Towards Realistic Remote Sensing Dataset Distillation with Discriminative Prototype-guided Diffusion

Pedram Ghamisi; Qihao Weng; Yonghao Xu

REVIEW 4 major objections 4 minor 59 references

Towards Realistic Remote Sensing Dataset Distillation with Discriminative Prototype-guided Diffusion

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a diffusion model guided by discriminative prototypes can distill large remote sensing datasets into compact synthetic sets that beat existing generative distillation baselines across eight classifier architectures.

desk verdict Solid first application of diffusion-based dataset distillation to RS scene classification, but the headline accuracy claim is confounded by unequal access to real data and the arXiv abstract doesn't match the paper. read the letter →

arxiv 2601.15829 v2 pith:D4BZRT54 submitted 2026-01-22 cs.CV

Yonghao Xu , Pedram Ghamisi , Qihao Weng This is my paper

classification cs.CV

keywords datasetdistillationremotesensingsceneclassificationdiffusionmodelstext-to-imagegenerationvisualprototypelatentspaceclusteringsynthetictrainingdataprivacy-preservinglearning

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 tries to establish that dataset distillation—replacing a large real training set with a small set of synthetic images—works for remote sensing scene classification, and that a diffusion-based method with two added guidance signals is the way to do it. The first signal is a classification consistency loss from a pretrained classifier, injected into diffusion training so synthesized images stay class-discriminative. The second is a vision-language prototype extraction step that clusters training images in the VAE latent space, picks a margin-maximizing prototype per cluster, and aggregates cluster captions, then initializes denoising from the prototype latent. On UCM, AID, and NWPU-RESISC45, the method consistently outperforms four competing methods across all IPC settings and eight classifier architectures. If true, it means compact fully synthetic remote sensing training sets are practical, with storage and privacy benefits.

What carries the argument

The carrying object is the DPD pipeline: a latent diffusion model fine-tuned with LoRA on a Stable Diffusion 2 backbone, with frozen VAE and CLIP text encoders; a classification consistency loss that pushes the decoded image to match the ground-truth label under a pretrained ResNet-18; and a vision-language prototype extraction stage in which K-Means on VAE latents within each category yields K clusters, a margin criterion selects the most representative latent per cluster, and a VLM-plus-LLM pair turns raw captions into one aggregated caption per cluster. At distillation time, reverse diffusion starts from the prototype latent rather than pure noise and is conditioned on the aggregated capt

What would settle it

A decisive experiment is to hold the entire DPD pipeline fixed but replace the margin-selected prototype with a random member of the same K-Means cluster, then measure downstream classifier accuracy; if random selection matches DPD, the prototype criterion is not carrying the argument. A complementary check is to compute the cluster purity of the VAE latent space against ground-truth class labels on a held-out remote sensing split.

Watch

Extended reading notes

Core claim

The paper's central claim is that realistic dataset distillation for remote sensing can be achieved by a diffusion model that is (1) fine-tuned with a classification consistency loss so that reconstructed images are discriminative, and (2) run from representative visual prototypes selected by margin-based clustering in the VAE latent space, conditioned on aggregated cluster captions. The authors argue that this dual guidance is what previous generative baselines lack: they optimize visual realism rather than downstream classification utility, and they start from random noise rather than category-representative latents. Quantitatively, DPD's distilled sets train eight different classifiers to

Load-bearing premise

The whole prototype step assumes that two remote sensing images close to each other in the Stable Diffusion VAE latent space—a space trained on ordinary natural photos—also look alike as aerial scenes; if latent distance does not track aerial scene semantics, the prototype selection and the diversity it claims can collapse.

Editorial extensions

If this is right

A model trained only on DPD-distilled synthetic images—never seeing the original training pixels—can beat other generative distillation baselines on UCM, AID, and NWPU-RESISC45 across AlexNet, VGG16, Inception-v3, ResNet18/101, DenseNet121/201, and RegNetX-16GF.
The classification consistency loss is the dominant component: adding L_cls to the diffusion loss raises IPC=10 accuracy by roughly 20 points on UCM and AID and more than 25 points on NWPU.
Distilled samples transfer across architectures, suggesting the distilled set is not overfit to the pretrained classifier used during generation.
At IPC=20, using about 3.3% of NWPU-RESISC45 training data, DPD reaches 65–72% accuracy depending on network; the remaining gap to full-data training (87–95%) defines the headroom for future work.
Because the distilled dataset contains no original images, the method offers a privacy-preserving path for training classifiers on sensitive remote sensing categories.

