Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

A gated flow-matching model turns 10 m Sentinel-2 views into 2 m Maxar-style imagery of rare Arctic thaw slumps, beating six baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A gated dual-conditioning flow-matching model achieves 10 m→2 m cross-sensor super-resolution with a 38% FID reduction over the best baseline on a rare-landform (retrogressive thaw slump) benchmark.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Nice gating mechanism and a useful benchmark, but the numbers are unverifiable until the authors say where test-time captions come from; App. .3 suggests they come from the HR target, which would sink the SOTA claim. the 3 major comments →

arxiv 2510.23816 v3 pith:2KNUS5O5 submitted 2025-10-27 cs.CV

Semantic-Guided Cross-Sensor Super Resolution of Remote Sensing Images: A Gated Dual Conditioning Flow Matching Model

Forouzan Fallah , Wenwen Li , Chia-Yu Hsu , Hyunho Lee , Anna Liljedahl , Yezhou Yang This is my paper
classification cs.CV
keywords cross-sensor super-resolutionflow matchinggated ControlNetMonte Carlo dropout uncertaintyretrogressive thaw slumpsremote sensingSentinel-2 to Maxarperceptual metrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

RareFlow is an attempt to close the gap between freely available 10 m Sentinel-2 satellite imagery and expensive 2 m commercial Maxar imagery, in a scientific domain where the target features are rare: retrogressive thaw slumps, permafrost landslides in the Arctic. The paper claims that a flow-matching generative model with dual conditioning—a gated ControlNet that keeps scene geometry from the low-resolution input and text prompts that supply geological semantics—can synthesize high-resolution images that are both geometrically faithful and stylistically matched to the target sensor. On a new benchmark of roughly 800 paired images, it reports the best perceptual similarity scores (LPIPS 0.36, DISTS 0.30), the best realism score (FID 116.16, a 38% reduction over the next-best baseline), and expert judgments that its outputs are perceptually on par with 2 m Maxar ground truth. If the claims hold, the framework would let Arctic monitoring rely less on commercial high-resolution data, and would provide a template for super-resolving rare, out-of-distribution phenomena in other data-scarce sciences.

Core claim

RareFlow's central claim is that cross-sensor super-resolution can be cast as controlled generation in which the target sensor's style is a signal, not a nuisance. A frozen SD3 flow-matching backbone generates the image; only the ControlNet pathway and per-block scalar gates α_l(t,u) are trained. The gates, driven by diffusion timestep and Monte-Carlo dropout uncertainty, decide how much blurry low-resolution structure enters each transformer block, while vision–language-model text prompts supply the rare thaw-slump semantics. A combined loss—flow matching plus FFT frequency alignment, CIELAB color consistency, and LPIPS—anchors output to ground truth. The paper reports that this resolves th

What carries the argument

The load-bearing mechanism is the gated ControlNet residual with per-block scalars: α_l(t,u) = σ(p_l0 + p_lt·norm(t) + p_lu·u). A ControlNet consuming the low-resolution latent emits residual maps r_l for each block of the frozen SD3 transformer; before injection they are scaled by α_l, a logistic function of normalized diffusion time t and an uncertainty scalar u. u is computed from T Monte-Carlo dropout forward passes as clipped mean variance, so the model can down-weight structural influence when the low-resolution evidence is deemed unreliable. The gate is what lets the model navigate between spatial-only conditioning, which preserves coarse geometry but propagates blur, and semantic-onl

Load-bearing premise

The load-bearing premise is that the paired Sentinel-2 and Maxar images, acquired at different times and hence spatially shifted, differently illuminated, and changed in land cover (e.g., snow cover), still constitute valid supervision for learning a true 10 m-to-2 m mapping — if temporal-change artifacts dominate, the reported gains measure style matching and texture injection rather than genuine super-resolution.

