Pith. sign in

REVIEW 3 major objections 5 minor 21 references

SuoiAI: Building a Dataset for Aquatic Invertebrates in Vietnam

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read SuoiAI proposes that an end-to-end pipeline of underwater cameras, semi-supervised labeling, and machine learning can build Vietnam's first aquatic-invertebrate dataset and classify species from it.

desk verdict Honest two-page workshop proposal for an aquatic-invertebrate dataset pipeline in Vietnam; no results to verify, and the annotation-budget numbers don't add up, but the gap it targets is real. read the letter →

arxiv 2504.15252 v1 pith:TSHT7325 submitted 2025-04-21 cs.AI cs.CVcs.LG

Tue Vo , Lakshay Sharma , Tuan Dinh , Khuong Dinh , Trang Nguyen , Trung Phan , Minh Do , Duong Vu This is my paper · ORCID
classification cs.AIcs.CVcs.LG
keywords aquaticinvertebratesVietnambiodiversityunderwatercameramonitoringsemi-supervisedlearningteacher-studentpseudo-labelingfine-grainedspeciesclassificationdatasetconstructionpipelinefreshwaterecology
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

SuoiAI is a proposed pipeline that would build a labeled image dataset of aquatic invertebrates in Vietnam and use machine learning to classify the species. The paper argues that a few hundred expert-labeled images, combined with underwater cameras and a loop in which an initial model labels new images and later models learn from those labels, can grow into thousands of labeled samples without a large manual annotation budget. This matters because Vietnam is severely underdocumented: of more than 100,000 described freshwater invertebrate species worldwide, only about 2,000 have been identified in Vietnam, and 800 aquatic species lack systematic documentation. If the pipeline works, ecologists gain a scalable tool for monitoring water quality and biodiversity in a climate-sensitive tropical region. The paper is explicitly a proposal; it reports no dataset, model, or experimental results yet.

What carries the argument

The load-bearing mechanism is the teacher-student bootstrap labeling loop, assisted by clustering and prompt-based segmentation. A teacher model trained on a few hundred expert-labeled images pseudo-labels new unlabeled captures; student models are trained on the mixture, and the cycle repeats. MorphoCluster groups visually similar images so annotators label clusters rather than single images, and SAM produces segmentation masks from prompts without training. This loop is what turns a small human effort into the large, fine-grained dataset the classification stage requires.

What would settle it

Deploy the proposed cameras at Cat Tien and Cuc Phuong through a full wet-dry cycle, then measure the fraction of frames in which a freshwater biologist can identify at least one invertebrate. If that fraction is too low to collect a few hundred high-quality manually labeled images within a season, the pipeline's central feasibility claim fails.

Watch

Extended reading notes

Core claim

The claim, stated on the paper's own terms, is that SuoiAI can deliver what has been missing for Vietnamese stream ecology: a labeled image dataset large enough and fine-grained enough to train species-level classifiers for aquatic invertebrates. The envisioned pipeline couples automated 1080p underwater cameras at two national parks with a few hundred expert-labeled images, then expands the labels through a teacher-student semi-supervised loop assisted by clustering and prompt-based segmentation. The paper expects roughly 20-50 genus-level classes and 100-200 species-level classes, with about 1,000 labeled samples per class, growing to 135 sites and about three million data points per site per year. If that holds, species classification and population tracking for Vietnamese stream invertebrates become feasible without a large manual annotation budget.

Load-bearing premise

The pipeline stands or falls on whether underwater cameras in Vietnamese streams capture images in which a person or model can actually see the invertebrates; if the water is too cloudy, dark, or cluttered, no labeling or learning method can recover the dataset.

Editorial extensions

If this is right

  • Vietnam would obtain its first systematic aquatic-invertebrate image dataset, shrinking the gap between the roughly 2,000 documented species and the estimated diversity.
  • Annotation cost per usable image drops enough that biodiversity datasets become feasible in regions that cannot hire large labeling teams.
  • Repeated deployments at 135 sites would allow population and range tracking for stream invertebrates, giving conservation managers a direct signal for water quality and climate effects.
  • A pretrained aquatic-invertebrate foundation model could emerge from the curated labels, letting downstream users classify images from nearby tropical regions with little or no extra training.

