Although remote sensing image arbitrary-scale super-resolution (RSI-ASSR) methods support reconstruction at arbitrary or continuous scales, they predominantly condition on the relative scale factor and have not sufficiently investigated the absolute spatial resolution itself as a reconstruction prior, thereby disregarding the inherent spatial context of objects, which frequently leads to spatial-related distortions and compromised performance in downstream remote sensing applications. To address these challenges, we introduce SpatialSR, a SRe-centric arbitrary-scale RSISR method that leverages the spatial resolution (SRe) of remote sensing images (RSI) as an essential prior for super-resolution reconstruction. Distinct from existing SR methods, SpatialSR aims to restore high-resolution remote sensing images with a target SRe (typically set to 0.1m/pixel to align with the metadata of commonly used downstream task datasets). Specifically, we design a Spatial Resolution Estimation Module (SREM) to estimate the SRe of low-resolution (LR) images, thereby dynamically determining the required arbitrary scaling factor. Next, we employ a pretrained SRe-to-image diffusion model to extract SRe-specified texture-style codes, which are further incorporated into a carefully designed Resolution-aware Style Consistency Module (RSCM) to constrain the SRe-aware texture consistency of reconstructed images. Furthermore, to overcome the high computational expense inherent in diffusion-based SR methods, we wrap the iterative diffusion process as a Trajectory Consistency (TC) loss, which enables comparable reconstructed texture quality of the generative model in SpatialSR within a single step. Extensive experiments demonstrate that the proposed SpatialSR achieves state-of-the-art (SOTA) performance on commonly used RSISR and downstream (image classification and segmentation tasks) datasets.
Figures & tables
Figure 1: Illustration of gradient statistics and downstream task performance at different spatial resolutions (SRe). (a) Comparison of average radial Fourier magnitude spectra under two spatial resolution settings. The left subplot compares images at 0.1 m/px with their 4× downsampled counterparts at 0.4 m/px, while the right subplot compares images at 1.0 m/px with their 4× downsampled counterparts at 4.0 m/px. For each SRe, the curve represents the mean radially averaged log-amplitude spectrum computed over all images, with the horizontal axis indicating normalized spatial frequency. (b) Visual and quantitative comparisons of ×4 super-resolution methods at different SRe.
Figure 2: Overview of the proposed SpatialSR pipeline. Our pipeline consists of three core components. (a) The SREM estimates the source SRe from the LR image and then computes the scale factor together with the target SRe (Section 3.2 ). (b) The RSCM synthesizes the final image under the modulation of a style embedding derived from the SRe-to-img model (Section 3.3 ). (c) Finally, the TC loss refines texture details by aligning the diffusion trajectories (Section 3.4 ).
Figure 3: Resolution-aware Style Consistency Modulation. RSCM injects resolution-specific style information by modulating the decoder features. The style embedding is projected into scaling ( γ ) and bias ( β ) parameters, which are used to perform a residual affine transformation on the feature map.
Figure 4: Illustration of TC loss . Both the SR image X^0 and its ground-truth (GT) counterpart X0 are passed through the denoising process of a pretrained diffusion model. The TC loss minimizes the discrepancy between their denoising trajectories, thereby encouraging the SR image to serve as a more realistic input to the diffusion model.
Dataset
TSR
CSR
Metric
Generative SR
Implicit Neural Representation-based SR
ESRGAN
IDM
LDM
StableSR
ResShift
LIIF
ContSR
AnySR
SRNO
LMF
HIIF
Ours
LR
Vaihingen
0.1
0.5
PSNR↑
22.84
24.42
24.56
25.94
25.71
25.89
25.18
26.14
25.26
26.34
25.35
26.21
SSIM↑
0.6107
0.6873
0.6851
0.7589
0.7787
0.7795
0.7644
0.7841
0.7553
0.7851
0.7699
0.7913
LPIPS↓
0.3535
0.3583
0.3908
0.2479
0.2981
0.2687
0.2728
0.2744
0.2744
0.2746
0.2645
0.2606
DISTS↓
0.2011
0.2371
0.2606
0.1717
0.1936
0.1888
0.1898
0.1879
0.1879
0.1908
0.1942
0.1875
FID↓
99.78
96.26
119.89
57.03
58.86
67.18
69.23
64.95
76.74
66.23
73.80
58.44
Table 1: Quantitative comparison results of SR and downstream task performance on bicubic-degraded synthetic datasets.
