Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Bridging Domain Gaps for Fine-Grained Moth Classification Through Expert-Informed Adaptation and Foundation Model Priors

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A 28M-parameter model matches a 304M one on fine-grained moth classification by distilling foundation-model knowledge and adding limited field data.

desk verdict Useful empirical recipe for moth classification, but the headline claim rests on an unvalidated train/test split and single-run numbers. read the letter →

arxiv 2508.20089 v1 pith:LRERJBOC submitted 2025-08-27 cs.CV

Ross J Gardiner , Guillaume Mougeot , Sareh Rowlands , Benno I Simmons , Flemming Helsing , Toke Thomas H{o}ye This is my paper · ORCID
classification cs.CV
keywords mothclassificationdomainshiftknowledgedistillationfoundationmodelsBioCLIP2lightweightCNNinsectcameratrapsfine-grained
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

The paper asks whether a small, deployable neural network can match a huge foundation model at identifying moth species from noisy camera-trap images. It answers yes: by transferring BioCLIP2's feature representations into a ConvNeXt-tiny student through a mean-squared-error hint loss, and mixing in a little expert-labelled field imagery, the 28-million-parameter student reaches 89.4% top-1 accuracy with 50% target-domain training data. That matches BioCLIP2's 88.3% at zero target supervision, a model with 304 million parameters. The result suggests that ecologists with limited computing can get near-foundation-model accuracy by combining knowledge distillation with a small amount of expert annotation. The paper also documents that BioCLIP2 alone is the strongest zero-shot performer on this domain-shifted task.

What carries the argument

The load-bearing mechanism is feature-based knowledge distillation with a hint loss: the student ConvNeXt is trained to match the final embedding of the frozen BioCLIP2 teacher with a mean-squared-error term, weighted equally with the classification cross-entropy loss. This transfers broad, domain-robust feature priors into a small architecture. The second mechanism is train-time domain mixing: GBIF source images are augmented with a MixRes strategy that downscales some images, mimicking the variable quality and resolution of field crops, and the fraction of expert-labelled AMI images in the training set is varied from 0% to 50%. A third, less visible mechanism is the train/test split: an un

What would settle it

Compare the current split with a manually curated, individual-aware split for a subset of the 101 species: compute the nearest-neighbour embedding distance between each training image and the closest test image. If many near-duplicates of the same individual appear in both splits, or if training on the alternative split materially lowers the distilled student's accuracy, the claimed match to BioCLIP2 would not hold.

Watch

Extended reading notes

Core claim

The central claim is that knowledge distillation from a large, broadly pretrained vision encoder (BioCLIP2, 304M parameters) into a lightweight ConvNeXt-tiny network (28M parameters), combined with progressive mixing of expert-labelled target-domain images into training, closes most of the domain-shift gap between curated source images (GBIF) and in-situ camera-trap images (AMI). With only 50% target data in the training mix, the distilled student obtains 89.4% top-1 accuracy on the target test set, essentially matching BioCLIP2's 88.3% with no target supervision at all. The paper further shows that modest target supervision helps all models, and that distillation provides a consistent boost

Load-bearing premise

The reported accuracies rest on the assumption that the unsupervised clustering of ResNet-50 embeddings fully separates near-duplicate images of the same moth individual between the training and test sets; if any repeated individual leaks across the split, every number in Table 1 would be inflated.

Editorial extensions

If this is right

  • A 28M-parameter student can match a 304M teacher with half target supervision, making lightweight models viable for on-device moth classification in camera traps and reducing the need for cloud inference.
  • The clear decision rule: use BioCLIP2 when computing is plentiful and target data scarce; use distillation plus domain mixing when lightweight models are required.
  • The consistent +3.6% accuracy boost from distillation across all target-mix levels indicates that foundation-model priors and limited expert labels are complementary rather than redundant.
  • Even 5% target supervision yields large gains (13–15 points for ConvNeXt variants), suggesting a small amount of expert labelling can go a long way in domain-shifted insect classification.