Reading between the lines

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

  • Beyond the paper's explicit claims, the transferable core is the bootstrap annotation loop, not any particular camera or detector; the same teacher-student plus clustering pattern applies to other underdocumented taxa where images are cheap and labels expensive.
  • Because the design targets 10-50 mm specimens with 1-5 individuals per frame, it is implicitly a benthic-macroinvertebrate system; applying it to planktonic or swimming fauna would require different camera placement and capture rates, not just different models.
  • A decisive early test would be a per-class annotation-cost comparison: if pseudo-labeling does not reduce effort on rare species relative to manual labeling alone, the economic argument collapses even if image quality is adequate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SuoiAI, an end-to-end pipeline for collecting and labeling underwater images of aquatic invertebrates in Vietnam and for training object detection and fine-grained classification models on them. It describes field deployment in two national parks, a semi-supervised annotation strategy using a teacher-student bootstrap, and several modeling approaches. The paper is explicitly a proposal: it reports no dataset, no experimental results, and no trained models.

Significance. The problem is important: Vietnam's freshwater invertebrate fauna is under-documented, and a publicly available dataset would be a valuable resource for ecology and conservation. If the proposed pipeline works as described, it could substantially reduce the annotation cost of building such a dataset and serve as a template for other tropical regions. The paper is transparent about being a proposal, and it draws on a reasonable set of established techniques (SAM, CLIP, semi-supervised object detection, fine-grained classification, clustering-based annotation). The main concern is that the central claim of annotation-cost reduction rests on an internally inconsistent budget, and none of the pipeline components is demonstrated on any data.

major comments (3)
  1. [§2.2 and §3.3] The annotation strategy in §2.2 begins with manual labeling of 'a few hundred high-quality images' for genus and species identification, while §3.3 specifies a target of 100–200 species-level classes 'with about 1000 labeled samples per class.' If these are the same labels, the seed set is three orders of magnitude smaller than the final budget (100,000–200,000 labels), requiring the teacher model to bootstrap almost the entire dataset; if the 'few hundred' is meant per class, then the total manual effort is tens of thousands of labels, which contradicts the paper's emphasis on reducing annotation cost. The paper does not provide a feasibility analysis for either reading, so the cost-reduction claim is not supported.
  2. [Abstract and Title] The title 'Building a Dataset' and parts of the abstract imply that a dataset has been or is being built, but no empirical evidence of any kind is presented: there is no pilot dataset, no evaluation of the teacher-student semi-supervised loop, no detection or classification accuracy, and no field data from the two mentioned national parks. Without at least a proof-of-concept on a small number of classes, the reader cannot assess whether the proposed pipeline is viable, and the paper's contribution is difficult to evaluate. The authors should either add a small-scale validation or explicitly reframe the paper as a project plan.
  3. [§4 Practical Considerations] The paper acknowledges that 'the system must be robust in diverse aquatic conditions, addressing variables such as lighting, turbidity, and background noise,' but provides no evidence that the proposed underwater cameras will produce usable images at the planned sites, which is a prerequisite for the entire pipeline. A small pilot deployment at Cat Tien or Cuc Phuong, reporting image quality and specimen visibility, would strengthen the proposal; without it, the feasibility of data collection is an untested assumption.
minor comments (5)
  1. [§1 (footnote on 'Suoi')] The phrase 'Suoi means streams in Vietnamese' would be more precise as 'Suối (Suoi) means stream in Vietnamese'; the plural form is not needed in the name.
  2. [§2.1] The production estimate of 'around three million data points per site annually' is asserted without derivation; the authors should specify the assumed frame rate, active hours, and image size.
  3. [§2.2] The sentence 'We will also look to include existing datasets of regional invertebrates to create the initial dataset' is vague; please name the datasets and describe how class overlap with new images will be handled.
  4. [§3.3] The target of '100-200 species-level classes' with 'about 1000 labeled samples per class' is presented without discussion of class imbalance or how rare species will be treated in the long-tail distribution; this is relevant to the evaluation plan.
  5. [Figure 1] Figure 1 is not referenced in the text; the reader is not told where in the pipeline to look at it, and the individual components in the diagram are not explained in a caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a forward-looking proposal with no fitted parameters, no predictions, and no derivation chain that could reduce to its inputs.

full rationale

This paper describes a proposed pipeline for collecting, annotating, and classifying aquatic invertebrates in Vietnam, and it explicitly reports no dataset, no experiments, and no results. There is no equation in which an output quantity is defined in terms of itself, no fitted parameter that is later relabeled as a prediction, and no uniqueness or existence theorem imported from prior work to force a choice. The only self-citation is the reference to Albini et al. (2024), which includes co-author Khuong Dinh; it is cited merely as related work on automated imaging of zooplankton and is not load-bearing for any claim in the proposal. The teacher-student bootstrap described in Section 2.2 is a standard semi-supervised learning mechanism, not a circular derivation: the teacher's pseudo-labels are not presented as independent ground truth, and the paper makes no claim that the pipeline's outputs are guaranteed by its own inputs. Likewise, the annotation budget figures in Sections 2.2 and 3.3 raise a feasibility question, but a quantitative inconsistency is not circularity under the required standard. Because there is no derivation or prediction whose truth could be presupposed, the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper rests on domain assumptions about field deployment and on the sufficiency of a few hundred manual labels to bootstrap a teacher model for fine-grained classification of aquatic invertebrates. These assumptions are untested: the paper reports no pilot data, no annotated images, and no experimental validation.