Dataset
TSR
CSR
Metric
Generative SR
Implicit Neural Representation-based SR
ESRGAN
IDM
LDM
StableSR
ResShift
LIIF
ContSR
AnySR
SRNO
LMF
HIIF
Ours
Ori
UAVid
0.1
0.16
NIQE↓
5.11
6.10
4.99
5.34
4.74
5.54
5.22
5.53
5.39
5.16
5.44
4.92
MUSIQ↑
60.6
52.6
51.8
63.0
64.4
59.4
60.9
58.6
60.6
60.0
60.6
64.7
MANIQA↑
0.42
0.39
0.53
0.61
0.58
0.64
0.66
0.67
0.68
0.65
0.64
0.69
CLIPIQA↑
0.50
0.28
0.28
0.47
0.54
0.38
0.36
0.39
0.37
0.35
0.37
0.48
mIoU↑
72.2
71.7
72.2
74.6
74.7
73.4
73.5
73.4
73.4
73.6
73.4
75.8
72.5
Table 2: Quantitative comparison results of remote-sensing image SR under real-world scenarios.
Figure 5: Qualitative comparisons of super-resolution and downstream task results under bicubic downsampling. From top to bottom, the rows correspond to spatial resolution enhancement from 4 m/px to 0.5 m/px ( ×10 ) and from 1.0 m/px to 0.1 m/px ( ×5 ), respectively.
TSR
CSR
Method
LoveDA
PSNR ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
mIoU ↑
0.5
2
baseline
29.89
0.3057
6.49
42.16
30.2
Ours + SREM
30.07
0.2953
6.20
44.63
31.5
Ours + RSCM
30.01
0.2801
6.01
45.58
32.5
Ours + TC
30.19
0.2869
5.95
45.86
32.9
Ours (SpatialSR)
30.25
0.2798
5.84
46.08
34.3
Table 3: Ablation study of SpatialSR on the LoveDA dataset. We separately apply the proposed SREM, RSCM, and TC loss to the LMF baseline, which serves as the shared backbone for fair comparison.
TSR
CSR
Method
LoveDA
PSNR ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
mIoU ↑
0.5
2
SREM + Random sample
29.83
0.3156
6.32
42.89
30.7
SREM + Laplacian sample
29.90
0.3020
6.21
44.16
30.8
SREM + Haar sample
30.01
0.3087
6.18
44.99
31.3
SREM + Hybrid sample (ours)
30.07
0.2953
6.20
44.63
31.5
4
SREM + Random sample
24.76
0.5875
8.07
30.45
24.2
Table 4: Ablation study of SREM with different sampling strategies.
Target SRe
Current SRe
Method
LoveDa
PSNR ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
mIoU ↑
0.5
2
RSCM (random SRe)
29.95
0.3056
6.68
42.89
31.2
RSCM (w/o SRe)
29.84
0.3102
6.36
44.21
32.5
RSCM (correct SRe)
30.01
0.2801
6.01
45.58
32.5
4
RSCM (random SRe)
24.95
0.5956
8.10
34.62
26.7
RSCM (w/o SRe)
24.82
0.5634
7.99
35.81
27.1
Table 5: Ablation study of RSCM under different SRe style conditions.
Target SRe
Current SRe
Method
LoveDa
PSNR ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
mIoU ↑
0.5
2
Pixel loss
30.09
0.3079
6.69
43.99
30.6
Pixel + LPIPS
29.82
0.2799
6.22
44.01
31.5
Pixel + TC (Ours)
30.19
0.2869
5.95
45.86
32.9
4
Pixel loss
25.10
0.5921
7.99
33.02
24.9
Pixel + LPIPS
24.82
0.5266
7.85
33.11
26.1
Table 6: Ablation study of the TC loss.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
λ
PSNR ↑
SSIM ↑
Score ↑
0
28.9803
0.7842
2.6752
0.05
28.9543
0.7759
2.6651
0.1
28.9217
0.7683
2.8914
0.2
28.8542
0.7628
2.9286
0.3
28.8358
0.7616
3.0205
0.4
28.7858
0.7582
3.0857
Appendix
Table 7: Ablation study on the TC loss weight λ . Evaluated on UCMD for 4 × SR (1.2 m/px → 0.3 m/px).
