Pith. sign in

REVIEW 4 major objections 6 minor 40 references

A framework for river connectivity classification using temporal image processing and attention based neural networks

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that combining temporal luma averaging with a vision transformer lifts unseen-site river connectivity classification from 75% to 90% accuracy.

desk verdict Useful applied problem with a sound site-level holdout, but the 90% accuracy claim is unverifiable as reported: no results table, unclear beta in Eq. 8, and no data release. read the letter →

arxiv 2502.00474 v1 pith:EXIOEB6I submitted 2025-02-01 cs.CV cs.LGeess.IV

Timothy James Becker , Derin Gezgin , Jun Yi He Wu , Mary Becker This is my paper · ORCID
classification cs.CVcs.LGeess.IV MSC 68T4568T0768U10
keywords riverconnectivitytrailcameraimagerytemporalimageenhancementvisiontransformerclassificationdiffusionmodelaugmentationstreammonitoringunseensitegeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that an automated, site-agnostic classifier can replace manual curation of trail-camera images for river connectivity monitoring. The authors assemble a seven-filter image-quality screen, a temporal enhancement that averages the brightness channel across neighboring hourly images to suppress foliage shadows, and a vision transformer classifier, and they report that this pipeline raises accuracy on images from sites never seen in training from 75% to 90%. This matters because stream gauges cover fewer than 1% of U.S. rivers and streams, while trail cameras are cheap and already deployed; automating the labeling burden would make connectivity data available for small headwater streams where drying events are hardest to track. Along the way the paper shows that the same temporal enhancement helps both ResNet and vision transformer models, that six-class labels are too hard for the current setup, and that neither geometric nor diffusion-based augmentation improves the binary result.

What carries the argument

The load-bearing component is a temporal luma-enhancement identity expressed in Eqs. 7 and 8. Each image's Y (luma) channel in YCrCb space is averaged over $\alpha$ neighboring hourly images to form $\mu$, then the enhanced luma is computed as $X + \beta(X - \mu)$, which reduces local shadow and highlight variance from foliage while preserving water-connectivity gradients. Around this sits the vision transformer, which divides the image into patches and applies self-attention so the model can compare distant regions such as foliage, sky, riverbed, and water surface without site-specific feature engineering; a site-level partitioning scheme (Eqs. 2–6) keeps all images from held-out sites out of training so the reported accuracy is a genuine unseen-site estimate. Diffusion-model augmentation was also built, but the paper reports that the completed comparison did not improve on the temporal ViT.

What would settle it

Re-run the e2hr_no_aug_vit experiment on the same held-out sites with Eq. 8 disabled (no temporal luma averaging) but otherwise identical preprocessing and training; if overall accuracy stays near 90% or the disconnected-class F1 does not drop, the temporal enhancement is not the cause of the headline result. A still more direct check is a site-heavy in foliage shadow where the averaging is supposedly beneficial, comparing models with $\beta=0$, $\beta=1$, and the tuned value.

Watch

Extended reading notes

Core claim

The central claim, on the authors' own terms, is that the general river connectivity classification problem—deciding whether a trail-camera image shows a connected or disconnected stream at a site the model has not seen—can be solved to about 90% accuracy by combining temporal image enhancement with an attention-based vision transformer. The recipe is: filter seven classes of bad images, convert to YCrCb, average the luma channel over a window of neighboring hourly images, mix the averaged luma back into the original with a beta parameter, bottom-center crop and resize, and train a ViT on balanced data. The authors report that the 2-hour temporal window raises the ViT's combined F1 to roughly 0.87 and overall accuracy from a 75% base, with the same enhancement also improving a ResNet baseline, and they interpret the gain as richer visual detail in riverbanks and water surfaces rather than mere noise removal.

Load-bearing premise

The reported gain assumes that averaging the brightness channel across neighboring hourly images suppresses foliage shadows without erasing the water-connectivity signal; the paper supports this only with qualitative example images and gives no sensitivity analysis for the mixing weight $\beta$ or the window size $\alpha$.