Reading between the lines

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

  • The same distillation-plus-mixing recipe may transfer to other fine-grained taxa monitored by camera traps (e.g., bees, beetles), where expert labels are scarce and GBIF images are similarly domain-shifted; this could be tested by repeating the protocol on a different insect family.
  • The paper does not test whether the student's advantage persists with a species set well beyond 101 classes or under long-tailed class distributions; a natural extension is to measure accuracy-per-parameter under class imbalance.
  • Because the train/test split relies on ImageNet-pretrained ResNet-50 embeddings, the measured accuracies could be inflated if clustering fails to fully separate repeated individuals; a direct check would compare the split-based results against a manual, individual-aware split on a subset of species.
  • The distillation layer and alpha weighting are fixed at 0.5; a sensitivity study might show tuning these narrows the gap to BioCLIP2 further.
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

3 major / 4 minor

Summary. The paper addresses fine-grained moth classification from automated camera-trap images under domain shift between curated GBIF images and noisy AMI field imagery. It proposes training a lightweight ConvNeXt-tiny student with knowledge distillation (feature-based 'hint' loss) from the BioCLIP2 foundation model, while progressively mixing expert-labelled target-domain AMI images into the training set. The authors benchmark ConvNeXt-tiny, ConvNeXt-tiny+KD, BioCLIP, and BioCLIP2 at target-domain mixes of 0%, 1%, 5%, 10%, 20%, 25%, 33%, and 50% and report top-1 accuracies on held-out target and source test sets. The central claim is that ConvNeXt-tiny+KD at 50% target supervision (89.4%) matches BioCLIP2 without target supervision (88.3%), despite having roughly 10 times fewer parameters, and that knowledge distillation provides a consistent accuracy boost across all mix levels.

Significance. If the results hold, the paper offers a practical recipe for lightweight, edge-deployable insect classifiers that retain much of the accuracy of much larger foundation models. The study is useful in comparing BioCLIP and BioCLIP2 on a real, domain-shifted ecological dataset and in quantifying how little target-domain supervision is needed to substantially improve a small model. The paper is clearly written and the experimental design is straightforward to reproduce in principle. However, the manuscript provides no code or data release, no repeated-seed experiments, and no validation of the clustering-based train/test split. These omissions are consequential because the central numerical claim rests on a one-point difference from a single run of a small student model and a teacher that never sees target-domain training images.

major comments (3)
  1. [Section 4, Table 1] The load-bearing train/test split is not validated. The AMI split uses agglomerative clustering of ImageNet ResNet-50 embeddings with K = max(5, sqrt(N)) ~ 10 clusters for N=110 images, then collects the 10-image test set by walking through shuffled clusters. If clusters are impure, or if one cluster straddles the train/test cutoff, near-duplicate time-lapse frames of the same individual can appear in both training and test sets. This is asymmetric: BioCLIP2 at 0% target supervision only trains on GBIF images and is unaffected, whereas ConvNeXt+KD at 50% target supervision trains on 10,100 AMI images that may contain near-duplicates of test frames. The paper's central claim in Section 5 (89.4% vs 88.3%) could therefore be inflated by memorized repetition rather than genuine domain generalization. Please report cluster purity, the distribution of test images across clusters, and a tempora
  2. [Section 4, Table 1] All reported numbers come from a single training run with no seeds, error bars, or statistical tests. The comparisons that drive the conclusions are small: the KD boost is averaged at +3.6%, and the headline 'matched BioCLIP2' is a 1.1-point difference (89.4 vs 88.3). Without run-to-run variance, it is impossible to know whether these differences are meaningful or noise. Please provide results over at least three seeds with mean/std or confidence intervals, and state whether the qualitative trends (KD helps; BioCLIP2 is best) are stable across seeds.
  3. [Section 3.2] The fixed training budget of 10 epochs is not justified, and no convergence curves are shown. Since all models use the same schedule, the relative ranking may be fair, but the absolute claim that ConvNeXt+KD 'matched' BioCLIP2 could change if some models underfit or if a longer schedule benefits one architecture more than another. Please report learning curves or verify that 10 epochs is sufficient for each model, especially for the frozen-encoder linear probes.
minor comments (4)
  1. [Section 3.1] There is a grammar issue: 'training smaller student model is to match' should be 'training a smaller student model to match the features of a larger pre-trained teacher'.
  2. [Section 3.1, Eq. (2)] The knowledge distillation weight alpha is hand-set to 0.5 with no sensitivity analysis. A short grid or discussion of alpha's effect would strengthen the claim that the KD improvement is robust.
  3. [Section 2] The source/test split is described as random, but no seed is given. Please specify the random seed or state that the split is available for reproducibility.
  4. [Supplementary Section 8] The phrase 'this selective part is illustrated in Figure 4' is unclear; the figure appears to show a toy example of cluster shuffling. Clarify the figure caption and how the described procedure maps to the actual dataset.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: central results are empirical evaluations against held-out test sets; only minor non-load-bearing self-citation.