Figure 6: Qualitative comparisons of super-resolution and downstream task results under bicubic downsampling. From top to bottom, the rows correspond to spatial resolution enhancement from 2.0 m/px to 0.5 m/px ( ×4 ) ,from 4.0 m/px to 0.5 m/px ( ×8 ) and from 0.5 m/px to 0.1 m/px ( ×5 ), respectively.
Dataset
TSR
CSR
Metric
Generative SR
Implicit Neural Representation-based SR
ESRGAN
IDM
LDM
StableSR
ResShift
LIIF
ContSR
AnySR
SRNO
LMF
HIIF
Ours
LR
Potsdam
0.1
0.5
PSNR↑
26.91
26.16
28.80
28.14
30.22
32.69
32.32
32.71
31.70
33.05
32.45
33.22
SSIM↑
0.6185
0.6160
0.7735
0.7652
0.8065
0.8767
0.8742
0.8756
0.8678
0.8779
0.8746
0.8955
LPIPS↓
0.3926
0.4965
0.3752
0.3012
0.2436
0.2734
0.2732
0.2734
0.2751
0.2726
0.2677
0.2591
DISTS↓
0.1474
0.2514
0.1929
0.1526
0.1557
0.1256
0.1261
0.1222
0.1285
0.1276
0.1239
0.1183
FID↓
10.64
47.85
37.09
32.06
17.97
4.19
4.32
4.15
5.40
4.25
4.39
3.74
Appendix
Table 8: Quantitative comparison results of SR and segmentation performance on bicubic-degraded synthetic datasets. The best and second-best results are shown in bold and underlined, respectively.
128→512
ESRGAN
IDM
LDM
StableSR
ResShift
LIIF
ContSR
AnySR
SRNO
LMF
HIIF
Ours
Trainable Params(M)
16.7
116.6
123.6
150
40.6
1.57
1.71
6.43
2.02
1.39
2.96
1.76
Inference Time(S)
0.033
35.87
27.92
25.91
6.67
0.1084
0.1921
0.0638
0.056
0.0062
0.2041
0.0074
MACs(G)
293.71
3920.41
3918.92
4087.52
454.65
382.62
153.2
105.42
231.3
24.37
440.1
25.76
FLOPs(G)
587.87
7944.61
7936.11
8175.33
808.82
766.63
307.72
211.58
463.9
49.16
885.5
51.96
Inference Step
/
150
150
150
4
/
/
/
/
/
/
/
Appendix
Table 9: Complexity comparison among different methods. ContinuousSR is abbreviated as ContSR. For our method, the reported inference time includes the sliding-window cropping and patch-selection steps of SREM.
Figure 7: Ablation study of different sampling strategies under varying spatial-resolution ranges. (a) MAE, (b) R2 , and (c) Accuracy within ±10%. The horizontal axis denotes the spatial-resolution range (m/px). Hybrid sampling consistently achieves the best performance across all evaluation metrics.
Figure 8: Visualizations of sampling results under different sampling strategies in the SREM for SRe estimation. The Random strategy produces samples dominated by low-frequency information, thereby increasing the difficulty of resolution estimation. Although the Laplacian and Haar strategies capture more high-frequency details, they fail to preserve the global characteristics of spatial resolution, leading to reduced estimation accuracy.
Train
Test
MAE ↓
RMSE ↓
R2↑
Accuracy ↑
± 5%
± 10%
MSRUAV (UAV)
MSRUAV (UAV)
0.0312
0.0401
0.94
87.6
96.8
MSRUAV (UAV)
UCMD (satellite)
0.0487
0.0623
0.88
74.3
89.5
UCMD-urban
UCMD-urban
0.0298
0.0387
0.95
89.1
97.2
UCMD-urban
UCMD-forest
0.0441
0.0573
0.90
79.8
92.4
Appendix
Table 10: Cross-scenario and cross-domain generalization of SREM. Within-domain settings serve as upper-bound references.