Editorial extensions

If this is right

  • Agencies can use the trained pipeline to triage hourly trail-camera images automatically, flagging disconnected states for human review instead of labeling every frame.
  • Because temporal enhancement improves both ResNet and ViT, the benefit is not confined to one architecture, so downstream work can pick the lighter model for deployment.
  • The binary connected/disconnected model is a realistic product, while the six-category labels remain too confusable near adjacent classes for the current pipeline.
  • Neither geometric augmentation nor the completed diffusion-model comparison outperformed the raw 2-hour temporal ViT, so balancing strategies for this domain still need a different approach.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A sensitivity analysis varying $\beta$ (the luma mix weight) and the temporal window $\alpha$ would show how much of the 90% depends on the exact enhancement recipe; the paper only gives qualitative examples.
  • The confusion pattern near the 3–4 boundary suggests a three-class model (disconnected, transitional, connected) might give managers more reliable information than six classes.
  • Since the trained models are small enough for CPU-only inference, embedding the classifier in trail cameras for near-real-time alerts is a plausible next step that the paper only gestures at.
  • Sites with sparse or irregular hourly coverage would stress the temporal enhancement, so testing on such deployments would bound the method's real-world reach.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a pipeline for classifying river connectivity from trail camera images into six categories (and a binary disconnected/connected task). The pipeline consists of seven image-quality filters, a temporal luma-channel enhancement that averages hourly images within a window α and mixes the difference back with parameter β, bottom-center cropping and resizing, optional geometric and diffusion-model-based generative augmentation, and classification with ResNet or vision transformer (ViT) architectures. The central claim, stated in the abstract, is that temporal enhancement plus ViT raises unseen-site accuracy from 75% to 90%. The paper also describes a site-level data partitioning strategy intended to prevent site leakage and reports exploratory six-class confusion results.

Significance. If the reported figures are reliable, the framework would be a practical contribution to low-cost stream connectivity monitoring: it targets generalization to entirely unseen camera sites, which distinguishes it from prior single-site streamflow regression work. The site-level holdout partitioning described in Section 2A is a genuine strength, as it directly addresses the site-leakage problem that commonly inflates accuracy estimates. The temporal luma enhancement idea is domain-motivated and, if properly specified, could be a useful preprocessing technique. However, the paper's current reporting does not support the headline 75%→90% claim: there is no numeric results table, no confidence intervals, no site/image counts, the sign and value of β in Eq. (8) are unspecified, and the choice of the 2-hour window appears to be made after inspecting the experimental outcomes. The empirical claim is plausible but not yet established with the level of rigor required for a journal publication.

major comments (4)
  1. [Section 2B, Eqs. (7)-(8)] The sign and value of β are never stated. As written, Eq. (8) sets X_enh = X + β(X − μ_luma); with β > 0 this amplifies deviations from the luma mean and increases local luma variance, which contradicts the text's claim that the operation 'lowers the total variance of the luma channel.' With β < 0 the operation is variance-reducing, as the text describes, but the paper does not state which case is used or what value of β (and α) was chosen. Because the central accuracy gain is attributed to this enhancement, the paper must specify the sign, range, and chosen values of β and α, and include a sensitivity analysis.
  2. [Section 3, Temporal Enhancement Effect and Figure 6] The results section contains no numeric accuracy table. The only explicit F1 values in the paper are 0.87 (2-hr ViT) and 0.73 (2-hr ViT with generative augmentation) in the 'Generative Augmentation Effects' subsection; the abstract's 75%→90% accuracy figures cannot be located in any table or equation, and the bars in Figure 6 are unquantified. Without exact accuracy and F1 values for each condition, the number of sites and images in each partition, and confidence intervals, the central claim is not testable. Please provide a full results table with sample sizes and uncertainty estimates.
  3. [Section 3, Generative Augmentation Effects; Section 2A] The choice of the 2-hour temporal window as the 'medium' level appears to be made after comparing the 0-hour, 2-hour, and 4-hour results in Figure 6, and the paper does not describe a separate validation protocol for this selection. Combined with the absence of confidence intervals or repeated runs, the reported improvement could reflect validation-set tuning. Please state explicitly how α and β were chosen (fixed a priori or tuned on a defined validation split) and confirm that the final held-out site IDs were used only once for the reported metrics.
  4. [Section 3, Performance Metrics, Eq. (10)] The definition of the 'combined m class label F1 accuracy' is unreadable as typeset: 'F1 = 4∏...∑...' does not correspond to macro-averaged F1, micro-averaged F1, or any standard F1 aggregation. Since combined F1 is a primary reported metric in Figures 6 and 7, please rewrite this formula and define the averaging procedure over classes clearly.
minor comments (6)
  1. [Section 2A, Eq. (1)] The formula for P(𝕐=y) contains stray symbols and is not readable; please re-typeset it and define all notation consistently.
  2. [Section 2A, Eqs. (3)-(6)] The partitioning optimization notation is unclear: p_u and p_t are defined in Eq. (5) but do not appear in Eq. (6), and the constraint that the subsets union to the full site set is written ambiguously. Please clarify the algorithm and the role of the split parameter θ.
  3. [Section 2B and Section 3] The sentence 'We generalizable the removal to become a localized gradient' is ungrammatical, and Section 3's 'Basic Augmentation Effects' refers to histogram equalization, which is not described in the augmentation methods of Section 2D.
  4. [Figures 6 and 7] The bar charts do not have numeric axes or value labels, so the reader cannot read off even approximate accuracies or F1 scores. Please add numeric values or a corresponding table.
  5. [References] There are several citation inconsistencies, including 'Vanswani et al.' in the Introduction versus 'Vaswani et al.' in Section 2C, and 'Kan et al.' which should likely be 'Khan et al.'; please standardize all references.
  6. [General] Please include a data and code availability statement, since the empirical claims depend on the CT DEEP dataset and the exact preprocessing and training pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central accuracy claim is an empirical held-out measurement, not a derivation, and no load-bearing step reduces to its own inputs.