What would settle it

A decisive check would be to validate on temporally stable pairs: select Sentinel-2 and Maxar acquisitions of the same thaw slump within days, with matching season and no snow change, and test whether RareFlow's PSNR/SSIM/LPIPS gains over baselines persist when ground truth is truly aligned. A second check: synthetically inject a known temporal change (e.g., a simulated snow patch) into the HR target and see whether the FFT and color losses force the model to reproduce the artifact — if they do, the physics-aware losses are matching changed content rather than resolving genuine detail.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If RareFlow holds, the free, 5-day-revisit Sentinel-2 constellation can serve as a proxy source for 2 m-scale geomorphic monitoring, reducing dependence on costly commercial high-resolution acquisitions.
  • On the new RTS benchmark, the paper reports the lowest LPIPS (0.36) and DISTS (0.30) and the lowest FID (116.16)—a 38% reduction over the next-best method, AdcSR at 187.18—with geoscience experts rating outputs perceptually on par with 2 m Maxar ground truth.
  • The architecture transfers to the standard single-image super-resolution setting (4× downsampled HR→HR), where it cuts LPIPS by 27.8% and FID by 41.7% relative to SeeSR while staying within 4% on PSNR, indicating the design is not specific to cross-sensor style transfer.
  • Uncertainty gating gives the model a built-in way to flag out-of-distribution inputs: MC-dropout variance modulates conditioning strength and can directly identify unfamiliar scenes, mitigating feature hallucination.
  • The unified SR-plus-harmonization design means color and brightness statistics are matched to the target sensor within a single model, removing the need for separate radiometric preprocessing pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The uncertainty-gated design suggests a deployment pattern the paper does not develop: run the model on every Sentinel-2 revisit, keep the per-pixel MC-dropout variance as an anomaly layer, and use it to trigger targeted Maxar acquisitions only where uncertainty is high—spending commercial data budgets where they matter most.
  • If the claim that adapters on a frozen backbone suffice in low-data regimes transfers, the same gated-ControlNet recipe could be tried for other scarce scientific imagery (medical slides, archival photographs), where RGB-only data and sparse paired examples are the norm, as the paper itself hints.
  • The ablation tables imply a tunable safety–realism dial: a practitioner could bias the gate toward geometry for hazard mapping or toward realism for visualization, a flexibility the paper reports but does not quantify into a deployment recommendation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. RareFlow is a cross-sensor super-resolution method that translates 10 m Sentinel-2 images into 2 m Maxar-like imagery, targeting rare retrogressive thaw slumps. The architecture combines a frozen SD3 flow-matching backbone with a gated ControlNet for spatial conditioning and text captions for semantic conditioning, plus an MC-dropout uncertainty gate, a frequency-alignment loss, a CIELAB-based color loss, and LPIPS. The authors introduce a new paired RTS benchmark of about 800 images and report state-of-the-art perceptual metrics, including a 38% lower FID than the next-best baseline, and favorable expert evaluation. Ablations isolate the contributions of captions, gating, and losses.

Significance. If the quantitative claims are valid, the paper would be a useful step for generative SR in low-data, cross-sensor remote sensing: the dual-conditioning idea is sensible, the uncertainty-gated injection is novel, and the curated RTS benchmark plus expert evaluation are valuable assets. The paper also reports extensive ablations and uses standard, reproducible metric implementations. However, the central SOTA claim depends on a test-time captioning protocol that is not specified and, as described in the appendix, may leak HR target content into the generative model. Because the main quantitative contribution is the perceptual-quality comparison against baselines, this issue is load-bearing and must be resolved before the SOTA claim can be accepted.

