arxiv: 2605.10953 · v1 · submitted 2026-04-30 · ⚛️ physics.geo-ph · cs.CV· cs.LG

Recognition: 2 theorem links

· Lean Theorem

Parameter-Efficient Adaptation of Pre-Trained Vision Foundation Models for Active and Passive Seismic Data Denoising

Jiahua Zhao , Umair Bin Waheed , Jing Sun , Yang Cui , Nikos Savva , Eric Verschuur

Authors on Pith no claims yet

Pith reviewed 2026-05-13 07:55 UTC · model grok-4.3

classification ⚛️ physics.geo-ph cs.CVcs.LG

keywords seismic denoisingvision foundation modelsparameter-efficient fine-tuningLoRA adaptationunsupervised test-time adaptationDAS datageophysical signal processing

0 comments

The pith

Pre-trained vision models adapt to denoise seismic data across sites using few added parameters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that a vision foundation model pre-trained on natural images can be repurposed for suppressing noise in active and passive seismic recordings by updating only a small subset of parameters. It combines low-rank adaptation on a DINOv3 encoder with a test-time module that uses kurtosis to locate signal-rich patches and then performs unsupervised self-training on those patches. On public exploration datasets and DAS vertical seismic profiling records from Utah, the method matches or exceeds the accuracy of models trained from scratch on seismic data. Further tests on land surveys from China and a geothermal site in Germany confirm that the adapted model separates signal from noise on data never seen during training. This matters because collecting large paired clean-noisy seismic datasets is costly, so methods that transfer knowledge efficiently and self-calibrate on new sites reduce both data and compute demands.

Core claim

A DINOv3 encoder pre-trained on natural images, when adapted via Low-Rank Adaptation and equipped with a kurtosis-guided unsupervised test-time adaptation module, produces denoised seismic images that match or surpass the quality of domain-specific supervised models on exploration seismic data and DAS VSP records while generalizing to unseen cross-site acquisitions in China and Germany.

What carries the argument

Low-Rank Adaptation (LoRA) applied to a pre-trained DINOv3 vision encoder, paired with a kurtosis-based module that selects information-rich patches for unsupervised self-training of only the LoRA weights during inference.

If this is right

New seismic surveys can be processed without collecting large paired training sets or retraining entire networks from scratch.
The same adapted model handles both active-source exploration data and passive DAS monitoring records without task-specific redesign.
Cross-site generalization reduces the need to recollect site-specific labeled examples for every new acquisition geometry.
Signal fidelity remains high enough for downstream steps such as imaging and event detection on dense array data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same LoRA-plus-kurtosis recipe could be tested on other geophysical waveform tasks such as first-break picking or microseismic event detection.
If the kurtosis selection step proves stable across more environments, it could support fully online unsupervised pipelines for real-time field processing.
Lower parameter counts may allow the framework to run on edge devices deployed with temporary seismic arrays.

Load-bearing premise

Features extracted from natural-image pre-training transfer usefully to the structure of seismic waveform images, and kurtosis reliably identifies signal-rich patches without creating artifacts or suppressing real events.

What would settle it

On a held-out seismic dataset the adapted model produces visibly lower signal-to-noise ratios or coherent event suppression compared with a standard bandpass filter or an existing seismic-specific denoiser.

Figures

Figures reproduced from arXiv: 2605.10953 by Eric Verschuur, Jiahua Zhao, Jing Sun, Nikos Savva, Umair Bin Waheed, Yang Cui.

