SatDINO: A Deep Dive into Self-Supervised Pretraining for Remote Sensing

Ivan Gruber; Jakub Straka

REVIEW 4 major objections 5 minor 28 references

SatDINO: A Deep Dive into Self-Supervised Pretraining for Remote Sensing

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SatDINO shows that DINO-style contrastive pretraining beats masked autoencoders for remote sensing.

desk verdict A useful, mostly honest DINO-vs-MAE comparison for remote sensing, but the headline kNN gains may be an artifact of how MAE features were extracted. read the letter →

arxiv 2508.21402 v1 pith:JGEEDURM submitted 2025-08-29 cs.CV cs.LG

Jakub Straka , Ivan Gruber This is my paper · ORCID

classification cs.CVcs.LG

keywords self-supervisedlearningremotesensingDINOmaskedautoencodersgroundsampledistancemulti-scalerepresentationcontrastivesatelliteimagery

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 sets out to show that DINO, a contrastive self-supervised method that aligns local and global views of an image, is a better default than masked autoencoders (MAE) for pretraining on remote sensing imagery. The authors introduce SatDINO, which adds two components to DINO: uniform view sampling, which exposes the model to a wide range of ground-sample distances during training, and a GSD-encoding token with a regression head that lets the model infer resolution from pixels. Across six classification datasets, kNN features from SatDINO exceed those of the MAE-based Scale-MAE and SatMAE by up to nine percent, and linear probing on fMoW-RGB is also higher—all with a ViT-Small backbone instead of the baselines' ViT-Large. The implication, if correct, is that scale-aware contrastive alignment is more valuable than reconstruction for multi-scale satellite data.

What carries the argument

The mechanism is DINO's global-local view alignment: a teacher and student network are trained to output similar distributions for global views (large crops covering 25–100% of the image) and local views (small crops covering 5–25%). Because crops are resized to fixed sizes, this alignment forces the model to recognize the same scene across different ground sample distances. SatDINO strengthens this scale awareness with two additions. Uniform view sampling divides the local-crop scale range into subranges and draws one view from each, so a spectrum of GSDs is represented in every batch. GSD encoding adds a randomly initialized token (alongside the class token) and a linear regression head th

What would settle it

Train a ViT-Small masked-autoencoder model on the same fMoW-RGB split and for the same number of epochs as SatDINO, then evaluate both with the identical kNN protocol on the six classification datasets. If the MAE model matches or exceeds SatDINO's average accuracy, the claim that DINO yields stronger features in this remote-sensing setting is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that DINO pretraining learns stronger and more robust features than MAE pretraining for multi-scale remote sensing, despite MAE being the more common choice. Evidence: kNN classification with frozen features across multiple scales on six datasets, where SatDINO beats Scale-MAE on every dataset and SatMAE on nearly all, and linear probing on fMoW-RGB, where SatDINO's top-1 accuracy (72.75–73.53) surpasses every reported MAE baseline. SatDINO also achieves competitive fine-tuning and segmentation results with a much smaller backbone. The authors attribute the advantage to DINO's multi-scale view alignment, which matches satellite data structure, a

Load-bearing premise

The paper's comparison assumes that the evaluation protocol—its fixed dataset splits, its re-evaluation of SatMAE and Scale-MAE, and the exclusion of Cross-Scale MAE—is neutral, so the measured lead reflects the pretraining objective rather than artifacts of the setup.

Editorial extensions

If this is right

If DINO's scale-alignment objective is the reason for the gains, remote-sensing pretraining should shift away from MAE as the default and adopt contrastive global-local alignment.
SatDINO's small size and strong frozen-feature performance make it a practical embedded-feature extractor for image matching, retrieval, and other tasks that require compressing images to vectors.
The GSD-encoding token lets downstream models use resolution awareness even when ground-sample-distance metadata is missing, since the model learns to estimate it from pixels.
Uniform view sampling is a training-time-only modification with no inference cost, so it can be combined with other pretraining objectives or augmentations.
The paper observes signs of overfitting when scaling DINO to larger backbones, suggesting that DINO's data appetite grows with size and that larger remote-sensing pretraining corpora would be needed to match the MAE baselines at equal scale.

