When Do Simple Methods Suffice? Fair Baselines for Denoising, Trace Quality Control, and Acoustic-Texture Clustering of Chirp Sub-Bottom Profiler Data
IEEE Access · 2026 · 14 · 169474-169489
DOI: 10.1109/access.2026.3748171 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ganghoon Lee und 2 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Basic Research Project through Korea Institute of Geoscience and Mineral Resources, Ministry of Science and ICT, Republic of Korea, Korea Institute of Energy Technology Evaluation and Planning (KETEP) and the Ministry of Climate, Energy and Environment (MCEE), Republic of Korea.
Abstract (Originalsprache)
Training-data-free methods and frozen pretrained models are increasingly used in geophysical imaging, often without comparison with well-tuned classical baselines. This study tests whether these newer approaches provide a consistent advantage for processed chirp sub-bottom profiler (SBP) envelope sections. We compare classical and newer methods in denoising, trace-level quality control (QC), and label-free acoustic-texture clustering using task-specific controlled tests. Settings and evaluation procedures are defined before testing, and the survey campaign is treated as the unit of independence. The evaluation uses chirp SBP sections from Korean waters and a public DUNEX dataset from the U.S. Geological Survey. For denoising, classical filters remain competitive with training-data-free neural methods. Block-matching and 3D filtering (BM3D) gives the highest output signal-to-noise ratio in the main Gaussian-noise test and also leads under white and burst noise, whereas total variation (TV) or Gaussian smoothing performs better under band-limited and striping noise. For QC, performance depends on how well each test matches the trace problem, and no single test or combination performs best in all cases. For texture clustering, the classical gray-level co-occurrence matrix (GLCM) representation remains competitive with frozen pretrained features across the evaluated data. Together, these comparisons indicate that classical baselines remain competitive under the tested conditions, while the relative benefit of newer approaches varies by task and data conditions. The findings are limited to precomputed SBP envelope sections and do not extend directly to raw signed traces or pre-stack data.
Förderung
- Basic Research Project through Korea Institute of Geoscience and Mineral Resources
- Ministry of Science and ICT, Republic of Korea
- Korea Institute of Energy Technology Evaluation and Planning (KETEP) and the Ministry of Climate, Energy and Environment (MCEE), Republic of Korea
Zitieren
APA
Lee, G., Kim, Y., & Kwon, J. (2026). When Do Simple Methods Suffice? Fair Baselines for Denoising, Trace Quality Control, and Acoustic-Texture Clustering of Chirp Sub-Bottom Profiler Data. IEEE Access, 14, 169474-169489. https://doi.org/10.1109/access.2026.3748171
Vancouver
Lee G, Kim Y, Kwon J. When Do Simple Methods Suffice? Fair Baselines for Denoising, Trace Quality Control, and Acoustic-Texture Clustering of Chirp Sub-Bottom Profiler Data. IEEE Access. 2026;14:169474-169489. doi:10.1109/access.2026.3748171
BibTeX
@article{Lee2026,
author = {Lee, Ganghoon and Kim, Young-Jun and Kwon, Junseok},
title = {When Do Simple Methods Suffice? Fair Baselines for Denoising, Trace Quality Control, and Acoustic-Texture Clustering of Chirp Sub-Bottom Profiler Data},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {169474--169489},
doi = {10.1109/access.2026.3748171}
}
EUVIMED