full rationale

The paper's claimed result is an empirical accuracy measurement on a validation partition of site IDs that are fully excluded from training, so the main claim is not derived from its inputs by construction. The temporal enhancement in Eq. 8 is a preprocessing transform with hand-set parameters α and β; these parameters are not fitted to reproduce the reported accuracy, and the paper does not claim Eq. 8 is derived from the classifier output. The comparison of 0-hour, 2-hour, and 4-hour enhancement windows is an ablation experiment, not a circular definition. The ViT architecture is cited from external literature, and the dataset is cited to Bellucci et al. 2020, which includes an author overlap but serves as data provenance rather than load-bearing circular support. No uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. The paper does have reporting and reproducibility weaknesses: the abstract's 75% to 90% accuracy figure is not located in a numeric table, the sign and values of β in Eq. 8 are not specified, and the effect of temporal enhancement is only shown qualitatively in Figure 6. However, these are correctness, verifiability, and clarity concerns, not circularity. Therefore the circularity score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-set image processing parameters (alpha, beta, crop, filter thresholds) and on domain assumptions about label correctness and temporal stationarity of trail camera scenes. No new physical entities are introduced. The dataset itself is proprietary and not released, which is a major reproducibility limitation.

free parameters (7)
  • Temporal aggregation window alpha (hours before/after) = not reported (experiments at 0, 2, and 4 hours)
    Controls the number of neighboring hourly images averaged in Eq. 7; the best window is selected after seeing validation performance.
  • Enhancement mixing parameter beta = not reported
    Controls how much average luma is mixed into each image in Eq. 8; no sensitivity analysis is given.
  • Bottom-center crop fraction and resize dimensions = not reported
    Preprocessing changes the input geometry and image content; exact crop and resize parameters are absent.
  • Image quality filter thresholds = not reported
    The seven filters (bright, dark, blur, flare, triggered, black-and-white, bad timestamp) have unspecified thresholds that determine which images enter the dataset.
  • Augmentation rotation angle and resize factor = 5 degrees, 1.3x resize
    Chosen by observation to avoid upside-down rivers and excessive distortion; not tuned systematically.
  • Diffusion model training hyperparameters = not reported
    The DDM-based generative augmentation setup is described at a high level but key training details are missing.
  • Class balancing augmentation amounts = variable
    The paper states variable amounts of generative augmentation are used to balance labels, but exact quantities are not specified.
assumptions (5)
  • standard math Vision transformers and deep residual networks can learn image classification from labeled data.
    Invoked in Section 2C as the machine learning foundation; no proof needed, but their suitability for this specific task is an empirical assumption.
  • domain assumption The CT DEEP staff labels for the six connectivity categories are correct ground truth.
    The supervised formulation in Section 2A treats labels from Bellucci et al. (2020) as true labels; no inter-rater reliability or label verification is reported.
  • domain assumption Temporal luma averaging over neighboring hourly images removes foliage shadow while preserving the water connectivity signal.
    Eqs. 7-8 in Section 2B assume that averaging the Y channel across time suppresses shadows without erasing the water presence that defines the label; only qualitative examples support this.
  • domain assumption EXIF timestamps are reliable and images within the enhancement window are from the same camera perspective.
    Section 2B clusters hourly images by EXIF time; if cameras shift or deployments overlap, the temporal enhancement could mix images from different scenes.
  • standard math Diffusion models can generate synthetic images from the training distribution.
    Used in Section 2D based on Ho et al. (2020) and Dhariwal and Nichol (2021); the authors assume the generated images are useful for balancing the dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A framework for river connectivity classification using temporal image processing and attention based neural networks." pith.science (2026). https://pith.science/paper/EXIOEB6I

