Task-Guided Prompting for Unified Remote Sensing Image Restoration
Authors: Wenli Huang, Yang Wu, Xiaomeng Xin, Zhihong Liu, Jinjun Wang, Ye Deng
Organizations: School of Electronic and Information Engineering, Ningbo University of Technology, Ningbo, Zhejiang 315211, China · Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, China · University of Exeter, Exeter EX4 4PY, UK · Engineering Research Center of Intelligent Finance, Ministry of Education, and the School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics, China
Remote sensing image restoration (RSIR) is essential for recovering high-fidelity imagery from degraded observations, enabling accurate downstream analysis. However, most existing methods focus on single degradation types within homogeneous data, restricting their practicality in real-world scenarios where multiple degradations often across diverse spectral bands or sensor modalities, creating a significant operational bottleneck. To address this fundamental gap, we propose TGPNet, a unified framework capable of handling denoising, cloud removal, shadow removal, deblurring, and SAR despeckling within a single, unified architecture. The core of our framework is a novel Task-Guided Prompting (TGP) strategy. TGP leverages learnable, task-specific embeddings to generate degradation-aware cues, which then hierarchically modulate features throughout the decoder. This task-adaptive mechanism allows the network to precisely tailor its restoration process for distinct degradation patterns while maintaining a single set of shared weights. To validate our framework, we construct a unified RSIR benchmark covering RGB, multispectral, SAR, and thermal infrared modalities for five aforementioned restoration tasks. Experimental results demonstrate that TGPNet achieves state-of-the-art performance on both unified multi-task scenarios and unseen composite degradations, surpassing even specialized models in individual domains such as cloud removal. By successfully unifying heterogeneous degradation removal within a single adaptive framework, this work presents a significant advancement for multi-task RSIR, offering a practical and scalable solution for operational pipelines. The code and benchmark will be released at https://github.com/huangwenwenlili/TGPNet.
Figures & tables
Fig. 1: Conceptual comparison of different RSIR paradigms. (a) Examples of real-world degradations across diverse modalities, including optical (RGB/Multispectral), Synthetic Aperture Radar (SAR), and Thermal Infrared (TIR). (b) Conventional single-task restoration relies on specialized networks (e.g., EN,DN for denoising) for each degradation type. (c) Our unified approach employs a single, prompt-guided network ( Eall,Dall ) to handle diverse tasks, including denoising, deshadowing, declouding, despeckling, and deblurring. (d) The corresponding high-quality restored images.
Fig. 2: Architecture of the proposed Task-Guided Prompting Network (TGPNet) for unified remote sensing image restoration. (a) The overall architecture consists of an encoder-decoder backbone augmented by our Task-Guided Prompting (TGP) module, which is composed of Learnable Task-Specific Embedding (LTSE) and Hierarchical Feature Modulation (HFM). (b) The LTSE module processes a task-specific prompt ( FTP ) to produce a degradation-aware embedding ( FE ). (c) The HFM module uses this embedding to generate stage-wise affine parameters (scaling factors γ and shifting offsets β ), which modulate the hierarchical features ( FDi ) at each stage of the decoder.
Task
Dataset
Train
Test
Resolution
Denoising
UCMLUD [ 48 ]
1680
420
256×256
Declouding
RICE1 [ 49 ]
400
100
512×512
RICE2 [ 49 ]
558
148
512×512
SEN12MS-CR [ 50 ]
114,319
7,899
256×256
Deshadowing
SRD [ 51 ]
2680
408
640×480
UAV-TSS
2,200
400
1920×1080
TABLE I: Overview of the datasets comprising the benchmark for unified remote sensing image restoration (URSIR).
Stage
Operator
Param & Output
Input
-
H × W ×Cin
Stem
Conv (3×Cin×C)
H × W ×C
E1
[TB] ×1
Nh=2 , H × W ×C
Downsample
Conv, PixelUnshuffle(2)
H/2 × W/2 ×2C
E2
[TB] ×2
Nh=4 , H/2 × W/2 ×2C
Downsample
Conv, PixelUnshuffle(2)
H/4 × W/4 ×4C
TABLE II: Detail of our TGPNet architecture.
