Organizations: School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China · Josef Kittler Research Institute on Artificial Intelligence, Jiangnan University, Wuxi 214122, China · School of Electronic Engineering, Xidian University, Xi’an 710071, China · College of Artificial Intelligence, China University of Petroleum (Beijing), Beijing 102249, China · Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China · Department of Computer Science, University of California, Santa Barbara, CA 93106, USA
Remote sensing images (RSIs) are frequently degraded by haze, fog, and thin clouds, which obscure surface reflectance and hinder downstream applications. This study presents the first systematic and unified survey of RSIs dehazing, integrating methodological evolution, benchmark assessment, and physical consistency analysis. We categorize existing approaches into a three-stage progression: from handcrafted physical priors, to data-driven deep restoration, and finally to hybrid physical-intelligent generation, and summarize more than 30 representative methods across CNNs, GANs, Transformers, and diffusion models. To provide a reliable empirical reference, we conduct large-scale quantitative experiments on five public datasets using 12 metrics, including PSNR, SSIM, CIEDE, LPIPS, FID, SAM, ERGAS, UIQI, QNR, NIQE, and HIST. Cross-domain comparison reveals that recent Transformer- and diffusion-based models improve SSIM by 12%~18% and reduce perceptual errors by 20%~35% on average, while hybrid physics-guided designs achieve higher radiometric stability. A dedicated physical radiometric consistency experiment further demonstrates that models with explicit transmission or airlight constraints reduce color bias by up to 27%. Based on these findings, we summarize open challenges: dynamic atmospheric modeling, multimodal fusion, lightweight deployment, data scarcity, and joint degradations, and outline promising research directions for future development of trustworthy, controllable, and efficient (TCE) dehazing systems. All reviewed resources, including source code, benchmark datasets, evaluation metrics, and reproduction configurations are publicly available at https://github.com/VisionVerse/RemoteSensing-Restoration-Survey.
Figures & tables
Term
Physical Cause
Relation to This Review
Haze
Aerosol scattering and absorption veil surface radiance and reduce contrast.
Core focus . This survey fully covers haze as the primary visibility degradation.
Thin Cloud
Weak scattering by small droplets forms a semi-transparent layer.
Included . Treated as haze-like when optical thickness is low.
Thick Cloud, Cloud Shadows
Strong occlusion from dense clouds or shadows causes irreversible radiance loss.
Excluded . Belongs to cloud removal rather than dehazing.
Atmospheric Correction
Gas absorption and Rayleigh or Mie scattering alter top-of-atmosphere radiance.
Related but not covered . A radiometric calibration process rather than visibility restoration.
Table 1: The difference analysis of haze, thin clouds, discussed in this review, and thick clouds, atmospheric correction.
Figure 1: Illustrative scenarios for RSIs dehazing. Each column (a)–(d) shows a different land-cover type from riverine–agricultural to industrial–urban areas, with haze density increasing from light (a) to heavy (d).
Figure 3: Number of papers published in RSIs dehazing up to January 2026.
Publication
IF2025
h5
Num.
IEEE Trans. on Pattern Analysis and Machine Intelligence
18.6
196
6
Information Fusion
15.5
134
6
IEEE Trans. on Image Processing
13.7
150
16
ISPRS J. of Photogrammetry and Remote Sensing
12.2
104
12
IEEE Trans. on Circuits and Systems for Video Technology
11.1
94
11
IEEE Trans. on Geoscience and Remote Sensing
8.6
141
99
Table 2: Impact Factor (IF2025), h5-index, and Paper Count † of Selected Journals and Conferences.
Figure 4: The distribution ratio of high-quality journals (impact factor greater than 7) and conference articles in the past three years.
Figure 5: Diagram of Image Enhancement and Priors Modeling for RSIs Dehazing.
Figure 6: Diagram of Deep Convolution Methods for RSIs Dehazing.
Figure 7: Diagram of Adversarial Generation Methods for RSIs Dehazing.
Figure 8: Diagram of Vision Transformers Methods for RSIs Dehazing.
Figure 9: Diagram of Diffusion Generation Methods for RSIs Dehazing.