**Figure 2.** Figure 2: Overview of the proposed framework and its key components. (a) The encoder [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Figure 3: Visual comparison of denoised results across the five configurations on a challenging [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

**Figure 4.** Figure 4: Visual comparison of the removed noise profiles corresponding to Figure 3. (a) [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗

**Figure 5.** Figure 5: Visual comparison of the denoising results achieved by two state-of-the-art baselines [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗

**Figure 6.** Figure 6: Denoising results of our three adapted models evaluated on the representative DAS [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗

**Figure 7.** Figure 7: Comparison of denoising generalization on unseen land seismic data from China. (a) [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗

**Figure 8.** Figure 8: Generalization test on unseen DAS-VSP data from the Groß Sch¨onebeck site. (a) [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗

read the original abstract

The demand for high-resolution subsurface imaging and continuous Earth monitoring has driven rapid growth in active and passive seismic data from dense geophone deployments, distributed acoustic sensing (DAS) arrays, and large-scale 2D and 3D surveys. This expansion makes complex noise suppression increasingly challenging, especially when signal fidelity must be preserved. Conventional supervised deep learning methods are often task-specific, require large paired datasets, and can suffer from domain shift under new acquisition conditions. Foundation models offer a promising alternative, but pre-training seismic foundation models from scratch requires massive domain-specific data and substantial computation. We propose an efficient framework that repurposes general-purpose Vision Foundation Models (VFMs) for geophysical tasks through Parameter-Efficient Fine-Tuning. The architecture uses a pre-trained VFM, a DINOv3 encoder, adapted with Low-Rank Adaptation (LoRA) to enable effective feature adaptation with few additional parameters. To improve robustness under unseen field conditions without ground truth, we introduce a kurtosis-guided unsupervised test-time adaptation module that updates only LoRA parameters during inference. This module self-calibrates the model to site-specific noise by identifying information-rich regions via kurtosis and performing self-training without labeled data. Experiments on public exploration seismic images and DAS vertical seismic profiling data from the Utah FORGE site show that the framework matches or outperforms domain-specific models. Tests on unseen cross-site data from a land survey in China and the Gro{\ss} Sch\"onebeck geothermal site in Germany further demonstrate strong generalization and effective signal-noise separation. These results highlight the potential of adapting pre-trained VFMs to data-intensive problems in exploration seismology.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adapts DINOv3 via LoRA plus a kurtosis-guided test-time module for seismic denoising and claims good cross-site results, but the abstract gives no metrics or ablations so the performance edge is hard to assess.

read the letter

This paper takes a pre-trained vision model, DINOv3, and adapts it for active and passive seismic denoising with LoRA for low-parameter tuning plus a kurtosis-based unsupervised test-time adaptation step. The goal is to avoid training large domain-specific networks from scratch and to handle new acquisition conditions without labels. They report results on Utah FORGE DAS data and public exploration images, then test generalization on land data from China and the Groß Schönebeck site. The parameter-efficient route is the practical part that stands out, since it keeps added parameters small while claiming to match or beat specialized models on signal-noise separation. The cross-site tests are the main evidence offered for robustness. The soft spots are clear from the abstract alone. No quantitative metrics, baseline details, or error analysis appear, so it is impossible to judge how large the gains actually are or whether they hold under standard seismic metrics. The kurtosis module for picking signal-rich patches for self-training is the least secured piece; seismic sections contain impulsive noise and coherent trains whose kurtosis can overlap with real events, and nothing in the description shows an ablation or correlation check against labeled events to confirm the selection works. If high-kurtosis patches are mostly noise, the updates would reinforce artifacts rather than clean the data. This is aimed at geophysicists who process field seismic records and want to try foundation-model transfer without massive new training sets. A reader working on parameter-efficient adaptation for scientific imaging would find the setup straightforward to follow. It deserves a serious referee because the core idea is concrete and the application is timely, even though the current evidence is light; review would force the missing numbers and checks on the adaptation step.

Referee Report

2 major / 1 minor

Summary. The paper proposes a parameter-efficient framework that adapts a pre-trained Vision Foundation Model (DINOv3 encoder) via Low-Rank Adaptation (LoRA) for denoising both active and passive seismic data. It introduces a kurtosis-guided unsupervised test-time adaptation module that identifies information-rich patches for self-training on unseen sites without ground truth. Experiments on public exploration seismic images, Utah FORGE DAS VSP data, and cross-site tests from a China land survey and the Groß Schönebeck geothermal site are reported to match or outperform domain-specific models while demonstrating generalization and effective signal-noise separation.