Method
2 Types
3 Types
Declouding
Deshadowing
Average
Declouding
Deshadowing
Denoising
Average
on RICE1
on RICE2
on SRD
on UAV-TSS
on RICE1
on RICE2
on SRD
Avg. on UCMLUD
PromptIR [ 14 ]
30.32/0.9445
33.36/0.8960
24.77/0.8559
40.39/0.9911
32.21/0.9219
31.24/0.9502
32.57/0.8973
24.90/0.8570
31.20/0.8630
29.98/0.8919
AdaIR [ 15 ]
28.98/0.9417
33.65/0.8966
25.02/0.8563
39.97/0.9909
31.91/0.9214
24.28/0.9047
29.37/0.8817
23.62/0.8507
31.00/0.8596
27.07/0.8742
MOCE-IR [ 32 ]
30.98/0.9441
31.69/0.8844
24.17/0.8509
39.85/0.9869
31.67/0.9166
32.75/0.9509
32.77/0.8941
24.63/0.8562
31.04/0,8601
30.30/0.8903
PromptHSI [ 28 ]
32.22/0.9513
33.56/0.8888
24.89/0.8563
38.67/0.9831
32.34/0.9199
30.65/0.9507
33.81/0.8915
24.52/0.8512
23.57/0.6406
28.14/0.8335
TABLE III: Comparisons for unified remote sensing image restoration (RSIR) across two and three degradation types (PSNR/SSIM; higher = better; best in bold).
Method
Declouding
Deshadowing
Denoising on UCMLUD
SAR Despeckling
Average
on RICE1
on RICE2
on SRD
σ=15
σ=25
σ=50
On NRD
PromptIR [ 14 ]
32.38/0.9539
34.67/0.9067
26.33/0.8653
34.07/0.9224
31.78/0.8827
28.82/0.8102
24.91/0.8311
30.42/0.8818
AdaIR [ 15 ]
33.10/0.9566
35.12/0.9094
26.99/0.8679
34.14/0.9238
31.85/0.8844
28.89/0.8124
25.05/0.8262
30.73/0.8830
MOCE-IR [ 32 ]
31.85/0.9561
33.88/0.9105
26.14/0.8685
34.13/0.9240
31.83/0.8844
28.86/0.8113
25.15 /0.8310
30.26/0.8838
PromptHSI [ 28 ]
32.15/0.9522
34.22/0.9010
25.96/0.8610
24.94/0.8034
24.59/0.7058
21.57/0.4719
24.79/0.8213
26.89/0.7881
Art [ 56 ]
29.70/0.9183
30.73/0.8599
23.65/0.8188
29.05/0.8505
28.01/0.7998
25.93/0.7028
21.31/0.6537
26.91/0.8005
TABLE IV: Comparisons for unified remote sensing image restoration (RSIR) across four degradation types (PSNR/SSIM; higher = better; best in bold).
Fig. 3: Visual comparison of restored images for four degradation types on our URSIR benchmark. Key local details (in green boxes) are highlighted in red boxes, and zooming in is recommended for detailed inspection.
Method
Declouding
Deshadowing
Denoising
SAR Despeckling
Deblurring
Average
On RICE1
On RICE2
On SEN12MS-CR
On SRD
Avg. on UCMLUD
On NRD
Avg. on HIT-UAV
PromptIR [ 14 ]
32.30/0.9510
32.66/0.9025
28.73/0.8712
25.17/0.8592
31.06/0.8563
24.97/0.8093
31.61/0.8701
29.50/0.8742
AdaIR [ 15 ]
32.01/0.9517
33.72/0.9045
29.47/0.8789
25.53/0.8625
31.31/0.8656
25.10/0.8241
31.82/0.8722
29.85/0.8799
MOCE-IR [ 32 ]
29.40/0.9240
31.91/0.8956
29.07/0.8763
25.01/0.8596
30.85/0.8546
24.95/0.8121
31.26/0.8633
28.92/0.8694
Art [ 56 ]
29.85/0.8414
30.71/0.8411
27.82/0.8599
23.92/0.8103
27.36/0.7732
20.96/0.6542
28.93/0.8244
27.08/0.8006
CR-former [ 7 ]
35.46/0.9588
35.85/0.9109
29.35/0.8926
27.53/0.8719
31.43/0.8701
25.31 /0.8273
32.26/0.8771
31.03/0.8870
TABLE V: Comparisons for unified remote sensing image restoration (RSIR) across five degradation types (PSNR/SSIM; higher = better; best in bold).