full rationale

The paper's central claims—BioCLIP2's strong target-domain accuracy and the distilled ConvNeXt+KD model's comparable accuracy—are empirical measurements on held-out AMI and GBIF test sets that are not used to fit the teacher or to define the student's objective. The KD student is trained to match BioCLIP2's frozen feature embeddings, but BioCLIP2's own test accuracy is never an input to the student's loss; the student's accuracy is an independent outcome. The self-citations to [14] (AMI dataset and MixRes augmentation) are data/method choices, not theoretical premises, and they do not make the evaluation circular. The unsupervised clustering split described in Supplementary Section 8 is a potential data-leakage risk, but it is not a circularity: the split does not define the measured accuracy in terms of any fitted parameter or known result. The hyperparameter alpha=0.5 in Equation (2) is a free choice, not a fitted value, and the comparison at 50% target supervision versus 0% target supervision is a comparison across independent training configurations, not a derivation. Thus no 'prediction' reduces by construction to an input, and no load-bearing step is justified only by self-citation. Score 2 reflects one minor self-citation that is not load-bearing.

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

The paper is an empirical study, so the ledger consists of hand-chosen hyperparameters and unstated domain assumptions rather than invented entities. The KD weight alpha, the clustering parameters, and the fixed training schedule are load-bearing for the reported numbers. Validation of the split and the teacher choice would strengthen the central claims.

free parameters (5)
  • KD loss weight alpha = 0.5
    Set in Eq. (2) without ablation; determines the balance between classification and hint loss, so the reported KD benefit is conditional on this choice.
  • Minimum cluster count Kmin = 5
    Used in the train/test split clustering (Supplementary Section 8); affects how well repeated images are separated and thus the reported accuracies.
  • Train/test split test size = 10 images per class
    Withheld from the 110 per-class AMI images (Section 2); smaller than 10 would increase variance, larger would reduce training data.
  • Training epochs = 10
    All models trained for exactly 10 epochs (Section 4); no convergence check is reported.
  • Learning rate = 1e-3
    From Section 3.2; chosen without a reported sweep, affects final accuracy.
assumptions (5)
  • domain assumption Expert labels of AMI images are correct to species level.
    Section 2 states moths were labelled by an expert taxonomist; if labels are noisy, supervision is misdirected.
  • domain assumption GBIF images are a valid source-domain proxy for curated moth images.
    Section 2 downloads 224 unique occurrence images per species, filtered to imago; this assumes GBIF images represent the source-domain distribution.
  • domain assumption ImageNet-pretrained ResNet-50 embeddings cluster images such that repeated individuals are separated.
    Supplementary Section 8 relies on this to build leakage-free splits; no evaluation of cluster purity is provided.
  • domain assumption BioCLIP2's frozen embeddings are a useful target for feature distillation on the target domain.
    Section 3.1 defines the hint loss against BioCLIP2; this assumes the teacher's representations are meaningful for AMI moth crops.
  • ad hoc to paper Ten training epochs are sufficient for all compared models.
    Section 4 fixes 10 epochs for all models; no learning curves or convergence analysis is shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Domain Gaps for Fine-Grained Moth Classification Through Expert-Informed Adaptation and Foundation Model Priors." pith.science (2026). https://pith.science/paper/LRERJBOC