Significance. If the central claims hold, the work shows that general-purpose vision foundation models can be efficiently repurposed for geophysical denoising tasks with very few trainable parameters and no labeled adaptation data. The combination of LoRA with kurtosis-based self-calibration offers a practical route to rapid deployment on new acquisition geometries, which is valuable for large-scale seismic monitoring where collecting paired clean/noisy datasets is costly. The cross-site generalization results, if quantitatively substantiated, would be a notable strength.

major comments (2)

[Kurtosis-guided unsupervised test-time adaptation module] Kurtosis-guided test-time adaptation module: the assumption that high kurtosis reliably flags signal-rich patches (rather than impulsive or coherent noise) is load-bearing for the cross-site generalization claims on China land data and Groß Schönebeck, yet no ablation, correlation with labeled events, or kurtosis distribution analysis is provided to verify the selection criterion. Overlap between noise and signal kurtosis distributions in seismic waveforms could cause the self-training to reinforce artifacts, directly undermining the reported outperformance.
[Experiments on public and field datasets] Experimental evaluation: the abstract states that the framework 'matches or outperforms domain-specific models' on multiple datasets, but the description supplies no quantitative metrics (e.g., SNR, MSE, or perceptual scores), baseline details, or statistical error analysis. Without these, the strength of the central performance and generalization claims cannot be assessed.

minor comments (1)

The abstract contains a formatting issue with 'Groß Schönebeck' that should be corrected for consistency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address each major comment point-by-point below, providing clarifications and committing to specific revisions that strengthen the validation of our claims.

read point-by-point responses

Referee: Kurtosis-guided test-time adaptation module: the assumption that high kurtosis reliably flags signal-rich patches (rather than impulsive or coherent noise) is load-bearing for the cross-site generalization claims on China land data and Groß Schönebeck, yet no ablation, correlation with labeled events, or kurtosis distribution analysis is provided to verify the selection criterion. Overlap between noise and signal kurtosis distributions in seismic waveforms could cause the self-training to reinforce artifacts, directly undermining the reported outperformance.

Authors: We agree that the kurtosis-guided selection is a central assumption whose robustness must be demonstrated, particularly for the cross-site results. While our design draws on the established property that seismic signals tend to produce higher kurtosis than Gaussian noise, we recognize the risk of overlap with impulsive or coherent noise. In the revised manuscript we will add: (1) an ablation study that compares the full kurtosis-guided module against a version that selects patches uniformly or by energy; (2) histograms and statistical summaries of kurtosis values computed separately on manually labeled signal and noise patches from the Utah FORGE dataset; and (3) a quantitative correlation between selected high-kurtosis patches and independently detected events. These additions will directly address the possibility of artifact reinforcement and provide the missing empirical support for the generalization claims. revision: yes
Referee: Experimental evaluation: the abstract states that the framework 'matches or outperforms domain-specific models' on multiple datasets, but the description supplies no quantitative metrics (e.g., SNR, MSE, or perceptual scores), baseline details, or statistical error analysis. Without these, the strength of the central performance and generalization claims cannot be assessed.

Authors: We acknowledge that the main-text narrative could more explicitly present the quantitative results. The manuscript already contains tables and figures reporting SNR, MSE, and visual comparisons on the public exploration data and Utah FORGE DAS VSP, together with baseline implementations (traditional filters and several domain-specific CNNs). To make these results fully transparent, we will expand the experimental section with: (i) a consolidated table listing mean SNR/MSE values plus standard deviations across repeated runs; (ii) explicit descriptions of all baseline architectures and training protocols; and (iii) statistical significance tests (paired t-tests) for the reported improvements. The cross-site experiments on the China and Groß Schönebeck data will likewise be augmented with the same quantitative metrics and error analysis. revision: yes