Reading between the lines

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

The prototype selection assumes Euclidean distance in a natural-image-pretrained VAE latent space tracks semantic similarity for aerial scenes; a straightforward test is to measure K-Means cluster purity against human labels on a remote sensing split—low purity would weaken the diversity claim.
The privacy benefit should be probed adversarially: although original pixels are absent, prototypes and aggregated captions may encode recognizable scene statistics, and an inversion or nearest-neighbor attack on distilled samples could reveal whether real training images are recoverable.
The dual-guidance design seems transferable to dense tasks such as land cover segmentation or object detection, which the paper explicitly leaves open; the prototype-plus-caption structure should adapt to spatially structured outputs.
One could isolate the margin criterion by replacing it with a random cluster member while keeping everything else fixed: if downstream accuracy does not drop, the criterion is not the source of the gain.

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.

Desk Editor's Note

read the letter

Short version: this paper deserves a serious referee, but the central claim—DPD 'consistently outperforms all competing methods'—is not supported as stated. DPD gets two big advantages the baselines don't: a classifier pretrained on the full real training set, and real-image latent prototypes selected from that set. The ablation shows those two components drive most of the gain, so the comparison is not apples-to-apples. The arXiv abstract also describes a different method (hyperspherical anchors, logit-margin ranking) than the full text (classifier consistency loss, VAE-latent clustering, LLM caption aggregation). That needs fixing before the results can be taken at face value.

What's genuinely good: the experimental package is careful. Ten-run means and standard deviations across eight architectures on three benchmarks, plus ablations and a parameter analysis. That's more thorough than the average submission. The idea of using a classification-consistency loss during diffusion training and selecting prototypes in the VAE latent space is sensible, and the qualitative figures suggest the method does produce varied, category-respecting images. The computational cost table is a useful addition.

The soft spots, in proportion: (1) The comparison asymmetry is the load-bearing problem. The baselines condition on class labels or text templates; DPD also uses a ResNet-18 pretrained on R to compute L_cls and initializes generation from real latents selected from R. Table V shows those components account for roughly 20–25 and 29 points respectively on UCM, so the reported margins over baselines may come largely from unequal access to R. A controlled comparison—giving baselines the same real-data supervision, or removing it from DPD—is needed. (2) The abstract mismatch is serious presentation sloppiness; a reader cannot tell what is actually being proposed. (3) The assumption that the Stable Diffusion VAE latent space carries meaningful semantic geometry for remote sensing scenes is plausible and the ablations support it on these three datasets, but it's not validated beyond that; I'd call that a minor concern, not fatal. (4) The claim of 'without exposing any original data' overstates the privacy story—the diffusion model is trained on real data and initialized from real latents; only the final distilled set is synthetic.

Who is this for? Remote sensing teams wanting to train classifiers on small synthetic sets. It's a practical contribution to that subfield, not a field-level breakthrough. I'd bring it to the reading group because the confound and abstract mismatch make for a useful discussion about benchmarking practices.

Recommendation: send to peer review. The empirical work is solid enough to warrant referee time, but reviewers should push for a controlled baseline and a consistency rewrite of the abstract and introduction.

Referee Report

4 major / 4 minor

Summary. The paper proposes DPD, a latent-diffusion framework for remote sensing dataset distillation. The full-text method (Section III) fine-tunes a Stable Diffusion 2 backbone with a diffusion loss plus a classification-consistency loss from a ResNet-18 pretrained on the real training set, selects prototypes by K-Means clustering and a margin criterion in the VAE latent space, generates cluster-level text prompts with a VLM/LLM, and reconstructs distilled images from prototype-initialized reverse diffusion. Experiments on UCM, AID, and NWPU-RESISC45 across eight architectures and IPC ∈ {3,5,10,15,20} report repeated-run means and standard deviations, with DPD claimed to consistently outperform four baselines. The paper also presents ablation studies and computational cost comparisons.