@misc{pith2026250820089,
  author       = {Pith},
  title        = {Pith review of: Bridging Domain Gaps for Fine-Grained Moth Classification Through Expert-Informed Adaptation and Foundation Model Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LRERJBOC}},
  note         = {Machine review of arXiv:2508.20089}
}
read the original abstract

Labelling images of Lepidoptera (moths) from automated camera systems is vital for understanding insect declines. However, accurate species identification is challenging due to domain shifts between curated images and noisy field imagery. We propose a lightweight classification approach, combining limited expert-labelled field data with knowledge distillation from the high-performance BioCLIP2 foundation model into a ConvNeXt-tiny architecture. Experiments on 101 Danish moth species from AMI camera systems demonstrate that BioCLIP2 substantially outperforms other methods and that our distilled lightweight model achieves comparable accuracy with significantly reduced computational cost. These insights offer practical guidelines for the development of efficient insect monitoring systems and bridging domain gaps for fine-grained classification.

Figures

Figures reproduced from arXiv: 2508.20089 by the authors.

Figure 1
Figure 1. Data processing pipeline showing how moth images from AMI (target-domain) and GBIF (source-domain) are curated, labelled [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Foundation vision encoders (top) with a trainable classi [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Top-1 target-domain accuracy plotted over training [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: List of image embeddings with each showing its clus [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

  1. [1]

    Gbif home page, 2025. 1

    work page 2025

  2. [2]

    Bar-On, Rob Phillips, and Ron Milo

    Yinon M. Bar-On, Rob Phillips, and Ron Milo. The biomass distribution on earth. Proceedings of the National Academy of Sciences, 115(25):6506–6511, 2018. 1

    work page 2018

  3. [3]

    Kim Bjerge, Henrik Karstoft, and Toke T. Høye. Towards edge processing of images from insect camera traps. Remote Sensing in Ecology and Conservation. 1, 2

    work page

  4. [4]

    An automated light trap to monitor moths (lepidoptera) using computer vision-based tracking and deep learning

    Kim Bjerge, Jakob Bonde Nielsen, Martin Videbæk Sep- strup, Flemming Helsing-Nielsen, and Toke Thomas Høye. An automated light trap to monitor moths (lepidoptera) using computer vision-based tracking and deep learning. Sensors, 21(2), 2021. 1

    work page 2021

  5. [5]

    Kim Bjerge, Jamie Alison, Mads Dyrmann, Carsten Eie Frigaard, Hjalte M. R. Mann, and Toke Thomas Høye. Accu- rate detection and identification of insects from camera trap images with deep learning. PLOS Sustainability and Trans- formation, 2(3):1–18, 2023. 1

    work page 2023

  6. [6]

    Ob- ject detection of small insects in time-lapse camera record- ings

    Kim Bjerge, Carsten Eie Frigaard, and Henrik Karstoft. Ob- ject detection of small insects in time-lapse camera record- ings. Sensors, 23(16), 2023. 1

    work page 2023

  7. [7]

    Mann, Toke T

    Kim Bjerge, Quentin Geissmann, Jamie Alison, Hjalte M.R. Mann, Toke T. Høye, Mads Dyrmann, and Henrik Karstoft. Hierarchical classification of insects with multitask learning and anomaly detection. Ecological Informatics, 77:102278,

    work page

  8. [8]

    Kevin F. A. Darras, Marcel Balle, Wenxiu Xu, Yang Yan, Vincent G. Zakka, Manuel Toledo-Hern´andez, Dong Sheng, Wei Lin, Boyu Zhang, Zhenzhong Lan, Li Fupeng, and Thomas C. Wanger. Eyes on nature: Embedded vision cam- eras for terrestrial biodiversity monitoring. Methods in Ecol- ogy and Evolution, 15(12):2262–2275, 2024. 1

    work page 2024

Show all 37 references
  1. [9]

    The taxo- nomic impediment: a shortage of taxonomists, not the lack of technical approaches

    Michael S Engel, Luis M P Cer ´ıaco, Gimo M Daniel, Pablo M Dellap ´e, Ivan L ¨obl, Milen Marinov, Roberto E Reis, Mark T Young, Alain Dubois, Ishan Agarwal, Pablo Lehmann A., Mabel Alvarado, Nadir Alvarez, Franco An- dreone, Katyuscia Araujo-Vieira, John S Ascher, D ´elio Baˆ...

