated masks were more compact and concentrated on the actual noise cone, whereas LSVD masks were often fragmented or included non-noise regions.

5. Significance and Impact

Paradigm Shift: The paper bridges the gap between computer vision (Large Vision Models) and geophysical signal processing. It demonstrates that pre-trained vision models can effectively understand seismic physics through natural language prompts.
Solving the Overlap Problem: By accurately localizing noise regions before separation, the method overcomes the fundamental limitation of global filtering methods that damage useful signals.
Practical Applicability: The "training-free" nature makes the method highly attractive for industry applications where acquiring large amounts of labeled clean/noisy seismic pairs is difficult or impossible.
Future Direction: The authors acknowledge that if ground roll is extremely diffuse and indistinguishable from reflections, the initial mask might be imperfect. They suggest future work could involve a feedback loop where the separation result refines the mask iteratively.

In conclusion, the LVM-LRA framework offers a robust, accurate, and physically interpretable solution for ground-roll attenuation, outperforming traditional and deep learning-based baselines by leveraging the semantic understanding capabilities of Large Vision Models.