Reading between the lines

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

The comparison is not controlled for backbone size: SatDINO uses ViT-Small or ViT-Base while the main baselines are ViT-Large. If the MAE baselines were re-trained with ViT-Small and the same compute, the reported gap might shrink, so an equal-capacity comparison would sharpen the claim.
The authors re-evaluated SatMAE and Scale-MAE with their own fixed dataset splits and could not reproduce Cross-Scale MAE at all. An independent evaluation on the official splits of each baseline would test whether the protocol itself is driving the advantage.
The uniform-sampling idea could be tested in isolation: applying it to a masked-autoencoder objective would reveal whether the improvement is caused by GSD diversity or by DINO's contrastive loss specifically.
The GSD regression head is a scalable pretext task that could be ported to other contrastive frameworks (e.g., CLIP-style or MoCo-style) or to dense prediction tasks such as segmentation, where scale awareness matters.

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.

Desk Editor's Note

read the letter

The paper does one genuinely useful thing: it runs a careful, large-scale comparison of DINO pretraining against MAE-style methods on fMoW, and it ships code and checkpoints. That alone makes it worth reading for anyone working on frozen-feature remote sensing embeddings. The two proposed modifications—uniform sampling of local views to cover GSD diversity and a small GSD-prediction token—are modest, and the ablations suggest real but marginal gains. The ablation study is more honest than most: it tests augmentation strength, view counts, sampling strategies, and reports downstream numbers, and it openly notes that temporal augmentations from [8] don't help on downstream tasks.

The main strength is the evidence that DINO's frozen features beat MAE's on kNN and linear probing across several datasets and scales. But the soft spots are in exactly the places you'd worry about. First, there are no error bars anywhere; single runs for everything, which is normal for this kind of paper but shouldn't be overstated. Second, the linear-probing comparison pairs ViT-Small DINO against ViT-Large MAE baselines; if DINO wins with a quarter of the parameters, that is interesting, but it's not an apples-to-apples test of the pretraining objective. Fine-tuning and segmentation show parity, not superiority, so the strong claim rests entirely on frozen-feature evals.

The stress-test about the class token is the most serious point. The paper says features are extracted using the class token "for all models." MAE pretraining doesn't produce a class token; SatMAE and Scale-MAE typically use mean pooling over patch tokens for linear probing. If the released checkpoints were loaded into a ViT with a randomly initialized class token and that token was used for kNN, the baseline numbers would be near random. The paper doesn't report mean-pooled numbers or say how MAE features were extracted. This is a testable, fixable issue, but it directly affects the headline "9% improvement" claim. I'd want that resolved before using these numbers as the basis for recommending DINO over MAE.

Also minor: the GSD loss weight γ=0.1 was chosen partly based on downstream classification results (Table 8a), which is selection on the evaluation metric. Not disqualifying, but it means the reported combined gains are optimistic.

Overall: this is a genuine, reproducible empirical study of a real question. The central claim is plausible and the code release makes it verifiable. I'd send it to review, but I'd ask the authors to add error bars or multiple runs, fix the MAE feature-extraction question, and ideally include a ViT-Small MAE baseline trained on the same data. The paper is for practitioners who want to choose a pretraining objective for remote sensing embeddings; they should read it, but treat the magnitude of the improvement skeptically until the baseline extraction is clarified.

Recommendation: send to peer review, ask for revisions.