    2021

  2. [10]

    Gardiner, Sareh Rowlands, and Benno I

    Ross J. Gardiner, Sareh Rowlands, and Benno I. Simmons. Towards scalable insect monitoring: Ultra-lightweight cnns as on-device triggers for insect camera traps. Methods in Ecology and Evolution. 1

  3. [11]

    Wang, Xiaoliang Huo, Joakim B

    Zeming Gong, Austin T. Wang, Xiaoliang Huo, Joakim B. Haurum, Scott C. Lowe, Graham W. Taylor, and Angel X. Chang. Clibd: Bridging vision and genomics for biodiversity monitoring at scale. arXiv preprint, 2025. 2

    2025

  4. [12]

    Campolongo, Matthew J

    Jianyang Gu, Samuel Stevens, Elizabeth G. Campolongo, Matthew J. Thompson, Net Zhang, Jiaman Wu, et al. Bioclip 2: Emergent properties from scaling hierarchical contrastive learning. arXiv preprint, 2025. 2

    2025

  5. [13]

    Høye, Johanna ¨Arje, Kim Bjerge, Oskar L

    Toke T. Høye, Johanna ¨Arje, Kim Bjerge, Oskar L. P. Hansen, Alexandros Iosifidis, Florian Leese, Hjalte M. R. Mann, Kristian Meissner, Claus Melvad, and Jenni Raito- harju. Deep learning and computer vision will transform en- tomology. Proceedings of the National Academy of S...

    2021

  6. [14]

    Insect identification in the wild: The ami dataset

    Aditya Jain, Fagner Cunha, Michael James Bunsen, Juan Se- basti´an Ca˜nas, L´eonard Pasi, Nathan Pinoy, Flemming Hels- ing, JoAnne Russo, Marc Botham, Michael Sabourin, et al. Insect identification in the wild: The ami dataset. In European Conference on Computer Vision , pages...

    2024

  7. [15]

    A method to establish a synthetic image dataset of stored-product insects for insect detection

    Jiangtao Li, Yuwei Su, Zhaojun Cui, Jida Tian, and Huiling Zhou. A method to establish a synthetic image dataset of stored-product insects for insect detection. IEEE Access, 10: 70269–70278, 2022. 1

    2022

  8. [16]

    Data-free knowledge transfer: A survey

    Yuang Liu, Wei Zhang, Jun Wang, and Jianyong Wang. Data-free knowledge transfer: A survey. CoRR, abs/2112.15278, 2021. 2

    2021 arXiv

  9. [17]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 11976–11986,

  10. [18]

    Losey and Mace Vaughan

    John E. Losey and Mace Vaughan. The economic value of ecological services provided by insects. BioScience, 56(4): 311–323, 2006. 1

    2006

  11. [19]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3

    2017 arXiv

  12. [20]

    Montgomery, Michael W

    Graham A. Montgomery, Michael W. Belitz, Rob P. Gural- nick, and Morgan W. Tingley. Standards and best practices for monitoring and benchmarking insects. Frontiers in Ecol- ogy and Evolution, 8, 2021. 1, 2

    2021

  13. [21]

    A survey on knowledge distillation: Recent advancements

    Amir Moslemi, Anna Briskina, Zubeka Dang, and Jason Li. A survey on knowledge distillation: Recent advancements. Machine Learning with Applications, 18:100605, 2024. 2, 3

    2024

  14. [22]

    Insect- foundation: A foundation model and large-scale 1m dataset for visual insect understanding

    Hoang-Quan Nguyen, Thanh-Dat Truong, Xuan Bac Nguyen, Ashley Dowling, Xin Li, and Khoa Luu. Insect- foundation: A foundation model and large-scale 1m dataset for visual insect understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (...