major comments (3)
  1. [§3, §4.2, App. .3, Fig. 8] The paper never specifies how the text conditioning c is obtained at test time. App. .3 states that captions are generated from the HR reference images and shows an instance-level example ('bowl-shaped... dark brown soil... pale sandy streaks sliding downslope toward blue-green water'). If the evaluation uses such HR-derived captions, RareFlow receives direct information about the target content, which would inflate LPIPS/DISTS/FID relative to baselines that do not receive captions. This is not a small omission: Table 3 shows caption-only variants lowering FID from 206 to 145, and the full model to 116, consistent with target leakage. The authors must state exactly how captions are generated at inference (from LR, from fixed templates, or not used), or re-evaluate with an LR-only protocol. Without this, the reported SOTA in Table 1 is unverifiable.
  2. [§4.2, Fig. 3, §4.4] All full-reference metrics (PSNR, SSIM, LPIPS, DISTS) and the FFT/color training losses are computed against GT that the paper itself describes as blurred, compressed, and subject to 'sub-pixel shifts, dramatic variations in illumination, and stark land cover changes.' This makes it hard to know whether the reported gains measure true SR or successful style/texture transfer toward the Maxar distribution. The authors acknowledge this and de-emphasize fidelity metrics, but the central claim of 'super-resolution' and the FID reduction still rely on this GT. I ask for a robustness check: e.g., report metrics on a subset with minimal temporal change, or after registration, or using the synthetic HR-downsampled setting as the primary clean evaluation. This is necessary to certify that the cross-sensor gains are not dominated by temporal-change artifacts.
  3. [Table 1, App. .1] The headline FID is computed over roughly 800 small images (some as small as 30×40 pixels) in a paired benchmark. FID is known to be biased and high-variance at small sample sizes, and using Inception features on such small, non-natural patches further weakens its reliability. Since the '38% FID reduction' is one of the paper's strongest quantitative claims, the authors should provide confidence intervals or an alternative distributional metric (e.g., KID with a larger patch count) to confirm that the gap is not an artifact of the small evaluation set.
minor comments (5)
  1. [Eq. (12)] The loss weights λ_fft, λ_color, λ_lpips are never listed in the experiments. Please report their values and any sensitivity analysis, since the ablation table suggests the full loss is important.
  2. [§3.2, Eq. (5)] The MC-dropout rate p_do, the number of samples T, and the percentile threshold in Eq. (5) are not specified. The uncertainty gate is a distinguishing feature, so these settings should be given.
  3. [Table 3] Rows (3) and (5) have nearly identical LPIPS/DISTS/FID values despite (5) adding the perceptual losses. This is surprising and deserves a sentence of explanation, as it could indicate that the caption dominates the loss or that the added terms have little effect in that configuration.
  4. [App. .3] The example caption is shown next to 'input image given to VLM'; the figure should clarify whether this input is the HR reference or the LR Sentinel-2 image. The current wording implies HR, which is precisely the ambiguity that blocks the main claim.
  5. [Throughout] Several citations are malformed in the text (e.g., 'Wang et al. (2022b). First' and 'Dong et al. (2021); Dou et al. (2020) Second'). A careful proofread is needed.

Circularity Check

1 steps flagged

Target-derived captions are the only specified text-conditioning source; the headline SOTA metrics are therefore forced by HR-reference information at evaluation time.

specific steps
  1. self definitional [Sec. 3.3 (Eq. 7); Appendix .3; Fig. 8]
    "Sec. 3.3: 'For training pairs (x, y), define ŷ = D(f_θ(z_t, t, c, x))'. Appendix .3: 'This section details the complete system prompt provided to the Vision-Language Model GPT-5 API ... to generate descriptive captions from the HR reference images.' Example caption: 'A bowl-shaped thaw slump cuts into a rugged coastal hillside, exposing dark brown soil and crumbly earth, with pale sandy streaks sliding downslope toward blue-green water.'"

    The model output ŷ is conditioned on c (Eq. 7). The only caption-generation procedure described in the paper produces c from the HR reference image y, and no test-time caption source from the LR input is specified; Fig. 8 displays the caption in the evaluation interface. Thus, under the described protocol, test-time SR receives semantic content derived from the target y it is supposed to infer, and the Table 1 metrics compare that target-informed output to y. The reported SOTA gains are not attributable to SR from x alone; they are forced by the target-derived caption, i.e., the prediction is defined using the ground truth it claims to predict.

