From Sentinel-2 to 1-m Imagery: Real-World $\times 10$ Cross-Sensor Super-Resolution via Visual Autoregressive Modeling

Tailai Chen; Yuluxin Fu; Jie Deng; Shu Luo; Yanan Wu; Danfeng Hong

cs.CVOct 26, 2026

From Sentinel-2 to 1-m Imagery: Real-World $\times 10$ Cross-Sensor Super-Resolution via Visual Autoregressive Modeling

Authors: Tailai Chen, Yuluxin Fu, Jie Deng, Shu Luo, Yanan Wu, Danfeng Hong

Organizations: College of Water Resources and Intelligence Engineering, China Agricultural University, Beijing 100083, China · College of Information and Electrical Engineering, China Agricultural University, Beijing 100083, China · Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · University of Chinese Academy of Sciences, Beijing 100049, China ·

Abstract

Reconstructing 1-m imagery from Sentinel-2 combines two difficulties. The input and the reference are observed by different satellites, which differ in spatial resolution, spectral response, radiometry, acquisition time, and local geometry, so no synthetic degradation operator links them. The required magnification of ten in each spatial dimension also far exceeds the factors of two to four at which remote sensing super-resolution is usually developed. We study this real-world $\times 10$ problem using YREB-SR, which pairs 2022 Sentinel-2 RGB composites with 2022 Jilin-1 RGB references resampled to a 1-m grid over the Yangtze River Economic Belt. The benchmark contains 11,559 geographically matched scene pairs, with 6,409, 2,135, and 3,015 assigned to training, validation, and testing, respectively. We develop CrossVAR, a visual autoregressive model that predicts the 1-m latent as a nine-scale hierarchy conditioned on the Sentinel-2 input and estimates the remaining quantization residual with a diffusion refiner run without injected noise. Complete 1600 $\times$ 1600 scenes are reconstructed in overlapping windows that start on the Sentinel-2 pixel grid, as every training patch does, and every window is decoded at the size on which the decoder was trained before the windows are blended with Gaussian weights. On all 3,015 test scenes, CrossVAR achieves 27.60 dB PSNR, 0.7725 SSIM, 0.1451 LPIPS, and 7.15 FID, the best value of each metric among the compared methods. These results establish a strong autoregressive baseline for large-scale cross-sensor reconstruction. To facilitate reproducibility, the source code and implementation are publicly available at https://github.com/SHANOAL/RSSR.

Figures & tables

Fig. 1: PSNR and model time per complete

1600\times 1600

test scene of YREB-SR. Marker area grows with the number of parameters.

Fig. 2: Spatial distribution of the 11,559 cross-sensor scene pairs in YREB-SR across the Yangtze River Economic Belt. Colors indicate the training, validation, and test subsets, while blue lines represent the Yangtze River network.

Attribute	Setting
Cross-sensor image pairs
Sensor pair	Sentinel-2 RGB $\rightarrow$ 1-m RGB reference
Spectral channels	B4–B3–B2 $\rightarrow$ R–G–B
Nominal GSD	10 m $\rightarrow$ 1 m ( $\times 10$ )
Full-scene grids	$160\times 160\rightarrow 1600\times 1600$ pixels
Training grids	$40\times 40\rightarrow 400\times 400$ pixels

TABLE I: Summary of the YREB-SR Benchmark

Fig. 3: Overview of CrossVAR. Panels 1 to 3 reconstruct one window. The condition encoder maps the Sentinel-2 window, resized to the output grid, to one condition token per latent position, and the next-scale transformer predicts nine code maps from $1\times 1$ to $25\times 25$ . Deterministic residual refinement (DRR) estimates the residual of the codes without injected noise, and the frozen VQ decoder outputs the 1-m window. Aligned window decoding (AWD) in panel 4 starts every window of a complete scene on the input pixel grid, reconstructs it through panels 1 to 3, and blends the decoded windows with Gaussian weights.

