Post-stack Seismic Data Denoising via Dynamic Guided Learning

This is the official implementation of the paper Poststack Seismic Data Denoising via Dynamic Guided Learning The Leading Edge. This work introduces a dynamic guided learning workflow that utilizes a dynamic database to generate both clean and noisy patches during training. This database guides the learning process for a supervised enhancement task, improving generalization by reducing reliance on specific known distributions. As a result, the method eliminates the need for external datasets while enhancing generalization and adaptability.

Proposed method

Figure 1. The dynamic guided learning workflow with two main processes: (I) the dynamic database that generates X poststack seismic data using the generative model (1000 patches) and Y noisy poststack seismic data (4000 patches) using the degradation model containing 12 different types of noise. (II) The supervised enhancement task learns distinctive features of poststack seismic data during training. While process (II) is training, process (I) generates the next batch of images with different types of noise randomly selected, completing one cycle, with (I) in CPU and (II) in GPU to maximize computational efficiency.

---

How to use

To install the project packages, you can use Anaconda.

conda env create -f environment.yml

or python.

pip install -r requirements.txt

Metrics

PSNR (dB) and SSIM metrics across different simulated types of noise for enhancement methods, including the median filter, DIP, S2S-WTV, and Baseline. Bold text indicates the best result, while underlined text highlights the second-best.

Noises	SSIM (Medial filter)	PSNR (Medial filter)	SSIM (DIP)	PSNR (DIP)	SSIM (S2S-WTV)	PSNR (S2S-WTV)	SSIM (Baseline)	PSNR (Baseline)	SSIM (Proposed)	PSNR (Proposed)
Gaussian	0.783	22.947	0.965	31.992	0.952	29.021	0.975	31.658	0.985	33.457
Poisson	0.784	22.899	0.961	31.997	0.952	28.493	0.974	31.291	0.986	34.248
Speckle	0.563	15.787	0.908	27.273	0.781	16.804	0.807	19.370	0.934	27.336
Salt and pepper	0.712	20.833	0.881	28.028	0.731	20.742	0.825	22.551	0.992	35.922
Linear	0.560	18.867	0.947	29.238	0.839	23.620	0.926	26.580	0.994	36.859
Waves	0.770	22.124	0.957	30.986	0.949	27.884	0.976	31.602	0.991	35.319
Stripes	0.633	18.838	0.964	32.717	0.798	20.687	0.948	25.201	0.998	39.361
Correlated g₁	0.782	22.774	0.944	29.538	0.934	27.272	0.965	29.442	0.972	31.108
Correlated g₂	0.784	22.923	0.950	31.654	0.942	28.509	0.978	31.001	0.987	34.262
Blur	0.516	17.400	0.904	30.426	0.550	17.707	0.600	19.085	0.901	27.337
Correlated g₁₂	0.796	23.124	0.954	31.062	0.953	28.775	0.981	31.271	0.984	33.593
g₁₂ blur	0.683	19.403	0.948	30.456	0.770	20.496	0.851	22.129	0.981	31.543

Citation

If you find the Dynamic Guided Learning useful in your research, please consider citing:

@article{doi:10.1190/tle44090692.1,
author = { Javier Torres-Quintero  and  Paul Goyes-Peñafiel  and  Ana Mantilla-Dulcey  and  Luis Rodríguez-López  and  José Sanabria-Gómez  and  Henry Arguello },
title = {Poststack seismic data denoising via dynamic guided learning},
journal = {The Leading Edge},
volume = {44},
number = {9},
pages = {692-704},
year = {2025},
doi = {10.1190/tle44090692.1},
URL = { https://doi.org/10.1190/tle44090692.1}
}