Significance. If the claims hold, this would be one of the first systematic studies of diffusion-based dataset distillation for remote sensing scene classification, with a compact fully synthetic training set that transfers across multiple architectures. The paper has concrete strengths: the evaluation is an external benchmark rather than a fitted derivation; results are reported as means and standard deviations over ten runs; eight classifier architectures are tested; and the code and pretrained models are promised publicly. The ablation study in Table V gives a clear decomposition of the proposed components. However, the current manuscript has serious specification and comparison issues—an abstract that describes a different method, an asymmetric comparison against baselines, an unvalidated latent-space prototype assumption, and an underspecified generation procedure—so the central empirical claim is not yet supported as written.

major comments (4)

[Abstract vs. §III] The manuscript contains two incompatible method descriptions. The arXiv abstract introduces hyperspherical semantic anchors around prototypes and a logit-margin criterion with a latent classifier for candidate ranking; neither appears in Section III (Eqs. 5–7, Algorithm 1) or Algorithm 2. The full-text abstract instead describes classifier-driven guidance and latent-space clustering with VLM text aggregation. This makes the claim that 'DPD consistently outperforms all competing methods' refer to a method that is not uniquely specified. The authors must reconcile the abstract with the body: if hyperspherical anchors and logit-margin ranking are used, provide the equations, algorithm, and ablation; if not, remove them.
[§IV-D, Tables I–III, Eq. (6), Algorithm 2] The head-to-head comparison is asymmetric. DPD trains L_cls with a ResNet-18 pretrained on the full real training set R and initializes generation from real VAE latents selected from R (Algorithm 2). The four baselines receive only class labels or text templates and do not have access to this privileged supervision. Table V shows that adding L_cls alone improves OA by roughly 20–25 points on UCM/AID and by over 25 points on NWPU, and adding visual prototype guidance adds roughly another 29 points on UCM at IPC=10. Therefore the reported margins over baselines may be largely due to unequal access to the real dataset rather than to the proposed prototype-guided diffusion machinery. A controlled comparison is needed: either give the same classifier guidance and/or real-latent initialization to the baselines, or ablate those components from DPD while keeping the rest, and report the resultin
[§III-C, Eqs. (8)–(11)] The prototype-extraction mechanism assumes that Euclidean distance in the frozen Stable Diffusion VAE latent space—pretrained on natural images—reflects semantic similarity for remote sensing scenes. This premise is not validated. The margin criterion in Eq. (10) selects samples that are far from other cluster centroids and near their own centroid; in a natural-image latent space, aerial scenes may cluster by low-level color/texture rather than land-use semantics, and the selected prototype can be a boundary sample. Because Table V assigns a large performance gain to visual prototype guidance, the authors should provide quantitative validation: cluster purity per class, comparison of the margin-selected prototypes against random prototypes or class-mean latents, and sensitivity of results to the number of clusters K.
[§III-D, Algorithm 2] The generation procedure is not fully specified. Algorithm 2 initializes z0 with the real prototype and then says 'estimate the clean latent by reverse diffusion.' Reverse diffusion normally starts from a noised state z_T, not from a clean z0; if z0 is already clean, the reverse process has nothing to denoise. There is no explicit forward-noising or inversion step, no number of diffusion steps, and no noise schedule for the distillation stage, even though Fig. 8(b) varies T during distillation. Please specify the complete protocol, e.g., z_T = sqrt(αbar_T) z* + sqrt(1−αbar_T) ε followed by T denoising steps; as written, the method is not reproducible.

minor comments (4)

[§IV-F, Table V] On UCM at IPC=10, adding caption aggregation lowers OA from 79.38 to 78.30, yet the final DPD configuration includes caption aggregation. The text acknowledges this but still adopts the full configuration globally. Please either justify the global choice or select the best configuration per dataset and state this explicitly.
[§V, Algorithm 2] The privacy claim 'without exposing any original data' is stronger than what is demonstrated. Algorithm 2 stores real VAE latent prototypes and initializes generation from them; distilled images may retain identifiable content. At minimum, weaken the claim to 'reduces direct storage and exposure' and provide a simple leakage check (e.g., nearest-neighbor distance to real training images, membership-inference evaluation).
[§IV-C] The caption-aggregation ablation is reported only for the full method with IPC=10 and ResNet18. To support the cross-architecture generalization claim, at least one caption-aggregation ablation across another architecture or IPC would be useful.
[§IV-E, Figures 4–7] The statement that all qualitative results are shown 'without manual cherry picking' is not verifiable from a figure. Please include the full generated image grids in the supplementary material or repository.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: DPD's central result is an external benchmark; the main caveat is an asymmetric comparison, not circularity.