Method
Publication
Code Link
CR-GAN-PM [ 94 ]
ISPRS P&RS 2020
https://github.com/Neooolee/CR-GAN-PM
RSDehazeNet [ 50 ]
TGRS 2020
https://github.com/tianqiong123/RSDehazeNet
DCIL [ 219 ]
TGRS 2022
https://github.com/Shan-rs/DCI-Net
GLF-CR [ 199 ]
ISPRS P&RS 2022
https://github.com/xufangchn/GLF-CR
DehazeFormer [ 164 ]
TIP 2023
https://github.com/IDKiro/DehazeFormer
RSDformer [ 162 ]
GRSL 2023
https://github.com/MingTian99/RSDformer
Table 3: Selected publicly available codes for RSIs dehazing methods.
Datasets
Number
Real/Syn.
Sensor Type
Nspectrum
Image Size
Application
Link
RICE1 19 [ 105 ]
500
Real
Google Earth
3
512×512
Hazy, Thin cloud
GitHub
RICE2 19 [ 105 ]
736
Real
Landsat 8 OLI/TIRS
3
512×512
Hazy, Thin cloud
GitHub
SateHaze1k 20 [ 66 ]
1200
Syn.
GF-2
3
512×512
Hazy
GitHub
LHID 22 [ 219 ]
31017
Syn.
Google Earth
3
512×512
Hazy
GitHub
DHID 22 [ 219 ]
14990
Syn.
Aerial Camera
3
512×512
Hazy
GitHub
RS-HAZE 23 [ 164 ]
51300
Syn.
Landsat 8 OLI/TIRS
11
512×512
Hazy
GitHub
Table 4: Summary of representative remote-sensing haze datasets, including data scale, real/synthetic type, sensor source, spectral counts, resolution, and applications.
Methods
Pub.
Overview
Advantages
Limitations
SMIDCP [ 93 ]
GRSL 2018
A sphere model to improve the DCP and combining it with HF for image enhancement.
Strong capability in handling inhomogeneous conditions, dense haze, and thin clouds.
Limited adaptability to extreme scenarios.
MDCP [ 156 ]
GRSL 2021
Integrate DCP into the multiscale transform and process images by selecting appropriate fusion rules and an optimized Laplacian model.
Effectively removes thin clouds and supports multi-temporal reference images.
Long processing time and highly dependent on the quality of reference images.
RSC-Net [ 96 ]
ISPRS P&RS 2019
Employs the encoding-decoding framework to enable end-to-end thin cloud removal.
Symmetrical concatenation effectively preserves details of cloud-free regions.
Relies on specific sensor band configurations and lacks validation of adaptability to complex atmospheric conditions.
RSDehazeNet [ 50 ]
TGRS 2020
An end-to-end network combining channel attention mechanism and global residual learning.
Fast processing speed, preserves image details, and high realism of synthetic data.
The inadequate adaptability to extreme conditions or complex terrain.
FCTF [ 99 ]
GRSL 2020
Adopting a coarse-to-fine two-stage architecture to refine dehazing results and optimize image details.
Improvement in dehazing accuracy under complex scenes.
Insufficient generalization capability to complex haze distributions and diverse ground surface characteristics.
CNNIM [ 246 ]
JSTARS 2021
Utilizing U-Net and Slope-Net to subtract the thin cloud thickness map and obtain a clear image.
Capable of acquiring thin cloud information at different altitudes and for each band.
Exhibits significant cumulative estimation errors for real-world images.
Table 5: Summary of representative remote sensing image dehazing methods, including publication venues, core designs, technical advantages, and existing limitations.
Methods
Category
SH-TN
SH-M
SH-TK
RICE
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
DHIM 15 [ 137 ]
Traditional
19.445
0.891
19.916
0.917
16.595
0.810
19.240
0.882
GRS-HTM 17 [ 110 ]
Traditional
15.489
0.762
15.071
0.784
10.473
0.462
18.278
0.825
SMIDCP 18 [ 93 ]
Traditional
13.639
0.833
15.990
0.863
14.956
0.757
16.573
0.712
IDeRs 19 [ 205 ]
Traditional
15.048
0.772
14.763
0.785
11.754
0.702
15.750
0.611
EVPM 22 [ 53 ]
Traditional
20.426
0.891
20.656
0.918
16.647
0.787
15.217
0.742
Table 6: Quantitative performance at PSNR (dB) and SSIM of RSIs dehazing algorithms evaluated on the SateHaze1k and RICE datasets. Red and blue values indicate the best and second-best performance for each metric, respectively.