full rationale

The only load-bearing circularity is the text-conditioning path. The architectural training (Eq. 12) is ordinary supervised flow-matching with auxiliary FFT, perceptual, and color losses against paired HR targets, which is not circular per se. The RTS benchmark reuses public annotations (Yang et al., 2023) and does not import an unverified self-citation. However, because the paper's sole specified caption-generation pipeline (App .3) derives captions from the HR reference images and no inference-time alternative is given, the central quantitative claim in Table 1 rests on target-derived semantic input. The ablation trend (caption alone lowers FID from 187.6 to 145.4; full model to 116.2) is consistent with this leakage. This is a partial circularity of the evaluation rather than of the parameter-fitting equations themselves, so the score is 6 rather than higher.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The framework rests on five domain assumptions listed above, all testable by reimplementation; it introduces no invented physical entities. Six hand-chosen hyperparameters (loss weights, γ, dropout rate, MC samples, κ/τ, s_ctrl) and the normalization percentiles are unreported, hampering exact reproduction.

free parameters (6)
  • Loss weights λ_fft, λ_color, λ_lpips = not reported
    Eq. 12 balances flow-matching, FFT, CIELAB and LPIPS terms; values are hand-chosen and unablated individually (the 'Caption & Loss' row of Table 3 cannot be separated from the gate).
  • Frequency exponent γ in W(ρ)∝ρ^γ = not reported
    Eq. 9 radial emphasis on mid/high frequencies; controls which spectral magnitudes are penalized.
  • MC-dropout rate p_do and sample count T = not reported
    Eq. 4 uncertainty estimate; T and p_do determine the variance u that gates the ControlNet residuals.
  • Uncertainty scale κ (or τ, Eq. 5-6) = not reported
    Maps raw variance to the gate input u∈[0,1]; the scale choice directly sets when the model suppresses semantic guidance.
  • Normalization percentiles p_low/p_high = not reported
    Eqs. 13-15; the paper notes strong sensitivity to the 12-bit normalization choice (Fig. 3) but reports no ablation of percentile values or of the two normalizations.
  • ControlNet residual scale s_ctrl = not reported
    Eq. 3 global strength of the spatial conditioning path; a hand-chosen multiplier on all gated residuals.
axioms (5)
  • domain assumption Cross-temporal Sentinel-2/Maxar pairs, despite sub-pixel shifts, illumination change, and land-cover change, are valid supervision for pixel-aligned SR metrics.
    Fig. 3 and §4 challenge (1) document the misalignment; PSNR/SSIM/LPIPS and the losses (Eqs. 9–12) nevertheless compare output to this target.
  • domain assumption The frozen SD3 MMDiT backbone supplies a useful generative prior for Arctic remote-sensing scenes.
    §3 freezes f_θ; the entire method assumes this natural-image prior, when gated, improves OOD geomorphic synthesis rather than injecting irrelevant textures.
  • domain assumption MC-dropout variance in the trainable ControlNet approximates epistemic uncertainty well enough to modulate semantic guidance.
    §3.2 Eqs. 4-6; only supported in-distribution by Table 4 (NLL, ECE); no held-out OOD calibration experiment is reported.
  • domain assumption GPT-5 captions generated from HR references are scientifically accurate, resolution-aware descriptions of RTS morphology.
    §4.3 Stage 1 checks caption content against an HR reference, not against measured geomorphology; caption accuracy for LR-only inference is untested.
  • domain assumption The SD3 VAE preserves 12-bit radiometry after percentile normalization.
    §3 encodes with a VAE trained on 8-bit natural images; no radiometric validation of the autoencoder is reported.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic-Guided Cross-Sensor Super Resolution of Remote Sensing Images: A Gated Dual Conditioning Flow Matching Model." pith.science (2026). https://pith.science/paper/2KNUS5O5

@misc{pith2026251023816,
  author       = {Pith},
  title        = {Pith review of: Semantic-Guided Cross-Sensor Super Resolution of Remote Sensing Images: A Gated Dual Conditioning Flow Matching Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KNUS5O5}},
  note         = {Machine review of arXiv:2510.23816}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

