HADDL-DAS-Denoising

Official MATLAB implementation for the Hybrid Attention-Driven Deep Learning (HADDL) framework for Distributed Acoustic Sensing (DAS) data noise attenuation.

This repository provides a fully reproducible and comprehensive workflow that executes both the synthetic wavefield simulations and the diverse real-world field data applications presented throughout the manuscript. It allows researchers to seamlessly evaluate framework scalability from controlled synthetic tests to high-volume field records.

Reference

If you use this code, subroutines, or benchmark datasets in your research, please cite:

• Oboué, Y. A. S. I., Chen, Y., & Chen, Y. (2026). A Hybrid Attention-Driven Deep Learning Framework for Denoising DAS Data. Geophysics (Under Review).

---

Repository Architecture

The package is organized into four primary directories to ensure modularity and ease of reproduction:

• Demos/: Contains the core high-level execution scripts for computing and plotting.
• Field_data/: Storage for real-world DAS data matrices (.mat format).
• Synth_data/: Storage for generated synthetic test profiles.
• Subroutines/: The complete backend engine containing filtering operators, metrics, and network helpers.

---

Detailed File Descriptions

1. Main Plotting and Figure Generation Scripts (Demos/)

These scripts act as the wrapper environment to visualize the comparative performance across different benchmarking techniques:

• plot_haddl_synth_fig_5_8.m: Synthesizes and plots the multi-method benchmark (Figures 5, 6, 7, and 8) for synthetic datasets.
• plot_haddl_m..._fig_14_17.m: Generates benchmarking plots and comparative profiles for the microDAS dataset (Figures 14 to 17).
• plot_haddl_F...AS_fig_9_13.m: Generates benchmarking plots and processing profiles for the FORGE_DAS dataset (Figures 9 to 13).
• plot_haddl_jD..._fig_18_22.m: Generates benchmarking plots, removed noise profiles, and local similarity maps for the jDAS submarine dataset (Figures 18 to 22).

2. Framework Execution and Test Pipelines (Demos/)

Run these workflows to process the data matrices prior to plotting:

• Proposed Framework:

  • test_haddl_synth.m / test_haddl_synth_fig1.m / test_haddl_synth_fig_3_4.m: Runs the HADDL denoising pipeline on different synthetic data configurations.
  • test_haddl_microDAS.m: Applies HADDL to the high-frequency microDAS field records.
  • test_haddl_FORGE_DAS.m: Processes the deep geothermal reservoir DAS data from the FORGE site.
  • test_haddl_field_jDAS.m: Executes HADDL on the submarine jDAS dataset.

• MHA-RN Baseline Evaluations:

  • test_mharn_synth.m, test_mharn_microDAS.m, test_mharn_FORGE_DAS.m, test_mharn_field_jDAS.m: Run the Multi-Head Attention Residual Network processing baseline across all corresponding synthetic and field scenarios.

• BP+SGK Baseline Evaluations:

  • test_sgk_synth.m, test_sgk_microDAS.m, test_sgk_field_jDAS.m, test_sgk_field_FORGE_DAS.m: Execute the traditional Bandpass + Sliding Singular Value Decomposition (SGK) baseline filters.

💡 Note on SSDL Benchmark Reproducibility: The SSDL (Self-Supervised Deep Learning) comparative data fields (stored in Output_Synth_SSDL/ and Output_Field_*_SSDL/) were generated using the official Python/Jupyter Notebook pipeline provided by Saad et al. (2024). The denoised outputs were exported as .mat matrices for seamless integration and calculation within this MATLAB evaluation suite.

3. Subroutines and Backend Functions (Subroutines/)

Core operators that drive the signal enhancement framework:

• haddl_DL_Predict.m: Handles deep learning inference and model execution.
• haddl_localsimi.m: Computes the 2D local similarity maps used to track noise orthogonality.
• haddl_snr.m: High-accuracy Signal-to-Noise Ratio calculation utility.
• haddl_fk_dip.m / haddl_bandpass.m: Classical f-k domain and frequency bandpass operators.
• LO_adjnull.m / LO_banded_solve.m: Mathematical solvers for localized orthogonalization routines.
• haddl_patch.m / haddl_patch_inv.m (2D & 3D): Implements optimal spatial windowing and overlapping patch reconstruction to mitigate boundary artifacts.

---

Data Dependencies

Field Data (Field_data/)

• microDAS_data.mat (Compressed) / forgedAS_data.mat / jDAS_data.mat

⚠️ CRITICAL NOTE ON microDAS_data.mat: Due to GitHub's file size limitations, the microDAS_data.mat field dataset has been compressed into a standard zip archive (e.g., microDAS_data.zip). Users must extract/unzip this file inside the Field_data/ directory before executing any associated field processing or plotting scripts in MATLAB.

• microDAS_data.mat: High-density ambient and microseismic active records.
• forgedAS_data.mat: Wellbore monitoring dataset from the Utah FORGE geothermal site.
• jDAS_data.mat: Submarine dark fiber acoustic sensing profile.

Synthetic Data (Synth_data/)

• dsynthDAS3.mat / dsynthDAS4.mat: Clean synthetic forward-modeled seismic wavefields.
• dnhoriz.mat / dnoiseSynthDAS.mat: Pre-computed complex random and coherent noise vectors.

---

License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.