Metrics
SpA-GAN [ 134 ]
DCIL [ 219 ]
DehazeFormer [ 164 ]
PSMB-Net [ 167 ]
Trinity-Net [ 20 ]
LFD-Net [ 75 ]
ASTA [ 7 ]
DehazeXL [ 13 ]
ERGAS ↓
13.386
14.473
18.127
13.075
22.765
23.813
12.188
26.444
32.128
30.591
29.493
28.706
34.624
34.873
11.727
37.214
24.088
17.574
16.013
16.808
22.293
25.929
15.343
36.951
UIQI ↑
0.902
0.904
0.889
0.911
0.650
0.858
0.913
0.759
0.555
0.621
0.624
0.624
0.425
0.598
0.944
0.428
0.739
0.843
0.837
0.841
0.650
0.773
0.845
0.544
Table 7: Quantitative performance at SAM, ERGAS, UIQI, QNR. NIQE and HIST on the SateHaze1k datasets. Red and blue values indicate the best and second-best performance for each metric, respectively.
Methods
Category
LHID
DHID
CIEDE ↓
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
CIEDE ↓
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
DCP 10 [ 55 ]
Traditional
9.161
0.481
21.490
21.120
0.818
9.867
0.163
79.920
18.920
0.824
IDeRs 19 [ 205 ]
Traditional
19.563
0.633
44.650
14.680
0.627
19.942
0.344
120.17
13.330
0.575
FCTF-Net 20 [ 99 ]
CNN
5.423
0.459
21.210
25.850
0.876
12.794
0.221
128.490
17.210
0.739
DCIL 22 [ 219 ]
CNN
5.180
0.447
14.930
28.120
0.898
4.877
0.131
64.730
26.950
0.896
PSMB-Net 23 [ 167 ]
CNN
5.864
0.454
16.130
28.890
0.900
4.965
0.113
53.830
27.070
0.887
Table 8: Quantitative performance at CIEDE ( ↓ ), LPIPS ( ↓ ), FID ( ↓ ) PSNR ( ↑ ) and SSIM ( ↑ ) of SOTA methods evaluated on the LHID and DHID datasets. Red and blue values indicate the best and second-best performance for each metric, respectively.
Figure 10: Quantitative comparison of PSNR (dB) across eight SOTA dehazing models on SateHaze1k (SH) and RRSIHID. The error bars indicate performance variance.
Methods
PSNR
FLOPs(G)
FPS
#Param(M)
SpA-GAN 20 [ 134 ]
19.664
67.931
29.84
0.210
DCIL 22 [ 219 ]
22.641
107.445
52.35
26.509
DehazeFormer 23 [ 164 ]
22.624
24.817
41.48
0.686
PSMB-Net 23 [ 167 ]
23.953
392.754
18.54
12.488
Trinity-Net 23 [ 20 ]
23.723
123.889
14.40
20.139
LFD-Net 23 [ 75 ]
17.975
23.352
269.07
0.090
Table 9: Model efficiency and performance comparison on SateHaze1k datasets.
Figure 11: Comparison of representative RSIs dehazing methods in terms of dehazing accuracy (PSNR) and computational complexity (FLOPs and Parameter).
Figure 12: Radiometric consistency analysis of RSIs before and after dehazing.
Methods
SH-TN
SH-M
SH-TK
SAM ↓
ERGAS ↓
SAM ↓
ERGAS ↓
SAM ↓
ERGAS ↓
SpA-GAN [ 134 ]
2.670
13.386
4.246
32.128
3.725
24.088
DCIL [ 219 ]
2.462
14.473
4.304
30.591
2.816
17.574
DehazeFormer [ 164 ]
2.646
18.127
3.923
29.493
3.032
16.013
PSMB-Net [ 167 ]
2.980
13.075
3.565
28.706
3.440
16.808
Trinity-Net [ 20 ]
2.835
22.765
3.457
34.624
3.494
22.293
Table 10: Physical consistency comparison in terms of SAM and ERGAS on SateHaze1k.
Figure 13: Visual comparison of dehazing results on the SateHaze1k datasets.
Figure 14: Visual comparison of dehazing results on the RICE datasets.
Figure 15: Visual comparison of dehazing results on the RRSHID datasets.
Figure 16: Uncertainty maps of DehazeXL on the RRSHID-TN, RRSHID-M and SateHaze1k-M datasets.