High spatial resolution satellite imagery is critical for monitoring fine-scale Earth surface processes, but is often limited by cost and revisit time. This work studies cross-sensor super-resolution (SR) to reduce this gap by translating 10 m Sentinel-2 imagery into 2 m Maxar-like imagery in a data-scarce, domain-shifted setting, with a focus on rare geomorphic features such as retrogressive thaw slumps (RTS). We propose RareFlow, a semantic-guided generative AI framework for cross-sensor super-resolution based on a flow-matching formulation, designed to produce visually plausible and physically reliable high-resolution images. RareFlow uses dual conditioning to guide the generation process: (1) a gated ControlNet that preserves scene geometry from low-resolution (LR) input, and (2) text-based semantic guidance that injects contextual information when the target phenomenon is rare. To ensure high-fidelity outputs, we introduce a multifaceted loss function that anchors the output to the high-resolution (HR) ground truth by jointly enforcing frequency alignment, perceptual similarity, and color consistency. RareFlow's performance is systematically evaluated on a newly curated benchmark of multi-sensor satellite imagery for rare Earth feature detection, and its generalizability is demonstrated on two public remote sensing benchmarks, SEN2NAIP and BreizhSR. Human evaluation with domain experts is also conducted to further verify RareFlow's effectiveness in generating high-fidelity super-resolved images for scientific analysis.

Figures

Figures reproduced from arXiv: 2510.23816 by Anna Liljedahl, Chia-Yu Hsu, Forouzan Fallah, Hyunho Lee, Wenwen Li, Yezhou Yang.

Figure 1
Figure 1. Figure 1: When the LR input is blurry or semantically OOD, spatial-only guidance pre￾serves coarse morphology yet remains soft, while semantic-only guidance hallucinates plau￾sible—but incorrect—textures. RareFlow bal￾ances structural evidence from the LR image with textual semantics, suppressing hallucinations and preserving physically consistent geometry and spectra. Confronting this triad of challenges—blurry gui… view at source ↗
Figure 2
Figure 2. Figure 2: (a) The training architecture of RareFlow. The control path (orange) consumes LR latents and caption tokens to produce residual hints ri and predicts per-block scalars α l ∈[0, 1] that scale ri before injection into the frozen backbone (blue) features Fi ← Fi + αiri . (b) ControlNet MM-DiT block internals which produce αi (see Sec. 3.1). faithfulness by stabilizing the denoising trajectory and balancing LR… view at source ↗
Figure 3
Figure 3. Figure 3: Data challenges. Left to right: HR (Maxar), LR (Sentinel-2, Percentile Norm), LR (Sentinel-2, Fixed Norm). Row 1 shows color dis￾crepancies and the effect of normalization. Row 2 shows spatial misalignment. Row 3 shows tempo￾ral misalignment due to changes in snow cover. These challenges include: (1) Spatiotem￾poral misalignment between LR and HR images, acquired at different times, causes sub-pixel shifts… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on paired LR–HR data. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual Comparison of model variants. Our analysis of [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of input image given to VLM to gen￾erate an RTS-aware caption. An example of a caption generated for the given input image ( [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on HR-downsampled [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Our custom web-based interface for human evaluation. [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Asynchronous Remote Sensing Time-Series Fusion for Cloud Removal and Anytime Reconstruction

    cs.CV 2026-05 unverdicted novelty 7.0

    AGFlow is a spatiotemporal flow-matching model for asynchronous S1/S2 fusion that enables cloud removal, missing-frame reconstruction, and on-demand synthesis at user-specified timestamps, with 16-19% MAE/RMSE gains o...

  2. Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems

    cs.LG 2026-05 unverdicted novelty 6.0

    Distilled one-step consistency model from optimal-transport flow-matching teacher reconstructs high-fidelity dynamical system flows from low-fidelity data with 12x speedup, half the parameters, and 23.1% better SSIM t...

  3. Spatial Representation Learning Beyond Pixels: Unifying Raster Data and Vector Semantics for Human-Centric Geospatial Foundation Models

    cs.AI 2026-06 unverdicted novelty 3.0

    Perspective paper calling for unified spatial representation learning that integrates raster imagery with vector semantics in geospatial foundation models.