full rationale

The central claim is empirical and is tested against held-out test sets, so it does not reduce to the method's inputs by construction. The classification-consistency loss (Eq. 6) uses a ResNet-18 pretrained on the real training set R as a regularizer during diffusion training, but the downstream classifiers in Tables I–III are trained from scratch on the distilled set and evaluated on unseen test splits, so the reported OA is not definitionally tied to f_phi. The prototypes selected by Eqs. (8)–(11) are real-image VAE latents used as initialization for generation; this makes the method heavily dependent on R and weakens the 'without exposing any original data' privacy claim, but it is an input to the generative process, not a fitted parameter renamed as a prediction. Self-citations such as [17] (Txt2Img-MHN) appear as a baseline and as motivation for why visual realism alone is insufficient; they are not used as a load-bearing external uniqueness result. No equation in the paper is equivalent by construction to any reported result. The reviewer-rule flags are non-circular rigor concerns: (1) the comparison is asymmetric—DPD receives privileged real-data access (pretrained classifier, real-image prototypes, VLM-generated captions) that the baselines do not, so 'consistently outperforms all competing methods across all settings' (Section IV-D) is confounded; and (2) the abstract describes hyperspherical semantic anchors and latent-classifier logit-margin candidate ranking, neither of which appears in the full-text methodology. These affect the strength of the empirical claim but are not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on standard diffusion-model assumptions plus domain assumptions about the transferability of natural-image-pretrained VAE/CLIP/LLM features to remote sensing, and on the reliability of a classifier pretrained on the target data as a semantic guide. No new physical entities are postulated.

free parameters (2)

lambda (classification consistency weight) = 0.3
Tuned via parameter analysis on UCM/AID/NWPU with IPC=10 (Fig. 8a) and adopted in all experiments; no held-out validation reported.
T (number of diffusion steps during distillation) = 50
Selected via parameter analysis (Fig. 8b) as a practical trade-off; adopted in all experiments.

assumptions (4)

domain assumption Euclidean distance in the Stable Diffusion VAE latent space reflects semantic similarity of remote sensing scenes.
Used in K-Means clustering and margin-based prototype selection (Eqs. 8–11); no validation that latent geometry tracks scene semantics for aerial imagery.
domain assumption A classifier pretrained on the real training set provides gradients that improve the discriminative quality of diffusion-generated samples.
Central to L_cls (Eq. 6); the paper gives no analysis of how the pretrained classifier's own errors affect the distilled set.
domain assumption LLM-aggregated text prompts preserve the visual information needed for representative remote sensing image generation.
Used in prototype-guided generation (Section III-D); no quantitative check of caption fidelity against the images they summarize.
domain assumption The standard dataset distillation transfer objective holds: models trained on the distilled set generalize to the real test set.
This is the core assumption of Section III-A and is the paper's evaluation protocol, assumed rather than derived.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Realistic Remote Sensing Dataset Distillation with Discriminative Prototype-guided Diffusion." pith.science (2026). https://pith.science/paper/D4BZRT54

@misc{pith2026260115829,
  author       = {Pith},
  title        = {Pith review of: Towards Realistic Remote Sensing Dataset Distillation with Discriminative Prototype-guided Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4BZRT54}},
  note         = {Machine review of arXiv:2601.15829}
}

read the original abstract

Recent years have witnessed the remarkable success of deep learning in remote sensing image interpretation, driven by the availability of large-scale benchmark datasets. However, this reliance on massive training data also brings substantial storage and computational costs. To address this challenge, this study introduces the concept of dataset distillation into the field of remote sensing image interpretation for the first time. Specifically, we propose discriminative prototype-guided diffusion (DPD), a diffusion-based generative distillation framework that condenses a large-scale remote sensing dataset into a compact and representative distilled dataset. To improve the semantic fidelity and diversity of the synthesized samples, we extract representative prototypes for each category in the latent space. We then construct hyperspherical semantic anchors around the prototypes to guide the reverse denoising trajectory. Furthermore, to enhance the discriminative quality of the generated samples, multiple candidates are generated for each prototype and ranked by a latent classifier using a logit-margin criterion, with the most discriminative candidates selected to form the final distilled dataset. Experiments on three high-resolution remote sensing scene classification benchmarks show that the proposed method can distill realistic, diverse, and discriminative samples for downstream model training. Code and pre-trained models are available online (https://github.com/YonghaoXu/DPD).