assumptions (3)
  • domain assumption Underwater cameras deployed in the proposed field sites will produce images usable for identification of 10-50mm invertebrates in natural conditions.
    This assumption is implicit in the data acquisition description in Section 2.1, where cameras are said to be deployed in rivers, lakes, and coastal areas, but no field test data are presented to support the feasibility.
  • domain assumption Manual labeling of a few hundred images by biologists will be sufficient to bootstrap a teacher model that can produce useful pseudo-labels at genus and species level.
    This is a core enabling assumption of the annotation strategy in Section 2.2. A few hundred images for 100 to 200 species would be very few per class, and no evidence is given that this is sufficient for the proposed semi-supervised bootstrap.
  • ad hoc to paper The proposed target counts of 1000 labeled samples per class and 20 to 50 genus classes are achievable within the project's operational scope.
    These targets in Section 3.3 are stated without a derivation or a collection cost estimate, and no evidence is provided that the field sites can supply that number of clear images.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SuoiAI: Building a Dataset for Aquatic Invertebrates in Vietnam." pith.science (2026). https://pith.science/paper/TSHT7325

@misc{pith2026250415252,
  author       = {Pith},
  title        = {Pith review of: SuoiAI: Building a Dataset for Aquatic Invertebrates in Vietnam},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSHT7325}},
  note         = {Machine review of arXiv:2504.15252}
}
read the original abstract

Understanding and monitoring aquatic biodiversity is critical for ecological health and conservation efforts. This paper proposes SuoiAI, an end-to-end pipeline for building a dataset of aquatic invertebrates in Vietnam and employing machine learning (ML) techniques for species classification. We outline the methods for data collection, annotation, and model training, focusing on reducing annotation effort through semi-supervised learning and leveraging state-of-the-art object detection and classification models. Our approach aims to overcome challenges such as data scarcity, fine-grained classification, and deployment in diverse environmental conditions.

Figures

Figures reproduced from arXiv: 2504.15252 by the authors.

Figure 1
Figure 1. An end-to-end diagram of our pipeline 3 Modeling Techniques 3.1 Baseline Object Detection We propose using conventional neural networks for object detection. Models like YOLO (Redmon et al. (2016)) or Faster R-CNN (Ren et al. (2016)) variants are older but small enough to be deployable on-device and therefore can serve as a reliable baseline. Models like SwinTransformer (Liu et al. (2021)) and DETR (Carion et al. (2… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 8 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Heimböck, Jan Heuschele, Janne E

    Dania Albini, Mathieu Lutier, Martin P. Heimböck, Jan Heuschele, Janne E. Søreide, Michelle C. Jackson, and Khuong V. Dinh. Temporal patterns in multiple stressors shape the vulnerability of overwintering arctic zooplankton. Ecology and Evolution, June 2024. doi:10.1002/ece3.11673. URL https://doi.org/10.1002/ece3.11673

    work page doi:10.1002/ece3.11673 2024

  3. [3]

    Classifying the unknown: Insect identification with deep hierarchical bayesian learning

    Sarkhan Badirli, Christine Johanna Picard, George Mohler, Frannie Richert, Zeynep Akata, and Murat Dundar. Classifying the unknown: Insect identification with deep hierarchical bayesian learning. Methods in Ecology and Evolution, 14 0 (6): 0 1515--1530, 2023. doi:https://doi.org/10.1111/2041-210X.14104. URL https://besjournals.onlinelibrary.wiley.com/doi/...

    work page doi:10.1111/2041-210x.14104 2023

  4. [4]

    Weiser, Kirsten de Beurs, Michael Kaspari, Cameron Siler, and Katie E

    Jarrett Blair, Michael D. Weiser, Kirsten de Beurs, Michael Kaspari, Cameron Siler, and Katie E. Marshall. Embracing imperfection: machine-assisted invertebrate classification in real-world datasets. bioRxiv, 2022. doi:10.1101/2021.09.13.460161. URL https://www.biorxiv.org/content/early/2022/02/07/2021.09.13.460161

    work page doi:10.1101/2021.09.13.460161 2022

  5. [5]