@misc{pith2026250200474,
  author       = {Pith},
  title        = {Pith review of: A framework for river connectivity classification using temporal image processing and attention based neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EXIOEB6I}},
  note         = {Machine review of arXiv:2502.00474}
}
read the original abstract

Measuring the connectivity of water in rivers and streams is essential for effective water resource management. Increased extreme weather events associated with climate change can result in alterations to river and stream connectivity. While traditional stream flow gauges are costly to deploy and limited to large river bodies, trail camera methods are a low-cost and easily deployed alternative to collect hourly data. Image capturing, however requires stream ecologists to manually curate (select and label) tens of thousands of images per year. To improve this workflow, we developed an automated instream trail camera image classification system consisting of three parts: (1) image processing, (2) image augmentation and (3) machine learning. The image preprocessing consists of seven image quality filters, foliage-based luma variance reduction, resizing and bottom-center cropping. Images are balanced using variable amount of generative augmentation using diffusion models and then passed to a machine learning classification model in labeled form. By using the vision transformer architecture and temporal image enhancement in our framework, we are able to increase the 75% base accuracy to 90% for a new unseen site image. We make use of a dataset captured and labeled by staff from the Connecticut Department of Energy and Environmental Protection between 2018-2020. Our results indicate that a combination of temporal image processing and attention-based models are effective at classifying unseen river connectivity images.

Figures

Figures reproduced from arXiv: 2502.00474 by the authors.

Figure 1
Figure 1. Disconnected state is shown in the top brown panel where we have Category 1, Category 2 and Category 3. Below the Connected state shown in green illustrates Category 4, Category 5 and Category 6 examples. All categories here are taken from the same site but at different time points (and deployments) showing the total span of all six categories across time. The related problem of stream flow regression has shown prog… view at source ↗
Figure 2
Figure 2. (A) The training phase filters low quality images before bottom-center auto-cropping and image enhancement is applied. (B) The data augmentation framework can generate more images for labels that are lower in frequency such as 1 or 6. (C) The augmented data is then given to the training model to fit and then saved for future inference. (D) The same preprocessing is used as in (A) but now the framework will not have … view at source ↗
Figure 3
Figure 3. [0] shows the temporal image before the currently processed image [1]. Image [0] has very low color variance and almost appears black and white in comparison to image [1] but it contains more details in the riverbank region to the right as well as more details in the water surface. These details take the form of luma gradients and by looking at differences between images (shown on the right as [0-1] luma diff, and t… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 23 canonical work pages

  1. [1]

    Increased extreme weather events associated with climate change can result in alterations to river and stream connectivity

    A framework for river connectivity classification using temporal image processing and attention based neural networks Timothy James Becker1, Derin Gezgin1, Jun Yi He Wu1, Mary Becker2 1Connecticut College, 2Connecticut Department of Energy and Environmental Protection Abstract Measuring the connectivity of water in rivers and streams is essential for effe...

    work page 2018

  2. [2]

    We looked to recent literature in real-time object tracking that used interframe images to remove shadows from the target object (people) [Wu et al