Fig. 4: Visual comparison of restored images for multispectral declouding on SEN12MS-CR and thermal deblurring on HIT-UAV. Key local details (in green boxes) are highlighted in red boxes, and zooming in is recommended for detailed inspection.
Fig. 5: Visual evaluation of TGPNet on unseen real-world imagery from the WHU-Shadow dataset [ 57 ] , demonstrating generalization to authentic remote sensing shadows.
Fig. 6: Visual comparison of restoration results for out-of-distribution composite degradations: direct vs. sequential processing. Direct approach (task embedding averaging) performs suboptimally. Sequential processing achieves effective restoration by decomposing composite degradations into ordered stepwise subtasks. Row 1: Denoising + Declouding; Row 2: Denoising + Deshadowing. Zoom in for details.
Method
Denoising+Declouding on RICE1
Denoising+Declouding on RICE2
Denoising+Deshadowing on SRD
Average
σ=15
σ=25
σ=50
σ=15
σ=25
σ=50
σ=15
σ=25
σ=50
PromptIR [ 14 ]
18.61/0.7062
18.47/0.6764
18.36/0.6312
22.32/0.7747
22.26/0.7709
22.22/0.7669
17.83/0.7622
17.70/0.7278
17.44/0.6583
19.47/0.7194
AdaIR [ 15 ]
18.54/0.7070
18.47/0.6754
18.34/0.6308
22.27/0.7730
22.27/0.7709
22.21/0.7676
17.85/0.7608
17.71/0.7282
17.46/0.6602
19.46/0.7193
MOCE-IR [ 32 ]
18.56/0.7129
18.47/0.6794
18.31/0.6307
22.31/0.7722
22.23/0.7693
22.16/0.7669
17.96/0.7622
17.72/0.7286
17.45/0.6597
19.46/0.7202
PromptHSI [ 28 ]
22.93/0.6268
20.02/0.4469
17.18/0.2411
30.06/0.8202
26.18/0.6869
20.61/0.3238
23.09/0.7046
20.66/0.5603
17.15/0.3320
21.99/0.5270
Art [ 56 ]
19.86/0.7022
18.43/0.6099
17.52/0.5374
25.58/0.7554
21.79/0.6675
20.61/0.6047
16.95/0.4374
16.61/0.4854
16.57/0.4202
19.32/0.5800
TABLE VI: Quantitative comparison on out-of-distribution two-composite-degradation tasks. Evaluated methods include RSIR baselines, our direct approach (task embedding averaging), and our sequential approach (ordered task-specific subtasks).
Method
Denoising+Deblurring+Declouding on RICE1
Denoising+Deblurring+Declouding on RICE2
Denoising+Deblurring+Deshadowing on SRD
Average
σ=15
σ=25
σ=50
σ=15
σ=25
σ=50
σ=15
σ=25
σ=50
PromptIR [ 14 ]
18.97/0.6514
18.61/0.6324
18.29/0.6011
22.97/0.7831
22.50/0.7708
22.20/0.7560
17.47/0.6483
17.32/0.6120
17.09/0.5467
19.49/0.6690
AdaIR [ 15 ]
19.02/0.6622
18.64/0.6415
18.39/0.6085
22.41/0.7773
22.30/0.7719
22.27/0.7642
17.46/0.6594
17.32/0.6205
17.08/0.5561
19.43/0.6735
MOCE-IR [ 32 ]
19.51/0.6572
18.96/0.6342
18.45/0.5862
22.61/0.7744
22.24/0.7607
22.03/0.7352
17.87/0.6543
17.44/0.6164
17.14/0.5434
19.58/0.6624
Art [ 56 ]
20.38/0.6778
19.90/0.5685
18.10/0.4843
28.51/0.7493
23.45/0.6350
20.98/0.5494
18.34/0.5145
16.99/0.4680
16.45/0.3962
20.34/0.5603
CR-former [ 7 ]
18.55/0.6720
18.44/0.6480
18.32/0.6154
22.46/0.7756
22.38/0.7736
22.28/0.7694
17.51/0.6649
17.31/0.6259
17.07/0.5621
19.37/0.6785
TABLE VII: Quantitative comparison on out-of-distribution three-composite-degradation tasks. Evaluated methods include RSIR baselines, our direct approach (task embedding averaging), and our sequential approach (ordered task-specific subtasks).