Figures

Figures reproduced from arXiv: 2601.15829 by the authors.

**Figure 2.** Overview of the proposed classification-consistent diffusion training pipeline. Given a remote sensing image, a VAE encoder maps it into the latent [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Illustration of the margin-based prototype selection, where the intra [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Qualitative comparison of randomly selected real images and distilled images generated by different dataset distillation methods on the UCM, AID, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Visualization of distilled images across 21 classes on the UCM dataset (IPC = 5). Note that all results are shown without manual “cherry picking”. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Visualization of distilled images across 30 classes on the AID dataset (IPC = 5). Note that all results are shown without manual “cherry picking”. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Visualization of distilled images across 45 classes on the NWPU dataset (IPC = 3). Note that all results are shown without manual “cherry picking”. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: Parameter analysis of the proposed method on the UCM, AID, and [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 6 linked inside Pith

[1]

Deep learning in remote sensing: A comprehensive review and list of resources,

X. X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, and F. Fraundorfer, “Deep learning in remote sensing: A comprehensive review and list of resources,”IEEE Geosci. Remote Sens. Mag., vol. 5, no. 4, pp. 8–36, 2017

2017
[2]

How will ai transform urban observing, sensing, imaging, and mapping?

Q. Weng, Z. Li, Y . Cao, X. Lu, P. Gamba, X. Zhu, Y . Xu, F. Zhang, R. Qin, M. Y . Yanget al., “How will ai transform urban observing, sensing, imaging, and mapping?”npj Urban Sustainability, vol. 4, no. 1, p. 50, 2024

2024
[3]

Scenenet: Remote sensing scene classification deep learning network using multi-objective neural evolution architecture search,

A. Ma, Y . Wan, Y . Zhong, J. Wang, and L. Zhang, “Scenenet: Remote sensing scene classification deep learning network using multi-objective neural evolution architecture search,”ISPRS J. Photogramm. Remote Sens., vol. 172, pp. 171–188, 2021

2021
[4]

Dota: A large-scale dataset for object detection in aerial images,

G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object detection in aerial images,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 3974–3983

2018
[5]

Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,

L. Wang, R. Li, C. Zhang, S. Fang, C. Duan, X. Meng, and P. M. Atkinson, “Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,”ISPRS J. Pho- togramm. Remote Sens., vol. 190, pp. 196–214, 2022

2022
[6]

Artificial intelligence for remote sensing data analysis: A review of challenges and opportunities,

L. Zhang and L. Zhang, “Artificial intelligence for remote sensing data analysis: A review of challenges and opportunities,”IEEE Geosci. Remote Sens. Mag., vol. 10, no. 2, pp. 270–294, 2022

2022
[7]

There are no data like more data: Datasets for deep learning in earth observation,

M. Schmitt, S. A. Ahmadi, Y . Xu, G. Tas ¸kin, U. Verma, F. Sica, and R. H¨ansch, “There are no data like more data: Datasets for deep learning in earth observation,”IEEE Geosci. Remote Sens. Mag., vol. 11, no. 3, pp. 63–97, 2023

2023
[8]

Research progress on few-shot learning for remote sensing image interpretation,

X. Sun, B. Wang, Z. Wang, H. Li, H. Li, and K. Fu, “Research progress on few-shot learning for remote sensing image interpretation,”IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 2387–2402, 2021

2021

Show all 59 references

[9]

Ai security for geoscience and remote sensing: Challenges and future trends,

Y . Xu, T. Bai, W. Yu, S. Chang, P. M. Atkinson, and P. Ghamisi, “Ai security for geoscience and remote sensing: Challenges and future trends,”IEEE Geosci. Remote Sens. Mag., vol. 11, no. 2, pp. 60–85, 2023. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, JANUARY 2022 13

2023
[10]

Responsible artificial intelligence for earth observation: Achievable and realistic paths to serve the collective good,

P. Ghamisi, W. Yu, A. Marinoni, C. M. Gevaert, C. Persello, S. Selvaku- maran, M. Girotto, B. P. Horton, P. Rufin, P. Hostert, F. Pacifici, and P. M. Atkinson, “Responsible artificial intelligence for earth observation: Achievable and realistic paths to serve the collective go...