    We found that the passed images still contained incredibly complex shadows and highlights from foliage, tree trunks, branches and sun reflections on water surfaces. We looked to recent literature in real-time object tracking that used interframe images to remove shadows from the target object (people) [Wu et al. 2020]. We generalizable the removal to beco...

    work page 2020

  3. [3]

    &',…,𝑋%"(

    Below the Connected state shown in green illustrates Category 4, Category 5 and Category 6 examples. All categories here are taken from the same site but at different time points (and deployments) showing the total span of all six categories across time. The related problem of stream flow regression has shown progress using Deep Convolutional Neural Net a...

    work page 2020

  4. [5]

    ImageNet Classification with Deep Convolutional Neural Networks

    as shown in luma+[1] and hue+[1] we can see that shadow and highlights are reduced and detail is increased in the riverbanks and throughout the entire image. 2C Methods: Machine Learning Framework We implemented and then tuned two well established methods for classification: deep residual convolutional neural networks (RESNET) and vision transformers (ViT...

    work page 2021

  5. [6]

    (D) The same preprocessing is used as in (A) but now the framework will not have labels

    (C) The augmented data is then given to the training model to fit and then saved for future inference. (D) The same preprocessing is used as in (A) but now the framework will not have labels. Instead, the model file is loaded and then proceeds to directly to inference in (E). 2B Methods: Image Enhancement To increase the image quality for classification, ...

    work page 2020

  6. [7]

    It is possible that better model training parameters would score higher but we did not have time to fully explore this area

    Instead, we trained a ViT with the medium level of temporal enhancement (2hr) which resulted in 0.73 F1 accuracy which was lower than our ViT with medium level of temporal enhancement (2hr) which scored 0.87 F1 accuracy. It is possible that better model training parameters would score higher but we did not have time to fully explore this area. 4 Discussio...

    work page 2018

  7. [14]

    arXiv:2105.05233

    Diffusion Models Beat GANs on Image Synthesis. arXiv:2105.05233. Retrieved from https://arxiv.org/abs/2105.05233 Maayan Frid-Adar, Idit Diamant, Eyal Klang, Michal Amitai, Jacob Goldberger, and Hayit Greenspan

    arXiv

  8. [15]

    A Data Scientist's Guide to Streamflow Prediction

    GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification. Neurocomputing 321, 10 (December 2018), 321-331. https://doi.org/10.1016/j.neucom.2018.09.013 Martin Gauch and Jimmy Lin. A data scientist guide to streamflow prediction. arXiv:2006.12975. Retrieved from https://arxiv.org/abs/2006.12975 Ian J. Good...

    work page Pith review arXiv 2018

Show all 40 references
  1. [20]

    arXiv:2006.11239

    Denoising Diffusion Probabilistic Models. arXiv:2006.11239. Retrieved from https://arxiv.org/abs/2006.11239 Hugh B.N. Hynes,

    2006 arXiv

  2. [22]

    arXiv:2006.05132

    A Survey on Generative Adversarial Networks: Variants, Applications, and Training. arXiv:2006.05132. Retrieved from https://arxiv.org/abs/2006.05132 Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah

    2006 arXiv

  3. [23]

    ACM Computing Surveys

    Transformers in Vision: A Survey. ACM Computing Surveys. 54, 10s, Article 200 (January 2022). https://doi.org/10.1145/3505244 Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton

    2022 doi

  4. [25]

    Journal of Big Data 5, 1 (November 2018), 5-42

    A survey on addressing high-class imbalance in big data. Journal of Big Data 5, 1 (November 2018), 5-42. https://doi.org/10.1186/s40537-018-0151-6 Bohan Li, Xiao Xu, Xinghao Wang, Yutai Hou, Yunlong Feng, Feng Wang, Xuanliang Zhang, Qingfu Zhu, and Wanxiang Che

    2018 doi

  5. [27]

    arXiv:2201.09865

    RePaint: Inpainting using Denoising Diffusion Probabilistic Models. arXiv:2201.09865. Retrieved from https://arxiv.org/abs/2201.09865 Megan C. Malish, Shang Gao, Darin Kopp, Yang Hong, Daniel C. Allen and Thomas Neeson

    arXiv

  6. [29]

    JAWRA Journal of the American Water Resources Association 43, 1 (February 2007), 86-103

    The Contribution of Headwater Streams to Biodiversity in River Networks. JAWRA Journal of the American Water Resources Association 43, 1 (February 2007), 86-103. https://doi.org/10.1111/j.1752-1688.2007.00008.x Simone Noto, Flavia Tauro, Andrea Petroselli, Ciro Apollonio, Gian...