Fig. 7: Visual comparison of restored images on composite degradation tasks under Gaussian noise ( σ=50 ). Zooming in is recommended for detailed inspection.
Fig. 8: Visual comparison of restored images for single-degradation declouding on RICE2. Key local details (in green boxes) are highlighted in red boxes, and zooming in is recommended for detailed inspection.
Method
PSNR ↑
SSIM ↑
MAE ↓
SAM ↓
SPA-GAN [ 58 ]
27.51
0.8177
0.0403
3.36
CVAE [ 59 ]
33.62
0.9079
0.0216
1.69
CMNet [ 8 ]
34.92
0.9121
0.0188
1.40
ACA-CRNet [ 55 ]
35.65
0.9126
0.0164
1.29
Restormer [ 45 ]
35.65
0.9124
0.0173
1.33
CR-former [ 7 ]
35.61
0.9123
0.0168
1.33
TABLE VIII: Comparisons for single declouding degradation on the RICE2 dataset. Higher PSNR and SSIM indicate better performance, while lower MAE and SAM are preferable. Best in bold.
Method
Params(M)
FLOPs(G)
GPU memory(MB)
Times(ms)
PromptIR [ 14 ]
35.59
177.15
1753
166.21
AdaIR [ 15 ]
28.79
165.89
1735
169.31
MOCE-IR [ 32 ]
11.48
44.12
1007
88.65
PromptHSI [ 28 ]
25.91
408.58
5227
218.22
Art [ 56 ]
37.79
4.90
973
7.16
CR-former [ 7 ]
20.86
71.22
1561
72.88
TABLE IX: Efficiency comparison of TGPNet with unified image restoration models across four key metrics (lower is better).
Method
Params(M)
FLOPs(G)
GPU memory(MB)
Times(ms)
SPA-GAN [ 58 ]
0.21
15.2
3749
26.00
CVAE [ 59 ]
15.42
37.1
1097
15.10
CMNet [ 8 ]
16.51
236.0
1607
197.00
ACA-CRNet [ 55 ]
20.39
1422.0
6417
223.70
Restormer [ 45 ]
26.13
155.0
1629
112.30
Ours
21.27
71.4
1643
60.84
TABLE X: Efficiency comparison of TGPNet with declouding models across four key metrics (lower is better).
Method
Declouding
Deshadowing
Denoising on UCMLUD
SAR Despeckling
Average
on RICE1
on RICE2
on SRD
sigma=15
sigma=25
sigma=50
On NRD
baseline
35.29/0.9606
36.07/0.9136
28.19/0.8744
34.25/0.9253
31.97 /0.8868
29.04/0.8166
24.88/0.8284
31.38/0.8865
w/ TGP (Relu)
35.68/0.9601
36.15/0.9143
28.44/0.8758
34.27/0.9255
31.97/0.8869
29.04/0.8162
24.96 /0.8283
31.50/0.8867
w/ TGP (Ours)
35.73/0.9608
36.35/0.9145
28.40/0.8747
34.27/0.9255
31.97 /0.8868
29.05/0.8169
24.88/ 0.8329
31.52/0.8874
TABLE XI: Ablation study for the proposed Task-Guided Prompt (TGP) module on the URSIR benchmark (PSNR/SSIM; higher = better; best in bold).
Fig. 9: Visualization of ablation study results comparing the baseline model and our TGP-enhanced unified architecture. Zooming in facilitates better visualization and comparison.
Method
Internal Metrics
External Metrics
Silhouette
CH
Dunn
ARI
AMI
FMI
baseline
0.1605
163.11
0.1450
0.4837
0.6348
0.6407
w/ TGP (Ours)
0.4388
572.76
0.2754
0.9610
0.9279
0.9736
TABLE XII: Clustering performance of decoder stage 2 features base and with TGP module. A higher metric is better.
Fig. 10: t-SNE visualization of decoder stage 2 features for the baseline vs. TGPNet across four degradation tasks.