2025
[11]

A comprehensive survey of dataset distillation,

S. Lei and D. Tao, “A comprehensive survey of dataset distillation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 1, pp. 17–32, 2023

2023
[12]

Dataset distillation: A comprehensive review,

R. Yu, S. Liu, and X. Wang, “Dataset distillation: A comprehensive review,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 1, pp. 150–170, 2023

2023
[13]

Dataset distillation,

T. Wang, J.-Y . Zhu, A. Torralba, and A. A. Efros, “Dataset distillation,” arXiv preprint arXiv:1811.10959, 2018

2018 arXiv
[14]

Efficient dataset distillation via minimax diffusion,

J. Gu, S. Vahidian, V . Kungurtsev, H. Wang, W. Jiang, Y . You, and Y . Chen, “Efficient dataset distillation via minimax diffusion,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2024, pp. 15 793–15 803

2024
[15]

Taming diffusion for dataset distillation with high representativeness,

L. Zhao, Y . Wu, X. Jiang, J. Gu, Y . Wang, X. Xu, P. Zhao, and X. Lin, “Taming diffusion for dataset distillation with high representativeness,” inProc. Int. Conf. Mach. Learn., 2025

2025
[16]

Deep learning for remote sensing data: A technical tutorial on the state of the art,

L. Zhang, L. Zhang, and B. Du, “Deep learning for remote sensing data: A technical tutorial on the state of the art,”IEEE Geosci. Remote Sens. Mag., vol. 4, no. 2, pp. 22–40, 2016

2016
[17]

Txt2img- mhn: Remote sensing image generation from text using modern hopfield networks,

Y . Xu, W. Yu, P. Ghamisi, M. Kopp, and S. Hochreiter, “Txt2img- mhn: Remote sensing image generation from text using modern hopfield networks,”IEEE Trans. Image Process., vol. 32, pp. 5737–5750, 2023

2023
[18]

Text-to-remote-sensing-image generation with structured generative adversarial networks,

R. Zhao and Z. Shi, “Text-to-remote-sensing-image generation with structured generative adversarial networks,”IEEE Geosci. Remote Sens. Lett., vol. 19, pp. 1–5, 2021

2021
[19]

Crs-diff: Controllable remote sensing image generation with diffusion model,

D. Tang, X. Cao, X. Hou, Z. Jiang, J. Liu, and D. Meng, “Crs-diff: Controllable remote sensing image generation with diffusion model,” IEEE Trans. Geos. Remote Sens., 2024

2024
[20]

Improved techniques for training GANs,

T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training GANs,” inProc. Neural Inf. Process. Syst., vol. 29, 2016

2016
[21]

Gans trained by a two time-scale update rule converge to a local nash equilibrium,

M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” inProc. Neural Inf. Process. Syst., vol. 30, 2017

2017
[22]

Energy and policy consid- erations for modern deep learning research,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy consid- erations for modern deep learning research,” inProc. AAAI Conf. Artif. Intell., vol. 34, no. 09, 2020, pp. 13 693–13 696

2020
[23]

Flexible dataset distillation: Learn labels instead of images,

O. Bohdal, Y . Yang, and T. M. Hospedales, “Flexible dataset distillation: Learn labels instead of images,” inWorkshop on Meta-Learning at NeurIPS, 2020

2020
[24]

Soft-label dataset distillation and text dataset distillation,

I. Sucholutsky and M. Schonlau, “Soft-label dataset distillation and text dataset distillation,” inProc. Int. Joint Conf. Neural Netw.IEEE, 2021, pp. 1–8

2021
[25]

Dataset condensation with gradient matching,

B. Zhao, K. R. Mopuri, and H. Bilen, “Dataset condensation with gradient matching,” inProc. Int. Conf. Learn. Representations, 2021

2021
[26]

Dataset distillation by matching training trajectories,

G. Cazenavette, T. Wang, A. Torralba, A. A. Efros, and J.-Y . Zhu, “Dataset distillation by matching training trajectories,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 4750–4759

2022
[27]

Dataset condensation with distribution match- ing,

B. Zhao and H. Bilen, “Dataset condensation with distribution match- ing,” inProc. IEEE Winter Conf. Appl. Comput. Vis., 2023, pp. 6514– 6523

2023
[28]

Dim: Distilling dataset into generative model,

K. Wang, J. Gu, J. Gu, H. Zhang, D. Zhou, Z. Zhu, W. Jiang, and Y . You, “Dim: Distilling dataset into generative model,” inProc. Eur. Conf. Comput. Vis.Springer, 2024, pp. 42–59