Reference graph

Works this paper leans on

24 extracted references · 3 canonical work pages · cited by 3 Pith papers

  1. [1]

    Comparison of pansharpening algorithms: Outcome of the 2006 grs-s data-fusion contest

    Luciano Alparone, Lucien Wald, Jocelyn Chanussot, Claire Thomas, Paolo Gamba, and Lori Mann Bruce. Comparison of pansharpening algorithms: Outcome of the 2006 grs-s data-fusion contest. IEEE Transactions on Geoscience and Remote Sensing, 45(10):3012–3021,

    2006

  2. [2]

    do no spectral harm

    ,(19) whereµ,σ 2, andσ xy are local (Gaussian-windowed) statistics;C 1, C2 stabilize division. While these methods improved the evaluation of structural integrity, they still couldn’t fully capture the complex textures and nuanced details that make an image look realistic. This limitation paved the way for learned perceptual metrics. The Learned Perceptua...

    2018

  3. [6]

    Antoni G Lewkowicz and Robert G Way

    doi: 10.1109/CVPR.2017.19. Antoni G Lewkowicz and Robert G Way. Extremes of summer climate trigger thousands of thermokarst landslides in a high arctic environment.Nature communications, 10(1):1329,

    doi:10.1109/cvpr.2017.19 2017

  4. [7]

    URLhttps://www.mdpi.com/2072-4292/17/5/927

    doi: 10.3390/rs17050927. URLhttps://www.mdpi.com/2072-4292/17/5/927. Qinwei Lin, Xiaopeng Sun, Yu Gao, Yujie Zhong, Dengjie Li, Zheng Zhao, and Haoqian Wang. TASR: Timestep-aware diffusion model for image super-resolution.arXiv preprint arXiv:2412.03355,

    Pith/arXiv arXiv 2072

  5. [10]

    Satdiffmoe: A mixture of estimation method for satellite image super-resolution with latent diffusion models.arXiv preprint arXiv:2406.10225,

    Zhaoxu Luo, Bowen Song, and Liyue Shen. Satdiffmoe: A mixture of estimation method for satellite image super-resolution with latent diffusion models.arXiv preprint arXiv:2406.10225,

    Pith/arXiv arXiv

  6. [11]

    completely blind

    URLhttps://resources.maxar.com/ data-sheets/worldview-3. Accessed: 2025-09-24. Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer.IEEE Signal processing letters, 20(3):209–212,

    2025

  7. [12]

    URLhttps://tc

    doi: 10.5194/tc-18-4787-2024. URLhttps://tc. copernicus.org/articles/18/4787/2024/. Ingmar Nitze, Konrad Heidler, Nina Nesterova, Jonas K¨upper, Emma Sch¨utt, Tobias H¨olzer, Sophia Barth, Mark J Lara, Anna K Liljedahl, and Guido Grosse. Darts: Multi-year database of ai- detected retrogressive thaw slumps in the circum-arctic permafrost region.Scientific ...

    work page doi:10.5194/tc-18-4787-2024 2024

  8. [14]

    doi: 10.1109/TPAMI.2022.3204461. Mehdi S. M. Sajjadi, Bernhard Sch ¨olkopf, and Michael Hirsch. Enhancenet: Single image super- resolution through automated texture synthesis. InIEEE International Conference on Computer Vision (ICCV),

    arXiv 2022

  9. [15]

    DiffFuSR: Super-resolution of all sentinel-2 multispectral bands using diffusion models.arXiv preprint arXiv:2506.11764,

    Muhammad Sarmad, Arnt-Børre Salberg, and Michael Kampffmeyer. DiffFuSR: Super-resolution of all sentinel-2 multispectral bands using diffusion models.arXiv preprint arXiv:2506.11764,

    arXiv

  10. [16]

    ControlSR: Taming diffusion models for consistent real-world image super resolution.arXiv preprint arXiv:2410.14279,

    Yuhao Wan, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, Ming-Ming Cheng, and Bo Li. ControlSR: Taming diffusion models for consistent real-world image super resolution.arXiv preprint arXiv:2410.14279,

    Pith/arXiv arXiv

  11. [17]

    Sam-diffsr: Structure-modulated diffusion model for image super-resolution.arXiv preprint arXiv:2402.17133,

    Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image super-resolution.arXiv preprint arXiv:2402.17133,

    Pith/arXiv arXiv

  12. [18]