Referee Report

4 major / 5 minor

Summary. The paper presents SatDINO, a DINO-based self-supervised pretraining method adapted for remote sensing imagery. The proposed modifications include uniform sampling of local views to cover a wider range of ground sample distances, and a GSD-prediction auxiliary task with a dedicated token and loss weight gamma. SatDINO is pretrained on fMoW-RGB and evaluated with kNN, linear probing, fine-tuning, and semantic segmentation across multiple classification and segmentation datasets. The central claim is that SatDINO learns stronger and more robust features than MAE-based methods such as SatMAE and Scale-MAE, based on consistently higher kNN accuracy and competitive linear-probing and fine-tuning results. The paper also reports an extensive ablation study of augmentations, number of views, view scale ranges, and GSD encoding.

Significance. If the central comparison is valid, the results would be a useful data point for the remote sensing community, suggesting that contrastive DINO-style pretraining can outperform the more commonly used masked-autoencoder pretraining on multi-scale satellite imagery. The paper covers many datasets, provides a broad ablation study, and releases code and checkpoints, which are practical strengths. However, the support for the central claim depends on a fair and technically correct comparison protocol, and several load-bearing issues in the evaluation prevent the current version from fully establishing the claimed superiority.

major comments (4)

[Sec. 4.1, Table 2, Figure 3] The kNN comparison is the primary evidence for the central claim. The text states that features are extracted using the class token for all models. MAE-based pretraining (SatMAE, Scale-MAE) does not optimize a class token; loading such checkpoints into a ViT with a randomly initialized [CLS] token would make the kNN features for the baselines essentially random for that token. The paper must state whether mean-pooled patch embeddings were used for the MAE checkpoints; if they were not, the kNN results in Table 2 and Figure 3 are not a valid comparison of learned representations and the claimed improvements of up to 9% would be an artifact of feature extraction. This is load-bearing and needs to be corrected or the conclusion revised.
[Sec. 5.3, Table 8, Sec. 3.2] The GSD loss weight gamma=0.1 was selected based on performance on downstream evaluation datasets (RESISC45, WHU-RS19, SIRI-WHU) and on fMoW, and Section 3.2 states the final SatDINO configuration was selected as the best balance across evaluation tasks. Because the same downstream datasets are then used to compare SatDINO against MAE baselines, the comparison is not neutral: hyperparameters have been tuned on the test data. This selection bias affects the magnitude and credibility of the reported improvements. Please use a held-out validation set for model/hyperparameter selection, or re-evaluate the baselines under the same selection procedure.
[Sec. 4.2, Table 3] Table 3 mixes results from different sources, protocols, and backbone sizes, and the note concedes that the linear-probing and fine-tuning settings may differ between models. Consequently, the statement that SatDINO outperforms all others in linear-probing Top-1 is not established by a controlled comparison. A fair comparison would require re-evaluating all models with the same linear-probe and fine-tuning protocol on the same splits, ideally with matched backbone sizes or with explicit controls for capacity. Without this, the linear-probing claim should be weakened or removed.
[Tables 6-9] All ablations appear to be based on a single training run, with no error bars or repeated seeds. Table 9 reports improvements as small as 0.1-0.4 percentage points (e.g., uniform scale + GSD encoding gives +0.31 kNN, +0.21 RESISC45), which are within typical run-to-run variation. The component-level conclusions about uniform sampling, GSD encoding, and their combination are therefore not supported. Please provide multiple seeds or confidence intervals for at least the key configurations, including the final SatDINO model.

minor comments (5)

[Sec. 5, Evaluation setup] Typos: 'Liner probing' should be 'Linear probing', 'learning rare' should be 'learning rate'. Also 'UperNet' in Section 4.3 should be 'UPerNet'.
[Table 5] Use decimal point consistently: entries such as '64,81' use a comma; other tables use a period.
[Sec. 4.1, Table 2] The text says 'some of the classification datasets' are resized to multiple scales, but Table 2/Figure 3 report averages across scales for all six datasets. Please clarify which datasets were evaluated at all four scales.
[References] References [9] and [17] are the same fMoW paper; consolidate to avoid duplication.
[Table 6a] The caption should define the augmentation intensity levels 'none', 'soft', and 'mid' (e.g., scaling probabilities by 0, 0.25, 0.75) for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SatDINO is an empirical comparison against externally trained MAE checkpoints, with no prediction that reduces by construction to a fitted input or self-citation.