    2021

  7. [32]

    arXiv:2010.02502

    Denoising Diffusion Implicit Models. arXiv:2010.02502. Retrieved from https://arxiv.org/abs/2010.02502 Rachel Stubbington, Michael T. Bogan, Núria Bonada, Andrew J. Boulton, Thibault Datry, Catherine Leigh, and Ross Vander V orste

    2010 arXiv

  8. [33]

    arXiv:2302.07944

    Effective Data Augmentation With Diffusion Models. arXiv:2302.07944. Retrieved from https://arxiv.org/abs/2302.07944 (USGS) United States Geological Survey

    arXiv

  9. [34]

    (March 2023)

    The Past, Present and Future of USGS Streamgages. (March 2023). Retrieved January 11, 2025 from https://www.usgs.gov/news/featured-story/past-present-and-future-usgs-streamgages (USGS) United States Geological Survey

    2023

  10. [35]

    Geological Survey National Water Information System database (January 2025)

    USGS water data for the Nation: U.S. Geological Survey National Water Information System database (January 2025). Retrieved January 11, 2025 from https://doi.org/10.5066/F7P55KJN Robin L. Vannote, G. Wayne Minshall, Kenneth W. Cummins, James R. Sedell, and Colbert E. Cushing

    2025 doi

  11. [37]

    In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17)

    Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17). Curran Associates Inc., Red Hook, NY , USA, 6000–6010. https://dl.acm.org/doi/10.5555/3295222.3295349 Minghu Wu, Rui Chen, Ying Tong. Shadow Elimi...

  12. [38]

    Progress in Physical Geography: Earth and Environment 41, 3 (June 2017), 345–362

    Connectivity in rivers. Progress in Physical Geography: Earth and Environment 41, 3 (June 2017), 345–362. https://doi.org/10.1177/03091333177149 Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang

    2017 doi

  13. [39]

    arXiv:2101.05278

    GAN Inversion: A Survey. arXiv:2101.05278. Retrieved from https://arxiv.org/abs/2101.05278 Margaret A. Zimmer, Kendra E. Kaiser, Joanna R. Blaszczak, Samuel C. Zipper, John C. Hammond, Ken M. Fritz, Katie H. Costigan, Jacob Hosen, Sarah E. Godsey, George H. Allen, Stephanie Ka...

    arXiv

  14. [40]

    WIREs Water 7, 3 (May/June 2020), 25 pages

    Zero or not? Causes and consequences of zero-flow stream gage readings. WIREs Water 7, 3 (May/June 2020), 25 pages. https://doi.org/10.1002/wat2.1436 Samuel C Zipper, John C Hammond, Margaret Shanafield, Margaret Zimmer, Thibault Datry, C Nathan Jones, Kendra E Kaiser, Sarah E...

    2020 doi

  15. [41]

    Environmental Research Letters 16, 8 (July 2021), 16 pages

    Pervasive changes in stream intermittency across the United States. Environmental Research Letters 16, 8 (July 2021), 16 pages. https://doi.org/10.4211/hs.fe9d240438914634abbfd-cfa03bed863

    2021 doi

  16. [270]

    International Conference on Learning Representations, Vienna Austria,

    https://doi.org/10.1890/100053 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby, An Image is worth 16x16 Words: Transforme...

    doi

  17. [363]

    Deep Residual Learning for Image Recognition

    https://doi.org/10.1145/3530190.3534805 Kaiming He; Xiangyu Zhang; Shaoqing Ren; Jian Sun. Deep Residual Learning for Image Recognition. Conference on Computer Vision and Pattern Recognition. Las Vegas, USA,

  18. [1975]

    SIL Proceedings 1922-2010 19, 1 (October 1975), 1–15

    The Stream and Its Valley: With 4 Figures and 2 Tables in the Text. SIL Proceedings 1922-2010 19, 1 (October 1975), 1–15. https://doi.org/10.1080/03680770.1974.11896033 Abdul Jabbar, Xi Li, and Bourahla Omar