Circularity Check

0 steps flagged

Empirical adaptation of external pre-trained VFM shows no derivation circularity

full rationale

The paper proposes a practical framework that repurposes an external pre-trained DINOv3 vision foundation model via LoRA fine-tuning plus a kurtosis-based unsupervised test-time adaptation heuristic. All performance claims are grounded in experimental results on public seismic datasets (Utah FORGE, China land survey, Groß Schönebeck) rather than any closed mathematical derivation. No equations, fitted parameters, or self-citations are shown to reduce the reported denoising metrics to quantities defined inside the same paper by construction. The method is therefore self-contained against external benchmarks and pre-trained weights.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard transfer-learning assumptions rather than new physical postulates. Two key domain assumptions are required for the central claim to hold.

free parameters (2)

LoRA rank
Dimensionality of the low-rank update matrices; chosen as a hyperparameter to balance adaptation capacity and parameter count.
TTA learning rate and steps
Controls how far the LoRA parameters move during unsupervised inference-time updates.

axioms (2)

domain assumption Features extracted by a DINOv3 encoder pre-trained on natural images remain useful after LoRA adaptation to seismic waveform images
Invoked when the paper states that the pre-trained VFM can be repurposed for geophysical tasks.
domain assumption Kurtosis values computed on local patches reliably indicate the presence of coherent seismic signal versus noise
Used to select regions for self-training during test-time adaptation without ground truth.

pith-pipeline@v0.9.0 · 5621 in / 1388 out tokens · 44444 ms · 2026-05-13T07:55:13.788907+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

kurtosis-guided unsupervised test-time adaptation module that updates only the LoRA parameters during inference... identifies information-rich, high-kurtosis patches
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

LoRA... low-rank matrices A and B... h = W0x + BAx

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 7 internal anchors

[1]

Physica A: Statistical Mechanics and its Applications , volume=

Extremum criteria for optimal suppression of coherent noise in seismic data using the Karhunen--Loeve transform , author=. Physica A: Statistical Mechanics and its Applications , volume=. 2006 , publisher=

work page 2006
[2]

Annual Review of Earth and Planetary Sciences , volume=

Fiber-optic seismology , author=. Annual Review of Earth and Planetary Sciences , volume=. 2021 , publisher=

work page 2021
[3]

Sensors , volume=

Seismic applications of downhole DAS , author=. Sensors , volume=. 2021 , publisher=

work page 2021
[4]

Applied optics , volume=

Fading noise reduction in distributed acoustic sensing using an optimal weighted average algorithm , author=. Applied optics , volume=. 2021 , publisher=

work page 2021
[5]

Interpretation , volume=

Distributed acoustic sensing coupling noise removal based on sparse optimization , author=. Interpretation , volume=. 2019 , publisher=

work page 2019
[6]

Sensors , volume=

Diffusion model for DAS-VSP data denoising , author=. Sensors , volume=. 2023 , publisher=

work page 2023
[7]

Geophysics , volume=

Denoising distributed acoustic sensing data using unsupervised deep learning , author=. Geophysics , volume=. 2023 , publisher=

work page 2023
[8]

Geophysics , volume=

Continuously time-variable recursive digital band-pass filters for seismic signal processing , author=. Geophysics , volume=. 1983 , publisher=

work page 1983
[9]

SEG technical program expanded abstracts 1986 , pages=

FXDECON and complex Wiener prediction filter , author=. SEG technical program expanded abstracts 1986 , pages=. 1986 , publisher=

work page 1986
[10]

The Leading Edge , volume=

Noise elimination and the Radon transform, part 1 , author=. The Leading Edge , volume=. 1990 , publisher=

work page 1990
[11]

The Leading Edge , volume=

Noise elimination and the Radon transform, part 2 , author=. The Leading Edge , volume=. 1990 , publisher=

work page 1990
[12]

Geophysics , volume=

Ground-roll suppression using the wavelet transform , author=. Geophysics , volume=. 1997 , publisher=

work page 1997
[13]

International Conference on Medical image computing and computer-assisted intervention , pages=

U-net: Convolutional networks for biomedical image segmentation , author=. International Conference on Medical image computing and computer-assisted intervention , pages=. 2015 , organization=

work page 2015
[14]