full rationale

The paper's central claim—that DINO-style pretraining yields stronger features than MAE-based pretraining—is an empirical result. SatDINO is trained from scratch on fMoW-RGB and compared with publicly released SatMAE and Scale-MAE checkpoints using fixed evaluation splits. No equation in the paper defines a predicted quantity in terms of a fitted parameter; the GSD-encoding token and uniform view sampling are trained components evaluated on held-out data, not fit-then-renamed-as-prediction. The ablations do select hyperparameters (e.g., γ=0.1) using validation performance, which is a model-selection bias rather than circularity, and the final numbers on downstream datasets are not forced by that selection. There are no load-bearing self-citations: the cited prior work on DINO, SatMAE, and Scale-MAE is external, and the paper does not invoke a uniqueness theorem from its own authors. The concern that MAE baselines may have been evaluated with an untrained class token is a potential experimental-correctness flaw in the comparison, not a circularity in the derivation chain, because the paper's claim does not reduce to that assumption by definition. Thus, the derivation is self-contained as an empirical study and receives a circularity score of 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central comparison rests on hyperparameters tuned on validation/test data, background assumptions about the effectiveness of DINO-style objectives, and the neutrality of the evaluation protocol. The GSD token is a new latent construct without external evidence.

free parameters (4)

GSD loss weight gamma = 0.1
Chosen from Table 8a, balancing fMoW and downstream classification; selected after inspecting downstream datasets, not an a priori value.
Number of local views = 10
Peak in Table 7a on fMoW-RGB; part of final recipe.
Local/global view scale ranges = [5,25] and [25,100]
Default DINO ranges; ablations in Tables 6b and 7b show these are best on fMoW, though wider local ranges help some downstream tasks.
Augmentation strength = default (probabilities unmodified), no temporal augmentations
Selected from Table 6a; temporal augmentations reduced downstream performance.

assumptions (4)

domain assumption DINO's local-global self-distillation objective yields transferable representations when applied to satellite imagery.
The entire method relies on this inherited from [7]; the paper tests it but assumes the objective is sufficient.
ad hoc to paper Ground sample distance is predictable from a single image crop and adding an auxiliary GSD regression task improves representation quality.
Introduced in Section 3.1; no external evidence supports this, and the paper's own ablation shows mixed, small gains.
domain assumption kNN accuracy and linear probing on the chosen datasets are reliable proxies for feature quality in remote sensing.
Evaluation methodology in Section 3.3; assumes these frozen-feature metrics correlate with downstream usefulness.
domain assumption The MAE baselines are evaluated fairly under the authors' protocol, and the inability to reproduce Cross-Scale MAE does not distort the comparison.
Section 4.1 acknowledges re-evaluated numbers differ from original papers and Cross-Scale MAE is excluded; the comparison presumes neutrality.

invented entities (1)

GSD prediction token
purpose: An extra learnable token alongside the class token; its output is fed to a linear layer that regresses ground sample distance, adding L_GSD to the DINO loss.
New architectural addition with no external validation; the paper only shows small internal performance differences, so this token's benefit is not independently established.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SatDINO: A Deep Dive into Self-Supervised Pretraining for Remote Sensing." pith.science (2026). https://pith.science/paper/JGEEDURM

@misc{pith2026250821402,
  author       = {Pith},
  title        = {Pith review of: SatDINO: A Deep Dive into Self-Supervised Pretraining for Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGEEDURM}},
  note         = {Machine review of arXiv:2508.21402}
}

read the original abstract

Self-supervised learning has emerged as a powerful tool for remote sensing, where large amounts of unlabeled data are available. In this work, we investigate the use of DINO, a contrastive self-supervised method, for pretraining on remote sensing imagery. We introduce SatDINO, a model tailored for representation learning in satellite imagery. Through extensive experiments on multiple datasets in multiple testing setups, we demonstrate that SatDINO outperforms other state-of-the-art methods based on much more common masked autoencoders (MAE) and achieves competitive results in multiple benchmarks. We also provide a rigorous ablation study evaluating SatDINO's individual components. Finally, we propose a few novel enhancements, such as a new way to incorporate ground sample distance (GSD) encoding and adaptive view sampling. These enhancements can be used independently on our SatDINO model. Our code and trained models are available at: https://github.com/strakaj/SatDINO.