Method	Batch	Steps
EDSR-baseline	16	150k
SwinIR	8	150k
TTST	4	150k
Real-ESRGAN+	4	150k

TABLE II: Training Budgets of the Retrained Baselines

Reconstruction and continuous-representation methods
Method	Representation	PSNR (dB) $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
Bicubic	Interpolation	14.5051	0.4626	0.9534	215.9498
EDSR-baseline [ 10 ]	CNN	24.0986	0.6596	0.6831	68.2941
SwinIR [ 11 ]	Transformer	24.5078	0.6706	0.6371	58.0257
TTST [ 12 ]	RS transformer	24.7157	0.6739	0.6201	48.9027
Thera [ 19 ]	Neural field	25.2978	0.6855	0.5679	37.8334

TABLE III: Full-Scene

\times 10

Reconstruction on YREB-SR. All Reported Scores Use the Same 3,015 Test Pairs. Best and Second-Best Results Are Bold and Underlined.

Baseline	$\Delta$ PSNR (dB)	$\Delta$ SSIM	$\Delta$ LPIPS	Better (%)
EDSR-baseline	+3.50	+0.113	$-$ 0.538	100
SwinIR	+3.09	+0.102	$-$ 0.492	100
TTST	+2.89	+0.099	$-$ 0.475	100
Thera	+2.30	+0.087	$-$ 0.423	100
Real-ESRGAN+	+5.64	+0.155	$-$ 0.182	100
Satlas-SR	+4.30	+0.140	$-$ 0.070	100

TABLE IV: Scene-Wise Comparison of CrossVAR with the Retrained Baselines on the 3,015 Test Scenes. Differences Are CrossVAR Minus Baseline, and the Last Column Gives the Share of Scenes on Which CrossVAR Is Better in All Three Metrics.

Fig. 4: Scene-wise analysis on the 3,015 test scenes. (a) Radially averaged power spectrum of the reconstructions relative to the Jilin-1 reference. The dotted line marks the Nyquist frequency of the Sentinel-2 input. (b) Mean PSNR in quartiles of the mean gradient magnitude of the reference. (c) PSNR gain of CrossVAR over Thera at the scene centers.

Method	Parameters (M)	Precision	Model time	I/O-inclusive	Peak memory
Bicubic	–	CPU	26.44	127.91	–
EDSR-baseline	1.39	FP32	2.92	384.07	124
SwinIR	11.99	FP32	106.41	440.12	499
TTST	21.76	FP32	189.63	503.57	1,632
Thera	5.85	FP32	303.50	717.46	4,534
Real-ESRGAN+	20.32	FP32	44.96	446.78	2,283

TABLE V: Full-Scene Efficiency. Times Are Milliseconds per Scene and Memory Is Peak MiB.

Fig. 5: Coarse-to-fine reconstruction. (a) Four test windows decoded after the $13\times 13$ , $18\times 18$ , and $25\times 25$ code maps and after deterministic residual refinement, with PSNR against the Jilin-1 reference. (b) Mean PSNR after every scale on 427 validation scenes, with the interquartile range shaded.

Panel of Figure 11 — Fig. 6: Complete-scene and local-detail comparisons on four test scenes. Seven learned methods are shown between the Sentinel-2 input and the Jilin-1 reference, and red and blue boxes mark the enlarged regions. Only the Sentinel-2 input is contrast-stretched for display.

Fig. 7: PSNR change of single windows on 427 validation scenes when the window start is moved away from the Sentinel-2 pixel grid. The outlined cell is the aligned start.