    Cross-sensor do- main adaptation for high spatial resolution urban land-cover mapping: From airborne to space- borne imagery.Remote Sensing of Environment, 277:113058, 2022a

    Junjue Wang, Ailong Ma, Yanfei Zhong, Zhuo Zheng, and Liangpei Zhang. Cross-sensor do- main adaptation for high spatial resolution urban land-cover mapping: From airborne to space- borne imagery.Remote Sensing of Environment, 277:113058, 2022a. doi: 10.1016/j.rse. 2022.113058. URLhttps://www.sciencedirect.com/science/article/pii/ S0034425722001729. Xintao...

    arXiv 2022

  13. [20]

    Se- mantic segmentation prior for diffusion-based real-world super-resolution.arXiv preprint arXiv:2412.02960,

    Jiahua Xiao, Jiawei Zhang, Dongqing Zou, Xiaodan Zhang, Jimmy Ren, and Xing Wei. Se- mantic segmentation prior for diffusion-based real-world super-resolution.arXiv preprint arXiv:2412.02960,

    Pith/arXiv arXiv

  14. [21]

    URL https://www.nature.com/articles/s41598-024-74781-y

    doi: 10.1038/s41598-024-74781-y. URL https://www.nature.com/articles/s41598-024-74781-y. Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment.IEEE transactions on Image Processing, 20(8):2378–2386,

    work page doi:10.1038/s41598-024-74781-y

  15. [23]

    APPENDIX .1 DATASETS& CHALLENGES The dataset covers seven Arctic regions, including the Yamal and Gydan Peninsulas, Lena River, and Kolguev Island in Russia, along with Herschel Island, Horton Delta, Tuktoyaktuk Peninsula, and Banks Island in Canada Yang et al. (2023). These sites provide diverse land-cover characteristics, including tundra, ice-rich perm...

    2023

  16. [2011]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543,

    Pith/arXiv arXiv

  17. [2016]

    doi: 10.1007/978-3-319-46475-6

    doi:10.1007/978-3-319-46475-6

  18. [2017]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter

    doi: 10.3390/rs9060584. Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30,

    work page doi:10.3390/rs9060584

  19. [2018]

    A review of image super-resolution approaches based on deep learning and applications in remote sensing.Remote Sensing, 14(21):5423, 2022b

    Xuan Wang, Jinglei Yi, Jian Guo, Yongchao Song, Jun Lyu, Jindong Xu, Weiqing Yan, Jindong Zhao, Qing Cai, and Haigen Min. A review of image super-resolution approaches based on deep learning and applications in remote sensing.Remote Sensing, 14(21):5423, 2022b. Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion n...

    Pith/arXiv arXiv

  20. [2020]

    S2 mission — sentiwiki (psf/mtf information).https:// sentiwiki.copernicus.eu/web/s2-mission

    10 Under review European Space Agency (ESA). S2 mission — sentiwiki (psf/mtf information).https:// sentiwiki.copernicus.eu/web/s2-mission. Accessed: 2025-09-07. Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion En- glish, Kyl...

    Pith/arXiv arXiv 2025

  21. [2022]

    Semantic-guided diffusion model for single-step image super-resolution.arXiv preprint arXiv:2505.07071,

    Zihang Liu, Zhenyu Zhang, and Hao Tang. Semantic-guided diffusion model for single-step image super-resolution.arXiv preprint arXiv:2505.07071,

    Pith/arXiv arXiv

  22. [2023]

    Advancing image super-resolution techniques in remote sensing: A comprehensive survey.arXiv preprint arXiv:2505.23248,

    Yunliang Qi, Meng Lou, Yimin Liu, Lu Li, Zhen Yang, and Wen Nie. Advancing image super-resolution techniques in remote sensing: A comprehensive survey.arXiv preprint arXiv:2505.23248,

    arXiv

  23. [2024]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Pith/arXiv arXiv

  24. [2025]

    Accessed: 2025-09-24

    URL https://www.esa.int/Applications/Observing_the_Earth/Copernicus/ Sentinel-2. Accessed: 2025-09-24. Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. InICCV,

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.