Figures

Figures reproduced from arXiv: 2508.21402 by the authors.

**Figure 1.** Visualization of the SatDINO model’s feature representations. (a) Original input image. (b) Attention map from the last layer, where different colors represent individual heads, and the head with the highest attention is shown for each patch. (c) PCA projection of patch embeddings. formance on these downstream applications. In computer vision, unsupervised pretraining methods can be broadly classified into generati… view at source ↗

**Figure 2.** Visualization of random and uniform sampling of the local views. The left side shows the input image with local views outlined by dashed lines, where the number inside each square represents the area of the view. The right side shows the cropped views. In the first strategy, views are randomly sampled from the entire range. In the second strategy, the range is divided into subranges and assigned to individual views… view at source ↗

**Figure 3.** kNN classification performance across multiple datasets and scales. We compare SatDINO-Small with SatMAE and ScaleMAE by evaluating feature robustness using kNN accuracy at different scales. Our model consistently outperforms Scale-MAE across most datasets, demonstrating superior feature quality and generalization. from various locations worldwide. The images are classified into 62 categories, span a wide range of… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 25 canonical work pages

[8]

Extending global-local view alignment for self- supervised learning with remote sensing imagery,

X. Wanyan, S. Seneviratne, S. Shen, and M. Kir- ley, “Extending global-local view alignment for self- supervised learning with remote sensing imagery,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) Work- shops, pp. 2443–2453, June 2024. 1, 2, 3, 7

work page 2024
[1]

Masked autoencoders are scalable vi- sion learners,

K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vi- sion learners,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pp. 16000–16009, 2022. 1

work page 2022
[2]

Learning transferable visual mod- els from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. , “Learning transferable visual mod- els from natural language supervision,” in Interna- 8 tional conference on machine learning , pp. 8748– 8763, PMLR, 2021. 1

work page 2021
[3]

Geo- clip: Clip-inspired alignment between locations and images for effective worldwide geo-localization,

V . Vivanco Cepeda, G. K. Nayak, and M. Shah, “Geo- clip: Clip-inspired alignment between locations and images for effective worldwide geo-localization,” Ad- vances in Neural Information Processing Systems , vol. 36, 2024. 1

work page 2024
[4]

Con- trastive ground-level image and remote sensing pre- training improves representation learning for natural world imagery,

A. V . Huynh, L. E. Gillespie, J. Lopez-Saucedo, C. Tang, R. Sikand, and M. Exp ´osito-Alonso, “Con- trastive ground-level image and remote sensing pre- training improves representation learning for natural world imagery,” in European Conference on Com- puter Vision, pp. 173–190, Springer, 2024. 1

work page 2024
[5]

Mo- mentum contrast for unsupervised visual representa- tion learning,

K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Mo- mentum contrast for unsupervised visual representa- tion learning,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pp. 9729–9738, 2020. 1

work page 2020
[6]

Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data,

O. Manas, A. Lacoste, X. Gir ´o-i Nieto, D. Vazquez, and P. Rodriguez, “Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data,” in Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pp. 9414–9423, 2021. 1

work page 2021
[7]

Emerging properties in self-supervised vision transformers,

M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF international conference on com- puter vision, pp. 9650–9660, 2021. 1

work page 2021

Show all 28 references

[9]

Functional map of the world,

G. Christie, N. Fendley, J. Wilson, and R. Mukher- jee, “Functional map of the world,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6172–6180, 2018. 2