DRR	AWD	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
		25.93	0.6960	0.1796	10.29
✓		26.27	0.7076	0.1695	9.81
	✓	27.01	0.7509	0.1539	6.38
✓	✓	27.60	0.7725	0.1451	7.15

TABLE VI: Ablation of Deterministic Residual Refinement (DRR) and Aligned Window Decoding (AWD) on the 3,015 Test Scenes

Windows	Fusion	Decoding	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
On grid	None	Scene	26.39	0.7406	0.1804	10.98
Off grid	Latent	Scene	26.27	0.7076	0.1695	9.81
Off grid	Latent	Window	26.51	0.7145	0.1595	8.94
Off grid	RGB	Window	26.45	0.7131	0.1558	7.96
On grid	Latent	Scene	27.26	0.7645	0.1609	9.08
On grid	Latent	Window	27.58	0.7708	0.1506	8.21

TABLE VII: Complete-Scene Assembly Rules with

\tau=0

on the 3,015 Test Scenes

Remote Sensing Image Super-Resolution

Remote sensing image super-resolution reconstructs fine spatial detail in satellite and aerial imagery from lower-resolution observations.

See all 35 papers

Mind the Gap: Quantifying the Domain Gap in Cross-Sensor Diffusion Super-Resolution
Jun 26, 2026Dawid Kopeć, Katarzyna Jabłońska, Wojciech Kozłowski +1Synthetic-to-Real Domain Adaptation Diffusion Models for Image Restoration
WUST, Wybrzeże Stanisława Wyspiańskiego 27, 50-370 Wrocław, Poland · Tooploox, Tęczowa 7, 53-601 Wrocław, Poland
Flow matching for Sentinel-2 super-resolution: implementation, application, and implications
May 1, 2026Dakota Hester, Vitor S. Martins, Lucas B. Ferreira +2Flow Matching Land Cover Classification
Department of Agricultural and Biological Engineering, Mississippi State University · North Mississippi Research and Extension Center, Mississippi State University
InfScene-SR: Seamless Super-Resolution of Arbitrarily Large Remote-Sensing Scenes via Variance-Preserving Joint Denoising
Feb 23, 2026Shoukun Sun, Zhe Wang, Xiang Que +2Image Super-Resolution Remote Sensing Image Super-Resolution
University of Idaho, Moscow, Idaho, USA · NCEAS, UC Santa Barbara, Santa Barbara, California, USA · Fujian Agriculture and Forestry University, Fuzhou, China

Jun 26, 2026J/K move · Enter open · S save

Mind the Gap: Quantifying the Domain Gap in Cross-Sensor Diffusion Super-Resolution

Dawid Kopeć, Katarzyna Jabłońska, Wojciech Kozłowski, Maciej Zięba

WUST, Wybrzeże Stanisława Wyspiańskiego 27, 50-370 Wrocław, Poland · Tooploox, Tęczowa 7, 53-601 Wrocław, Poland

Synthetic-to-Real Domain Adaptation Diffusion Models for Image Restoration Remote Sensing Image Super-Resolution Image Super-Resolution

Open paper AskPDF arXiv

Demand for high-resolution satellite imagery has increased interest in super-resolution (SR) to bridge the spatial resolution gap between freely available missions such as Sentinel-2 and commercial systems like PlanetScope. Because no sensor provides true paired low- and high-resolution observations, SR models are usually trained on synthetically degraded data, creating a domain gap on real cross-sensor imagery. In this work, we provide the first systematic study of how this synthetic-to-real mismatch affects the performance of modern diffusion-based SR models. Using a large, geometrically and temporally aligned dataset of Sentinel-2 and PlanetScope imagery, we evaluate five state-of-the-art diffusion architectures under controlled experimental settings. We also introduce LPIPS-Sat, a domain-adapted perceptual metric based on Sentinel-2 self-supervised features. Our results show two persistent challenges: synthetically trained models degrade sharply on real pairs, while models trained on real cross-sensor data exhibit optimisation difficulties and struggle to adapt to the physical and radiometric diversity. These findings highlight a key limitation of current SR and motivate methods that disentangle super-resolution from domain adaptation.

Abstract

Remote Sensing Image Super-Resolution

Explore similar work

Mind the Gap: Quantifying the Domain Gap in Cross-Sensor Diffusion Super-Resolution

Flow matching for Sentinel-2 super-resolution: implementation, application, and implications

InfScene-SR: Seamless Super-Resolution of Arbitrarily Large Remote-Sensing Scenes via Variance-Preserving Joint Denoising