Journal of Petroleum Science and Engineering , volume=

Seismic random noise suppression by using deep residual U-Net , author=. Journal of Petroleum Science and Engineering , volume=. 2022 , publisher=

work page 2022
[15]

Geophysics , volume=

Quadratic Unet for seismic random noise attenuation , author=. Geophysics , volume=. 2025 , publisher=

work page 2025
[16]

Geophysical Journal International , volume=

PhaseNet: a deep-neural-network-based seismic arrival-time picking method , author=. Geophysical Journal International , volume=. 2019 , publisher=

work page 2019
[17]

IEEE access , volume=

Unsupervised seismic random noise attenuation based on deep convolutional neural network , author=. IEEE access , volume=. 2019 , publisher=

work page 2019
[18]

Frontiers in Earth Science , volume=

Attenuating coherent environmental noise in seismic data via the U-net method , author=. Frontiers in Earth Science , volume=. 2023 , publisher=

work page 2023
[19]

First International Meeting for Applied Geoscience & Energy , pages=

Machine learning for seismic processing: The path to fulfilling promises , author=. First International Meeting for Applied Geoscience & Energy , pages=. 2021 , organization=

work page 2021
[20]

IEEE Transactions on Geoscience and Remote Sensing , volume=

Seismic signal denoising and decomposition using deep neural networks , author=. IEEE Transactions on Geoscience and Remote Sensing , volume=. 2019 , publisher=

work page 2019
[21]

Geophysics , volume=

Deep learning for denoising , author=. Geophysics , volume=. 2019 , publisher=

work page 2019
[22]

arXiv preprint arXiv:2506.22018 , year=

Seismic resolution enhancement via deep Learning with Knowledge Distillation and Domain Adaptation , author=. arXiv preprint arXiv:2506.22018 , year=

work page arXiv
[23]

Seismological Research Letters , volume=

Generalization of deep-learning models for classification of local distance earthquakes and explosions across various geologic settings , author=. Seismological Research Letters , volume=. 2024 , publisher=

work page 2024
[24]

Frontiers in Earth Science , volume=

Transfer learning for self-supervised, blind-spot seismic denoising , author=. Frontiers in Earth Science , volume=. 2022 , publisher=

work page 2022
[25]

Geophysics , volume=

Ground-truth-free deep learning for 3D seismic denoising and reconstruction with channel attention mechanism , author=. Geophysics , volume=. 2024 , publisher=

work page 2024
[26]

Interpretation , volume=

Synthetic seismic data for training deep learning networks , author=. Interpretation , volume=. 2022 , publisher=

work page 2022
[27]

GPT-4 Technical Report

Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[28]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Segment anything , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

work page
[29]

Geophysics , volume=

Seismic foundation model: A next generation deep-learning model in geophysics , author=. Geophysics , volume=. 2025 , publisher=

work page 2025
[30]

DINOv3

Dinov3 , author=. arXiv preprint arXiv:2508.10104 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[31]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Swin transformer v2: Scaling up capacity and resolution , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[32]

Journal of Geophysical Research: Machine Learning and Computation , volume=

Cross-domain foundation model adaptation: Pioneering computer vision models for geophysical data analysis , author=. Journal of Geophysical Research: Machine Learning and Computation , volume=. 2025 , publisher=

work page 2025
[33]

arXiv preprint arXiv:2503.24166 , year=

Foundation Models For Seismic Data Processing: An Extensive Review , author=. arXiv preprint arXiv:2503.24166 , year=

work page arXiv
[34]

, author=

Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

work page
[35]

IEEE Transactions on Geoscience and Remote Sensing , year=

Unsupervised Deep Learning for DAS-VSP Denoising Using Attention-Based Deep Image Prior , author=. IEEE Transactions on Geoscience and Remote Sensing , year=

work page
[36]

The thrity-seventh asilomar conference on signals, systems & computers, 2003 , volume=

Multiscale structural similarity for image quality assessment , author=. The thrity-seventh asilomar conference on signals, systems & computers, 2003 , volume=. 2003 , organization=

work page 2003
[37]

Comparison between distributed acoustic sensing and geophones: Downhole microseismic monitoring of the