2018
[10]

Satmae: Pre- training transformers for temporal and multi-spectral satellite imagery,

Y . Cong, S. Khanna, C. Meng, P. Liu, E. Rozi, Y . He, M. Burke, D. Lobell, and S. Ermon, “Satmae: Pre- training transformers for temporal and multi-spectral satellite imagery,” Advances in Neural Information Processing Systems, vol. 35, pp. 197–211, 2022. 2, 3

2022
[11]

Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning,

C. J. Reed, R. Gupta, S. Li, S. Brockman, C. Funk, B. Clipp, K. Keutzer, S. Candido, M. Uyttendaele, and T. Darrell, “Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning,” in Proceedings of the IEEE/CVF Interna- tional Conference on Co...

2023
[12]

Cross- scale mae: A tale of multiscale exploitation in remote sensing,

M. Tang, A. Cozma, K. Georgiou, and H. Qi, “Cross- scale mae: A tale of multiscale exploitation in remote sensing,” Advances in Neural Information Processing Systems, vol. 36, 2024. 2, 3

2024
[13]

Self- supervised vision transformers for joint sar-optical representation learning,

Y . Wang, C. M. Albrecht, and X. X. Zhu, “Self- supervised vision transformers for joint sar-optical representation learning,” in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Sympo- sium, pp. 139–142, IEEE, 2022. 2

2022
[14]

Skysense: A multi-modal remote sensing foundation model to- wards universal interpretation for earth observation imagery,

X. Guo, J. Lao, B. Dang, Y . Zhang, L. Yu, L. Ru, L. Zhong, Z. Huang, K. Wu, D. Hu, et al., “Skysense: A multi-modal remote sensing foundation model to- wards universal interpretation for earth observation imagery,” in Proceedings of the IEEE/CVF Confer- ence on Computer Visio...

2024
[15]

An image is worth 16x16 words: Trans- formers for image recognition at scale,

D. Alexey, “An image is worth 16x16 words: Trans- formers for image recognition at scale,”arXiv preprint arXiv: 2010.11929, 2020. 2

2010 arXiv
[16]

Imagenet: A large-scale hierarchical im- age database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical im- age database,” in 2009 IEEE conference on computer vision and pattern recognition , pp. 248–255, Ieee,

2009
[17]

Functional map of the world,

G. Christie, N. Fendley, J. Wilson, and R. Mukherjee, “Functional map of the world,” in CVPR, 2018. 3

2018
[18]

In- troducing eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,

P. Helber, B. Bischke, A. Dengel, and D. Borth, “In- troducing eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,” in IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium , pp. 204–207, IEEE,

2018
[19]

Remote sensing image scene classification: Benchmark and state of the art,

G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classification: Benchmark and state of the art,” Proceedings of the IEEE , vol. 105, pp. 1865–1883, Oct 2017. 3, 4

2017
[20]

Bag-of-visual-words and spatial extensions for land-use classification,

Y . Yang and S. Newsam, “Bag-of-visual-words and spatial extensions for land-use classification,” in Pro- ceedings of the 18th SIGSPATIAL international con- ference on advances in geographic information sys- tems, pp. 270–279, 2010. 3, 4

2010
[21]

Satellite image classification via two-layer sparse coding with biased image representa- tion,

D. Dai and W. Yang, “Satellite image classification via two-layer sparse coding with biased image representa- tion,” IEEE Transactions on Geoscience and Remote Sensing, vol. 8, no. 1, pp. 173–176, 2011. 3, 4

2011
[22]

Feature significance- based multibag-of-visual-words model for remote sensing image scene classification,

L. Zhao, P. Tang, and L. Huo, “Feature significance- based multibag-of-visual-words model for remote sensing image scene classification,” Journal of Ap- plied Remote Sensing , vol. 10, no. 3, pp. 035004– 035004, 2016. 3, 4

2016
[23]