Figure 9: Visualizations of the same scene across continuous spatial resolutions and their corresponding diffusion timesteps T . The visualization at T=45 preserves both texture details and weak semantic structures, making its latent code the most suitable for style modulation.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
FID ↓
(a) Pretrained CNN feature
32.81
0.885
0.271
0.126
4.18
(a) CLIP feature
32.74
0.882
0.274
0.128
4.23
(b) Misaligned style
32.69
0.879
0.283
0.133
4.41
(c) w/o SRe fine-tuning
32.77
0.883
0.275
0.129
4.31
(d) GT SRe + diffusion
33.29
0.896
0.256
0.117
3.70
Ours (est. SRe + diffusion)
33.22
0.895
0.259
0.118
3.74
Appendix
Table 11: Ablation study on the diffusion prior in RSCM. “GT SRe” uses the ground-truth spatial resolution; “est. SRe” uses the SREM prediction.
Generative diffusion priors have recently achieved state-of-the-art performance in natural image super-resolution, demonstrating a powerful capability to synthesize photorealistic details. However, their direct application to remote sensing image super-resolution (RSISR) reveals significant shortcomings. Unlike natural images, remote sensing images exhibit a unique texture distribution where ground objects are globally stochastic yet locally clustered, leading to highly imbalanced textures. This imbalance severely hinders the model's spatial perception. To address this, we propose TexADiff, a novel framework that begins by estimating a Relative Texture Density Map (RTDM) to represent the texture distribution. TexADiff then leverages this RTDM in three synergistic ways: as an explicit spatial conditioning to guide the diffusion process, as a loss modulation term to prioritize texture-rich regions, and as a dynamic adapter for the sampling schedule. These modifications are designed to endow the model with explicit texture-aware capabilities. Experiments demonstrate that TexADiff achieves superior or competitive quantitative metrics. Furthermore, qualitative results show that our model generates faithful high-frequency details while effectively suppressing texture hallucinations. This improved reconstruction quality also results in significant gains in downstream task performance. The source code of our method can be found at https://github.com/ZezFuture/TexAdiff.
High-resolution remote sensing images (RSIs) are crucial for Earth observation applications, yet acquiring them is often limited by sensor constraints and costs. In recent years, generative super-resolution (SR) methods, particularly diffusion models, have made significant progress. However, they typically require slow iterative inference with 40--1000 steps and exhibit limited flexibility in continuous-scale SR settings. To address these issues, we propose FlowGS, a generative reconstruction framework for arbitrary-scale SR of RSIs. FlowGS models the high-frequency detail representations between high- and low-resolution images and learns a continuous probability flow from noise to detail priors via flow matching (FM) constrained by shortcut consistency, thereby reducing generative complexity and improving inference efficiency. Additionally, we employ 2D Gaussian splatting to construct a continuous feature field, thereby enabling flexible reconstruction at arbitrary query locations. Experimental results show that FlowGS delivers competitive perceptual quality compared with existing methods in both continuous-scale and fixed-scale SR settings, with substantially improved inference efficiency.
Jiangwei Mo, Xi Lu, Hanlin Wu
School of Information Science and Technology, Beijing Foreign Studies University, Beijing 100089, China
Diffusion-based methods demonstrate significant potential for remote sensing image super-resolution at large scaling factors, particularly in reference-based super-resolution (RefSR), where high-resolution reference images provide critical fine-grained texture priors. However, existing methods often suffer from a trade-off between over-reliance on reference information, which leads to texture artifacts, and under-utilization of such information, which results in insufficient detail recovery. To address these issues, we propose DS-DiT, a Decoupled Siamese Diffusion Transformer that decouples the interaction between low-resolution (LR) and reference (Ref) conditions within the attention mechanism. By allowing LR structural priors and Ref texture information to independently interact with the noisy latent, the framework effectively mitigates competition between the two conditional sources. To further compensate for the limited local modeling ability of global attention, we introduce a Patch-Level Weighting (PLW) module that adaptively modulates the fusion of conditional sources. In addition, the siamese architecture enables an inference-time autoguidance strategy that exploits the prediction discrepancy between strong and weak Ref conditions to improve generation quality without additional training. Experimental results across multiple datasets and scaling factors show that DS-DiT outperforms existing methods in both quantitative metrics and visual fidelity.
Bin Luo, Runmin Dong, Zhaoyang Luo +4
Tsinghua Shenzhen International Graduate School, Shenzhen, China · Sun Yat-sen University, Zhuhai, China · Tsinghua University, Beijing, China +1