    End-to-end object detection with transformers, 2020

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. URL https://arxiv.org/abs/2005.12872

    arXiv 2020

  6. [6]

    Mixed pseudo labels for semi-supervised object detection, 2023

    Zeming Chen, Wenwei Zhang, Xinjiang Wang, Kai Chen, and Zhi Wang. Mixed pseudo labels for semi-supervised object detection, 2023. URL https://arxiv.org/abs/2312.07006

    arXiv 2023

  7. [7]

    Singh, Soumik Sarkar, Nirav Merchant, Arti Singh, Baskar Ganapathysubramanian, and Chinmay Hegde

    Benjamin Feuer, Ameya Joshi, Minsu Cho, Shivani Chiranjeevi, Zi Kang Deng, Aditya Balu, Asheesh K. Singh, Soumik Sarkar, Nirav Merchant, Arti Singh, Baskar Ganapathysubramanian, and Chinmay Hegde. Zero-shot insect detection via weak language supervision. Plant Phenome Journal, May 2024. doi:10.1002/ppj2.20107. URL https://doi.org/10.1002/ppj2.20107

    work page doi:10.1002/ppj2.20107 2024

  8. [8]

    An affordable and automated imaging approach to acquire highly resolved individual data—an example of copepod growth in response to multiple stressors

    Jan Heuschele, Torben Lode, Tom Andersen, Katrine Borgå, and Josefin Titelman. An affordable and automated imaging approach to acquire highly resolved individual data—an example of copepod growth in response to multiple stressors. PeerJ, 7: 0 e6776, 2019. doi:10.7717/peerj.6776. URL https://doi.org/10.7717/peerj.6776

    work page doi:10.7717/peerj.6776 2019

Show all 21 references
  1. [9]

    The inaturalist species classification and detection dataset, 2018

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset, 2018. URL https://arxiv.org/abs/1707.06642

    2018 arXiv

  2. [10]

    Swin transformer: Hierarchical vision transformer using shifted windows, 2021

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. URL https://arxiv.org/abs/2103.14030

    2021 arXiv

  3. [11]

    A metric-based few-shot learning method for fish species identification with limited samples

    Jiamin Lu, Song Zhang, Shili Zhao, Daoliang Li, and Ran Zhao. A metric-based few-shot learning method for fish species identification with limited samples. Animals, 14 0 (5): 0 755, 2024. doi:10.3390/ani14050755

    2024 doi

  4. [13]

    Pham, Peijie Chen, Tin Nguyen, Seunghyun Yoon, Trung Bui, and Anh Totti Nguyen

    Thang M. Pham, Peijie Chen, Tin Nguyen, Seunghyun Yoon, Trung Bui, and Anh Totti Nguyen. Peeb: Part-based image classifiers with an explainable and editable language bottleneck, 2024. URL https://arxiv.org/abs/2403.05297

    2024 arXiv

  5. [14]

    Sam 2: Segment anything in images and videos, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenh...

    2024 arXiv

  6. [15]

    You only look once: Unified, real-time object detection, 2016

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2016. URL https://arxiv.org/abs/1506.02640

    2016 arXiv

  7. [16]

    Faster r-cnn: Towards real-time object detection with region proposal networks, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. URL https://arxiv.org/abs/1506.01497

    2016 arXiv

  8. [17]

    Morphocluster: Efficient annotation of plankton images by clustering

    Simon-Martin Schröder, Rainer Kiko, and Reinhard Koch. Morphocluster: Efficient annotation of plankton images by clustering. Sensors, 20 0 (11), 2020. ISSN 1424-8220. doi:10.3390/s20113060. URL https://www.mdpi.com/1424-8220/20/11/3060

    2020 doi

  9. [18]

    Thuaire, Y

    B. Thuaire, Y. Allanic, A. Hoang Viet, Q. Le Khac, T. Luu Hong, C. Nguyen The, and T. Nguyen Thi. Assessing the biodiversity in viet nam – analysis of the impacts from the economic sectors, 2021

    2021

  10. [19]

    Automatic underwater fish species classification with limited data using few-shot learning

    Sébastien Villon, Corina Iovan, Morgan Mangeas, Thomas Claverie, David Mouillot, Sébastien Villéger, and Laurent Vigliola. Automatic underwater fish species classification with limited data using few-shot learning. Ecological Informatics, 63: 0 101320, 2021. ISSN 1574-9541. do...

    2021

  11. [20]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  12. [21]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  13. [22]

    how to learn a new classification problem with only few images

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2021 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.