Dirichlet-derived multiple topic scene classification model for high spatial resolution remote sensing im- agery,

B. Zhao, Y . Zhong, G.-S. Xia, and L. Zhang, “Dirichlet-derived multiple topic scene classification model for high spatial resolution remote sensing im- agery,”IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 4, pp. 2108–2123, 2015. 3, 4 9

2015
[24]

International society for photogrammetry and re- mote sensing - potsdam and vaihingen

R. Franz, S. Gunho, G. Markus, and J. D. Weg- ner, “International society for photogrammetry and re- mote sensing - potsdam and vaihingen.” https:// www.isprs.org/education/benchmarks/ UrbanSemLab/default.aspx, 2013. 3, 4

2013
[25]

Loveda: A remote sensing land-cover dataset for do- main adaptive semantic segmentation,

J. Wang, Z. Zheng, A. Ma, X. Lu, and Y . Zhong, “Loveda: A remote sensing land-cover dataset for do- main adaptive semantic segmentation,” arXiv preprint arXiv:2110.08733, 2021. 3, 4

2021 arXiv
[26]

MMSegmentation: Openmm- lab semantic segmentation toolbox and bench- mark

M. Contributors, “MMSegmentation: Openmm- lab semantic segmentation toolbox and bench- mark.” https://github.com/open-mmlab/ mmsegmentation, 2020. 4, 6

2020
[27]

Decoupled weight decay regulariza- tion,

I. Loshchilov, “Decoupled weight decay regulariza- tion,” arXiv preprint arXiv:1711.05101, 2017. 4

2017 arXiv
[28]

Unified perceptual parsing for scene understanding,

T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” in Pro- ceedings of the European conference on computer vi- sion (ECCV), pp. 418–434, 2018. 6 10

2018

Pith tools

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

[1] [8]

Extending global-local view alignment for self- supervised learning with remote sensing imagery,

X. Wanyan, S. Seneviratne, S. Shen, and M. Kir- ley, “Extending global-local view alignment for self- supervised learning with remote sensing imagery,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) Work- shops, pp. 2443–2453, June 2024. 1, 2, 3, 7

work page 2024

[2] [1]

Masked autoencoders are scalable vi- sion learners,

K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vi- sion learners,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pp. 16000–16009, 2022. 1

work page 2022

[3] [2]

Learning transferable visual mod- els from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. , “Learning transferable visual mod- els from natural language supervision,” in Interna- 8 tional conference on machine learning , pp. 8748– 8763, PMLR, 2021. 1

work page 2021

[4] [3]

Geo- clip: Clip-inspired alignment between locations and images for effective worldwide geo-localization,

V . Vivanco Cepeda, G. K. Nayak, and M. Shah, “Geo- clip: Clip-inspired alignment between locations and images for effective worldwide geo-localization,” Ad- vances in Neural Information Processing Systems , vol. 36, 2024. 1

work page 2024

[5] [4]

Con- trastive ground-level image and remote sensing pre- training improves representation learning for natural world imagery,

A. V . Huynh, L. E. Gillespie, J. Lopez-Saucedo, C. Tang, R. Sikand, and M. Exp ´osito-Alonso, “Con- trastive ground-level image and remote sensing pre- training improves representation learning for natural world imagery,” in European Conference on Com- puter Vision, pp. 173–190, Springer, 2024. 1

work page 2024

[6] [5]

Mo- mentum contrast for unsupervised visual representa- tion learning,

K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Mo- mentum contrast for unsupervised visual representa- tion learning,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pp. 9729–9738, 2020. 1

work page 2020

[7] [6]

Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data,

O. Manas, A. Lacoste, X. Gir ´o-i Nieto, D. Vazquez, and P. Rodriguez, “Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data,” in Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pp. 9414–9423, 2021. 1

work page 2021

[8] [7]

Emerging properties in self-supervised vision transformers,

M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF international conference on com- puter vision, pp. 9650–9660, 2021. 1

work page 2021