2024
[29]

Dˆ 4: Dataset distillation via disentangled diffusion model,

D. Su, J. Hou, W. Gao, Y . Tian, and B. Tang, “Dˆ 4: Dataset distillation via disentangled diffusion model,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2024, pp. 5809–5818

2024
[30]

Zero-shot text-to-image generation,

A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. V oss, A. Radford, M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,” inProc. Int. Conf. Mach. Learn., 2021, pp. 8821–8831

2021
[31]

Vision-language models in remote sensing: Current progress and future trends,

X. Li, C. Wen, Y . Hu, Z. Yuan, and X. X. Zhu, “Vision-language models in remote sensing: Current progress and future trends,”IEEE Geosci. Remote Sens. Mag., vol. 12, no. 2, pp. 32–66, 2024

2024
[32]

Diffusion models meet remote sensing: Principles, methods, and perspectives,

Y . Liu, J. Yue, S. Xia, P. Ghamisi, W. Xie, and L. Fang, “Diffusion models meet remote sensing: Principles, methods, and perspectives,” IEEE Trans. Geos. Remote Sens., 2024

2024
[33]

Diffusionsat: A generative foundation model for satellite imagery,

S. Khanna, P. Liu, L. Zhou, C. Meng, R. Rombach, M. Burke, D. B. Lobell, and S. Ermon, “Diffusionsat: A generative foundation model for satellite imagery,” inProc. Int. Conf. Learn. Representations
[34]

Rs5m and georsclip: A large scale vision-language dataset and a large vision-language model for remote sensing,

Z. Zhang, T. Zhao, Y . Guo, and J. Yin, “Rs5m and georsclip: A large scale vision-language dataset and a large vision-language model for remote sensing,”IEEE Trans. Geos. Remote Sens., 2024

2024
[35]

Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model,

C. Liu, K. Chen, R. Zhao, Z. Zou, and Z. Shi, “Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model,”IEEE Geosci. Remote Sens. Mag., 2025

2025
[36]

Federated learning across decentralized and unshared archives for remote sensing image classi- fication: A review,

B. B ¨uy¨uktas ¸, G. Sumbul, and B. Demir, “Federated learning across decentralized and unshared archives for remote sensing image classi- fication: A review,”IEEE Geosci. Remote Sens. Mag., vol. 12, no. 3, pp. 64–80, 2024

2024
[37]

Federated deep learning with prototype matching for object extraction from very-high-resolution remote sensing images,

X. Zhang, B. Zhang, W. Yu, and X. Kang, “Federated deep learning with prototype matching for object extraction from very-high-resolution remote sensing images,”IEEE Trans. Geos. Remote Sens., vol. 61, pp. 1–16, 2023

2023
[38]

Fedrsclip: Federated learning for remote sensing scene classification using vision-language models,

H. Lin, C. Zhang, D. Hong, K. Dong, and C. Wen, “Fedrsclip: Federated learning for remote sensing scene classification using vision-language models,”IEEE Geosci. Remote Sens. Mag., 2025

2025
[39]

Local differential privacy based membership-privacy-preserving federated learning for deep-learning- driven remote sensing,

Z. Zhang, X. Ma, and J. Ma, “Local differential privacy based membership-privacy-preserving federated learning for deep-learning- driven remote sensing,”Remote Sensing, vol. 15, no. 20, p. 5050, 2023

2023
[40]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, 2020, pp. 6840– 6851

2020
[41]

High- resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 10 684–10 695

2022
[42]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inProc. Int. Conf. Mach. Learn.PmLR, 2021, pp. 8748–8763

2021
[43]

Scalable diffusion models with transformers,

W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProc. IEEE Int. Conf. Comput. Vis., 2023, pp. 4195–4205

2023
[44]

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,” inProc. SIGSPATIAL Int. Conf. Adv. Geo- graphic Inf. Syst.ACM, 2010, pp. 270–279

2010
[45]

Aid: A benchmark data set for performance evaluation of aerial scene classification,

G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y . Zhong, L. Zhang, and X. Lu, “Aid: A benchmark data set for performance evaluation of aerial scene classification,”IEEE Trans. Geosci. Remote Sens., vol. 55, no. 7, pp. 3965–3981, 2017

2017
[46]

Remote sensing image scene classifica- tion: Benchmark and state of the art,

G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classifica- tion: Benchmark and state of the art,”Proc. IEEE, vol. 105, no. 10, pp. 1865–1883, 2017