Lellouch, Ariel and Lindsey, Nathaniel J and Ellsworth, William L and Biondi, Biondo L , journal=. Comparison between distributed acoustic sensing and geophones: Downhole microseismic monitoring of the

work page
[38]

Seismological Research Letters , year=2023, volume=94, issue=1, pages=

Denoising of distributed acoustic sensing seismic data using an integrated framework , author=. Seismological Research Letters , year=2023, volume=94, issue=1, pages=

work page 2023
[39]

Decoupled Weight Decay Regularization

Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[40]

Sourab Mangrulkar and Sylvain Gugger and Lysandre Debut and Younes Belkada and Sayak Paul and Benjamin Bossan and Marian Tietz , howpublished =

work page
[41]

IEEE transactions on image processing , volume=

Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

work page 2004
[42]

Geophysics , volume=

Local seismic attributes , author=. Geophysics , volume=. 2007 , publisher=

work page 2007
[43]

Geophysics , volume=

Random noise attenuation using local signal-and-noise orthogonalization , author=. Geophysics , volume=. 2015 , publisher=

work page 2015
[44]

Solid Earth , volume=

3D deep geothermal reservoir imaging with wireline distributed acoustic sensing in two boreholes , author=. Solid Earth , volume=. 2022 , publisher=

work page 2022
[45]

Journal of Applied Geophysics , volume=

Noncausal f--x--y regularized nonstationary prediction filtering for random noise attenuation on 3D seismic data , author=. Journal of Applied Geophysics , volume=. 2013 , publisher=

work page 2013
[46]

arXiv preprint arXiv:2104.10972 , year=

Imagenet-21k pretraining for the masses , author=. arXiv preprint arXiv:2104.10972 , year=

work page arXiv
[47]

Advances in neural information processing systems , volume=

Attention is all you need , author=. Advances in neural information processing systems , volume=

work page
[48]

Artificial Intelligence in Geosciences , volume=

The potential of self-supervised networks for random noise suppression in seismic data , author=. Artificial Intelligence in Geosciences , volume=. 2021 , publisher=

work page 2021
[49]

Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

Accurate, large minibatch sgd: Training imagenet in 1 hour , author=. arXiv preprint arXiv:1706.02677 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[50]

SGDR: Stochastic Gradient Descent with Warm Restarts

Sgdr: Stochastic gradient descent with warm restarts , author=. arXiv preprint arXiv:1608.03983 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[51]

2024 , publisher=

Lapins, Sacha and Butcher, Antony and Kendall, J-Michael and Hudson, Thomas S and Stork, Anna L and Werner, Maximilian J and Gunning, Jemma and Brisbourne, Alex M , journal=. 2024 , publisher=

work page 2024
[52]

2024 , journal=

Seismic node arrays for enhanced understanding and monitoring of geothermal systems , author=. 2024 , journal=

work page 2024
[53]

2024 , journal=

Seismic-led exploration and characterisation of carbon storage sites , author=. 2024 , journal=

work page 2024
[54]

Natural Hazards , year=

A review of seismic detection using fiber optic distributed acoustic sensing: from telecommunication cables to earthquake sensors , author=. Natural Hazards , year=. doi:10.1007/s11069-025-07370-5 , url=

work page doi:10.1007/s11069-025-07370-5
[55]

Photonics , year=

Research Advances on Distributed Acoustic Sensing Technology for Seismology , author=. Photonics , year=. doi:10.3390/photonics12030196 , url=

work page doi:10.3390/photonics12030196
[56]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations (ICLR) , year=. doi:10.48550/arXiv.2010.11929 , url=. 2010.11929 , archivePrefix=

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2010
[57]

2009 IEEE conference on computer vision and pattern recognition , pages=

Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

work page 2009
[58]

DINOv2: Learning Robust Visual Features without Supervision

DINOv2: Learning Robust Visual Features without Supervision , author=. arXiv preprint arXiv:2304.07193 , year=

work page internal anchor Pith review Pith/arXiv arXiv