    2024

  15. [23]

    Kalkman, and Dan Stowell

    Rita Pucci, Vincent J. Kalkman, and Dan Stowell. Perfor- mance of computer vision algorithms for fine-grained clas- sification using crowdsourced insect images. IET Computer Vision, 19(1):e70006, 2025. 1

    2025

  16. [24]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  17. [25]

    Motion vectors and deep neural networks for video camera traps

    Miklas Riechmann, Ross Gardiner, Kai Waddington, Ryan Rueger, Frederic Fol Leymarie, and Stefan Rueger. Motion vectors and deep neural networks for video camera traps. Ecological Informatics, 69:101657, 2022. 1

    2022

  18. [26]

    D. B. Roy, J. Alison, T. A. August, M. B´elisle, K. Bjerge, J. J. Bowden, M. J. Bunsen, F. Cunha, Q. Geissmann, K. Gold- mann, A. Gomez-Segura, A. Jain, C. Huijbers, M. Larriv ´ee, J. L. Lawson, H. M. Mann, M. J. Mazerolle, K. P. McFarland, L. Pasi, S. Peters, N. Pinoy, D. Rol...

    1904

  19. [27]

    Insect detect: An open-source diy camera trap for automated insect monitoring

    Maximilian Sittinger, Johannes Uhler, Maximilian Pink, and Annette Herz. Insect detect: An open-source diy camera trap for automated insect monitoring. PLOS ONE, 19(4):1–28,

  20. [28]

    Thompson, Eliza- beth G

    Samuel Stevens, Jiaman Wu, Matthew J. Thompson, Eliza- beth G. Campolongo, Chan Hee Song, David E. Carlyn, et al. Bioclip: A vision foundation model for the tree of life. arXiv preprint, 2023. 2

    2023

  21. [29]

    Nigel E. Stork. How many species of insects and other ter- restrial arthropods are there on earth? Annual Review of Entomology, 63(V olume 63, 2018):31–45, 2018. 1

    2018

  22. [30]

    The inaturalist species classification and de- tection dataset

    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 de- tection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages...

  23. [31]

    Bowler, Konstantin B

    Roel van Klink, Diana E. Bowler, Konstantin B. Gongal- sky, Ann B. Swengel, Alessandro Gentile, and Jonathan M. Chase. Meta-analysis reveals declines in terrestrial but in- creases in freshwater insect abundances.Science, 368(6489): 417–420, 2020. 1

    2020

  24. [32]

    Towards zero-shot camera trap image categorization

    Ji ˇr´ı Vysko ˇcil and Lukas Picek. Towards zero-shot camera trap image categorization. In European Conference on Com- puter Vision, pages 37–53. Springer, 2025. 2

    2025

  25. [33]

    Wagner, Eliza M

    David L. Wagner, Eliza M. Grames, Matthew L. Forister, May R. Berenbaum, and David Stopak. Insect decline in the anthropocene: Death by a thousand cuts. Proceedings of the National Academy of Sciences , 118(2):e2023989118,

  26. [34]

    Deep visual domain adapta- tion: A survey

    Mei Wang and Weihong Deng. Deep visual domain adapta- tion: A survey. Neurocomputing, 312:135–153, 2018. 1

    2018

  27. [35]

    Deng, Andre Nakkab, Md Zahid Hasan, et al

    Chih-Hsuan Yang, Benjamin Feuer, Zaki Jubery, Zi K. Deng, Andre Nakkab, Md Zahid Hasan, et al. Biotrove: A large curated image dataset enabling ai for biodiversity. arXiv preprint, 2024. 2 Bridging Domain Gaps for Fine-Grained Moth Classification Through Expert-Informed Adapta...

    2024

  28. [36]

    Embedding-Based Clustering for Train/Test Splitting To obtain train/test splits from AMI data that are decou- pled, we first embed each AMI image using a ResNet-50 network pretrained on ImageNet-1K; this places visually similar images near one another in the embedding space. F...

  29. [37]

    Statistics of each domain-mixed training dataset at var- ious amounts of target-domain percentages

    Dataset Composition Target-domain Mix (%) Target-domain (AMI) Contribution Source-domain (GBIF) Contribution Total Dataset Size 0% 0 18573 18573 1% 187 18573 18760 5% 997 18573 19550 10% 2063 18573 20636 20% 4643 18573 23216 25% 6191 18573 24764 33% 9147 18573 27720 50% 10100 ...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.