2017
[47]

Imagenet classification with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,”Proc. Neural Inf. Process. Syst., vol. 25, pp. 1097–1105, 2012

2012
[48]

Very deep convolutional networks for large-scale image recognition,

K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

2014 arXiv
[49]

Rethinking the inception architecture for computer vision,

C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 2818–2826

2016
[50]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 770–778

2016
[51]

Densely connected convolutional networks,

G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017, pp. 4700–4708

2017
[52]

Designing network design spaces,

I. Radosavovic, R. P. Kosaraju, R. Girshick, K. He, and P. Doll ´ar, “Designing network design spaces,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2020, pp. 10 428–10 436

2020
[53]

Openclip,

G. Ilharco, M. Wortsman, R. Wightman, C. Gordon, N. Carlini, R. Taori, A. Dave, V . Shankar, H. Namkoong, J. Miller, H. Hajishirzi, A. Farhadi, and L. Schmidt, “Openclip,” Jul. 2021

2021
[54]

LAION-5b: An open large-scale dataset for training next generation image-text models,

C. Schuhmann, R. Beaumont, R. Vencu, C. W. Gordon, R. Wight- man, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, P. Schramowski, S. R. Kundurthy, K. Crowson, L. Schmidt, R. Kaczmar- czyk, and J. Jitsev, “LAION-5b: An open large-scale dataset for training next generat...

2022
[55]

Qwen2.5-vl technical report,

S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

2025 arXiv
[56]

Lora: Low-rank adaptation of large language models

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” inProc. Int. Conf. Learn. Representations, 2022

2022
[57]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

2017 arXiv
[58]

A method for stochastic optimization,

K. D. B. J. Adamet al., “A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, vol. 1412, no. 6, 2014

2014 arXiv
[59]

Qwen3 technical report,

A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

2025 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Deep learning in remote sensing: A comprehensive review and list of resources,

X. X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, and F. Fraundorfer, “Deep learning in remote sensing: A comprehensive review and list of resources,”IEEE Geosci. Remote Sens. Mag., vol. 5, no. 4, pp. 8–36, 2017

2017

[2] [2]

How will ai transform urban observing, sensing, imaging, and mapping?

Q. Weng, Z. Li, Y . Cao, X. Lu, P. Gamba, X. Zhu, Y . Xu, F. Zhang, R. Qin, M. Y . Yanget al., “How will ai transform urban observing, sensing, imaging, and mapping?”npj Urban Sustainability, vol. 4, no. 1, p. 50, 2024

2024

[3] [3]

Scenenet: Remote sensing scene classification deep learning network using multi-objective neural evolution architecture search,

A. Ma, Y . Wan, Y . Zhong, J. Wang, and L. Zhang, “Scenenet: Remote sensing scene classification deep learning network using multi-objective neural evolution architecture search,”ISPRS J. Photogramm. Remote Sens., vol. 172, pp. 171–188, 2021

2021

[4] [4]

Dota: A large-scale dataset for object detection in aerial images,

G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object detection in aerial images,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 3974–3983

2018

[5] [5]

Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,

L. Wang, R. Li, C. Zhang, S. Fang, C. Duan, X. Meng, and P. M. Atkinson, “Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,”ISPRS J. Pho- togramm. Remote Sens., vol. 190, pp. 196–214, 2022

2022

[6] [6]

Artificial intelligence for remote sensing data analysis: A review of challenges and opportunities,

L. Zhang and L. Zhang, “Artificial intelligence for remote sensing data analysis: A review of challenges and opportunities,”IEEE Geosci. Remote Sens. Mag., vol. 10, no. 2, pp. 270–294, 2022

2022

[7] [7]

There are no data like more data: Datasets for deep learning in earth observation,

M. Schmitt, S. A. Ahmadi, Y . Xu, G. Tas ¸kin, U. Verma, F. Sica, and R. H¨ansch, “There are no data like more data: Datasets for deep learning in earth observation,”IEEE Geosci. Remote Sens. Mag., vol. 11, no. 3, pp. 63–97, 2023

2023

[8] [8]

Research progress on few-shot learning for remote sensing image interpretation,

X. Sun, B. Wang, Z. Wang, H. Li, H. Li, and K. Fu, “Research progress on few-shot learning for remote sensing image interpretation,”IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 2387–2402, 2021

2021