Metrics
SpA-GAN [ 134 ]
DCIL [ 219 ]
DehazeFormer [ 164 ]
PSMB-Net [ 167 ]
Trinity-Net [ 20 ]
LFD-Net [ 75 ]
ASTA [ 7 ]
DehazeXL [ 13 ]
PSNR ↑
20.640
20.846
22.640
23.190
23.003
21.002
22.420
22.980
20.525
23.254
23.360
23.120
22.767
20.838
20.843
23.023
20.947
24.640
24.550
25.180
24.055
20.433
24.160
24.171
SSIM ↑
0.493
0.527
0.594
0.643
0.631
0.517
0.591
0.592
0.506
0.651
0.660
0.657
0.665
0.544
0.651
0.640
0.573
0.707
0.705
0.726
0.725
0.527
0.699
0.681
Table 11: Quantitative evaluation on the authentic RRSHID datasets.
Figure 17: SSIM curves of different methods on SateHaze1k and RRSHID (thin to thick haze), reflecting structural fidelity under varying haze conditions.
Table 16: Task-driven evaluation of detection-friendly dehazing on the HazyDet datasets.
Figure 21: Visual comparison of detection-friendly dehazing on the HazyDet datasets.
Methods
PSNR ↑
SSIM ↑
Trained on Clear Images
Trained on Dehazing Images
mIoU ↑
Acc ↑
F 1 -score ↑
mIoU ↑
Acc ↑
F 1 -score ↑
U-Net [ 149 ]
–
–
14.87
70.03
24.64
30.95
80.14
43.98
+ SpA-GAN [ 134 ]
13.45
0.639
13.65 -1.22
68.43 -1.60
23.06 -1.58
13.62 -17.33
74.17 -5.97
21.54 -22.44
+ DCIL [ 219 ]
13.60
0.680
18.17 +3.30
73.05 +3.02
29.41 +4.77
31.42 +0.47
80.05 -0.09
45.08 +1.10
+ DehazeFormer [ 164 ]
21.02
0.872
30.39 +15.52
80.77 +10.74
43.42 +18.78
36.18 +5.23
82.83 +2.69
49.95 +5.97
+ PSMB-Net [ 167 ]
13.14
0.752
17.72 +2.85
73.60 +3.57
29.07 +4.43
26.76 +4.19
77.98 +2.16
40.33 +3.65
Table 17: Task-driven evaluation of segmentation-friendly dehazing on the HazyUDD datasets.
Figure 22: Visual comparison of segmentation-friendly dehazing on the HazyUDD datasets.
Figure 23: Future prospects for RSIs dehazing: Trustworthy, controllable, and efficient (TCE) remote sensing dehazing system.
Dimension
Concrete Research Actions
Clear and Measurable Goals
Trustworthy
• Reliable parameter estimation based on differential equations, including transmittance, atmospheric light, and scattering intensity; • Radiation-consistency regularization via knowledge distillation; • Uncertainty modeling for extreme haze regions using depth estimation.
• Accuracy and stability of estimated physical variables across datasets and haze levels; • Radiometric metrics: SAM ↓ , ERGAS ↓ , QNR ↑ , HIST ↑ ; • Reduced color bias and spectral distortion under real-world conditions.
Controllable
• Physics-aware latent disentanglement separating shared physical factors and modality-private appearance; • Cross-modal semantic anchoring using high-level prior knowledge from large multimodal models such as CLIP; • Hierarchical and region-adaptive multimodal alignment based on segmenting anything model; • Self-supervised learning and test-time adaptation guided by SAR geometric location cues.
• Performance decay rate under cross-dataset and cross-sensor transfer; • Semantic consistency when modalities are added, removed, or substituted; • Cross-domain invariance of estimated physical variables under similar atmospheric conditions.
Efficient
• Lightweight Mamba architectures reducing redundant feature learning; • Parameter-efficient multimodal fusion via reparameterization; • Efficient diffusion-generative models with reduced inference steps; • Hardware-aware operator optimization for edge deployment.
• Model size, FLOPs, and inference latency; • Scalability to large-scale or wide-swath RS imagery; • Performance–efficiency trade-off (accuracy/physical consistency per computation cost).
Table 18: TCE framework with concrete research actions and measurable goals.