    1922

  19. [1980]

    Canadian Journal of Fisheries and Aquatic Sciences 37, 1 (January 1980), 130–137

    The River Continuum Concept. Canadian Journal of Fisheries and Aquatic Sciences 37, 1 (January 1980), 130–137. https://doi.org/10.1139/f80-01 Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin

    1980 doi

  20. [2002]

    Environmental Management 30, 4 (October 2002), 492–507

    Basic Principles and Ecological Consequences of Altered Flow Regimes for Aquatic Biodiversity. Environmental Management 30, 4 (October 2002), 492–507. https://doi.org/10.1007/s00267-002-2737-0 Daren M Carlisle, David M Wolock, and Michael R Meador

    2002 doi

  21. [2014]

    arXiv:1406.2661

    Generative Adversarial Networks. arXiv:1406.2661. Retrieved from https://arxiv.org/abs/1406.2661 Amrita Gupta, Tony Chang, Jeffrey Walker, and Benjamin Letcher

    arXiv

  22. [2015]

    arXiv:1505.04597

    U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv:1505.04597. Retrieved from https://arxiv.org/abs/1505.04597 Connor Shorten and Taghi M. Khoshgoftaar

    arXiv

  23. [2016]

    https://doi.org/10.1109/CVPR.2016.90 Jonathan Ho, Ajay Jain, and Pieter Abbeel

    2016 doi

  24. [2017]

    ImageNet classification with deep convolutional neural networks. Commun. ACM 60, 6 (June 2017), 84–90. https://doi.org/10.1145/3065386 Joffrey L. Leevy, Taghi M. Khoshgoftaar, Richard A. Bauder, and Naeem Seliya

    2017 doi

  25. [2018]

    arXiv:1711.04340

    Data Augmentation Generative Adversarial Networks. arXiv:1711.04340. Retrieved from https://arxiv.org/abs/1711.04340 Christopher J. Bellucci, Mary E. Becker, Melissa Czarnowski, and Corinne Fitting

    arXiv

  26. [2019]

    Journal of Big Data 6, 1 (July 2019), 6-60

    A survey on Image Data Augmentation for Deep Learning. Journal of Big Data 6, 1 (July 2019), 6-60. https://doi.org/10.1186/s40537-019-0197-0 Jiaming Song, Chenlin Meng, and Stefano Ermon

    2019 doi

  27. [2020]

    River Research and Applications 36, 8 (October 2020), 1504-1514

    A novel method to evaluate stream connectivity using trail cameras. River Research and Applications 36, 8 (October 2020), 1504-1514. https://doi.org/10.1002/rra.3689 András Béres

    2020 doi

  28. [2021]

    https://doi.org/10.48550/arXiv.2010.11929 Prafulla Dhariwal and Alex Nichol

    2010 doi arXiv

  29. [2022]

    Towards Continuous Streamflow Monitoring with Time-Lapse Cameras and Deep Learning. In Proceedings of the 5th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, 29 June – 01 July 2022, Seattle WA USA, Association for Computing Machinery, New York, NY , USA, 353 -

    2022

  30. [2023]

    Ecosphere 14, 3 (March 2023), 16 pages

    Small increases in stream drying can dramatically reduce ecosystem connectivity. Ecosphere 14, 3 (March 2023), 16 pages. https://doi.org/10.1002/ecs2.4450 Judy L. Meyer, David L. Strayer, J. Bruce Wallace, Sue L. Eggert, Gene S. Helfman, and Norman E. Leonard

    2023 doi

  31. [2024]

    arXiv:2302.02070

    Semantic-Guided Generative Image Augmentation Method with Diffusion Models for Image Classification. arXiv:2302.02070. Retrieved from https://arxiv.org/abs/2302.02070 Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool

    arXiv

  32. [2025]

    arXiv:2407.04103

    Advances in Diffusion Models for Image Data Augmentation: A Review of Methods, Models, Evaluation Metrics and Future Research Directions. arXiv:2407.04103. Retrieved from https://arxiv.org/abs/2407.04103 Antreas Antoniou, Amos Storkey, and Harrison Edwards

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.