Real-world remote sensing image dehazing (RSID) remains challenging because atmospheric scattering, spatially non-uniform haze and colour distortion jointly degrade structural and spectral information. Most deep learning methods rely on RGB inputs and spatial-domain feature extraction, which limits their ability to separate global background haze from local surface details. Here, we propose DPSF-Net, a dual-prior spatial-frequency network built on MCAF-Net for real-world RSID. The network uses hazy RGB images and dark channel prior (DCP) maps as joint inputs, allowing physical degradation cues to guide end-to-end feature learning. A spatial-frequency residual interaction block introduces a FourierUnit branch into multi-directional spatial interaction to model large-scale haze components. A prior-guided feature attention module adaptively fuses prior and attention features to reduce colour shift and structural distortion. A selective kernel complementary fusion module screens multi-scale skip features through bidirectional residual complementary gating and selective kernel fusion. Extensive experiments demonstrate that DPSF-Net achieves state-of-the-art performance on the real-world RRSHID remote sensing image dehazing benchmark and remains competitive across multiple synthetic datasets. Moreover, the proposed method strikes a favourable balance among restoration quality, parameter count and computational complexity, supporting the effectiveness of dual-prior spatial-frequency modelling.
Mei Lu, Shangliang Shao, Shanliang Yao
School of Software Engineering, Jinling Institute of Technology, Nanjing 211100, China · School of Information Engineering, Yancheng Institute of Technology, Yancheng 224051, China
Single-image dehazing aims to recover a clear scene from a hazy image and is generally formulated as an image-to-image translation task; however, it faces two limitations. Its performance depends heavily on the haze-formation priors embedded in the model. Prevailing methods adopt the Atmospheric Scattering Model (ASM), whose assumptions of single scattering and homogeneous media are often violated, leading to residual haze and color drift. Moreover, large-scale real hazy/clear pairs are impractical to collect, and existing synthesis approaches fail to reproduce the full complexity of natural haze. To address these issues, we present RTE-FM-Dehazer, a novel dehazing approach, together with a scalable data pipeline. Unlike the ASM, the Radiative Transfer Equation (RTE) jointly accounts for both scattering and absorption, naturally accommodating the non-homogeneous, multiple-scattering media that characterize real hazy scenes. Motivated by the structural similarity between the RTE diffusion-absorption term and the ODE in flow matching, we introduce a diffusion-absorption regularizer derived from a reduced RTE, to steer the flow matching trajectory at each step. Next, leveraging modern vision-language models, we build an automated pipeline and release P-HAZE, a dataset of 50000 realistic hazy/clear pairs. Extensive evaluations demonstrate that RTE-FM-Dehazer, trained solely on P-HAZE, effectively eliminates artifacts like residual haze and color drift, exhibits strong cross-domain generalization, and achieves leading results on five real-world dehazing benchmarks.
Chenfeng Wei, Chun Wang, Boyang Zhao +3
Xi’an Jiaotong-Liverpool University, China · University of Liverpool, UK · Mashang Consumer Finance Co. Ltd, China +3
Existing diffusion-based methods have recently made significant progress in image dehazing. However, they typically neglect the physics of haze formation and reconstruct clean images from pure Gaussian noise, thereby limiting their restoration potential. To address this issue, we propose Haze-Noise Diffusion (HNDiff), a novel diffusion framework that embeds the atmospheric scattering model as an inductive bias. By grounding diffusion in physical principles, HNDiff ensures that the restoration aligns more closely with underlying mechanisms of haze formation. In its forward process, we introduce joint haze-noise diffusion with a haze-aware noise scheduler, which progressively adds both haze and noise to an image. Essentially, the scheduler adapts noise levels according to haze density, meaning that regions with heavier haze receive stronger noise injection to encourage content generation, while clearer regions receive lighter noise to better preserve details, which directly links the forward degradation process with the physics of haze. In the reverse process, we then derive a physically consistent dehazing-denoising process that simultaneously removes haze and noise to restore a clean image in a manner aligned with the forward degradation process. To further enhance practicality, we propose Latent HNDiff, which compiles clean latent priors that can be seamlessly integrated into existing dehazing networks to boost performance. Extensive experiments show that our work significantly improves leading dehazing backbones and achieves state-of-the-art results on benchmark datasets. The project page is available at https://jin-ting-he.github.io/HNDiff .
Jin-Ting He, Fu-Jen Tsai, Yan-Tsung Peng +3
Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan · National Tsing Hua University, Taiwan · National Chengchi University, Taiwan +1