Tasks
Method
Internal Metrics
External Metrics
Silhouette
CH
Dunn
ARI
AMI
FMI
4-task
before TGP
0.1621
209.77
0.1649
0.4669
0.6359
0.626
after TGP
0.1623
237.75
0.2046
0.5206
0.6714
0.6634
5-task
before TGP
0.3448
649.51
0.1908
0.707
0.784
0.7783
after TGP
0.4314
522.41
0.312
0.9124
0.8842
0.9338
TABLE XIII: Clustering performance of decoder stage 2 features before and after TGP module in both 4-task and 5-task settings. A higher metric is better.
Fig. 11: t-SNE visualization of decoder stage 2 features before and after application of the TGP module, shown for both 4-task and 5-task settings.
Method
Declouding
Deshadowing
Denoising
SAR Despeckling
Deblurring
Average
On RICE1
On RICE2
On SEN12MS-CR
On SRD
Avg. on UCMLUD
On NRD
Avg. on HIT-UAV
Random learnable prompts (Ours)
35.59/0.9591
35.93/0.9130
29.61 /0.8963
27.97/ 0.8739
31.55/0.8723
24.95 /0.8294
32.35/0.8789
31.14/0.8890
Language-based prompts (CLIP)
35.76/0.9601
36.22/0.9131
29.50/ 0.8973
28.31 /0.8736
31.60/0.8733
24.93/ 0.8312
32.41/0.8797
31.25/0.8898
TABLE XIV: Comparison of prompt strategies across five RS degradation types (PSNR/SSIM; higher = better; best in bold).
Train Iterations
Baseline
TGPNet
Avg. on 4 Tasks
SAR Despeckling
Avg. on 4 Tasks
SAR Despeckling
200,000
30.72/0.8844
25.08 /0.8275
31.32/0.8856
24.90/0.8278
300,000
31.27/0.8855
25.03/0.8309
31.35/0.8859
24.96 /0.8291
400,000
31.28/0.8859
24.94/0.8267
31.51/0.8874
24.85/0.8324
500,000
31.38/0.8865
24.88/ 0.8284
31.52/0.8874
24.88/ 0.8329
TABLE XV: Performance comparison between Baseline and TGPNet at different training iterations (PSNR/SSIM; higher = better; best in bold).
Task-driven image restoration aims to improve both image quality and downstream task performance. However, existing methods predominantly focus on single degradation type and struggle to handle the diverse degradations encountered in real-world scenarios. Different degradations impose distinct restoration demands, and insufficient restoration may leave residual degradations and artifacts that impair object boundaries and semantic cues, thereby compromising downstream task performance. To address these challenges, we propose TaskIR, a two-stage task-driven unified image restoration framework that integrates degradation-adaptive restoration with task feedback refinement. In Stage I, a Degradation Representation Module (DRM) extracts degradation representations, enabling a Degradation-Guided Transformer Block (DGTB) to dynamically modulate feature transformations for adaptive restoration. In Stage II, a Task-to-Restoration Feedback Generation module (TRFG) transforms heterogeneous task features into restoration feedback by modeling task-representation discrepancies associated with the current restoration. Subsequently, a Selective Task Feedback Refinement module (STFR) assesses feedback relevance and selectively refines intermediate restoration features to mitigate interference with well-restored content. Extensive experiments demonstrate that TaskIR achieves competitive restoration quality and downstream task performance across diverse degradations and tasks.
Yanjie Tu, Qingsen Yan, Axi Niu +4
Northwestern Polytechnical University · Shenzhen Research Institute of Northwestern Polytechnical University · Xi’an University of Architecture and Technology
Remote sensing imagery suffers from clouds, haze, noise, resolution limits, and sensor heterogeneity. Existing restoration and fusion approaches train separate models per degradation type. In this work, we present Language-conditioned Large-scale Remote Sensing restoration model (LLaRS), the first unified foundation model for multi-modal and multi-task remote sensing low-level vision. LLaRS employs Sinkhorn-Knopp optimal transport to align heterogeneous bands into semantically matched slots, routes features through three complementary mixture-of-experts layers (convolutional experts for spatial patterns, channel-mixing experts for spectral fidelity, and attention experts with low-rank adapters for global context), and stabilizes joint training via step-level dynamic weight adjustment. To train LLaRS, we construct LLaRS1M, a million-scale multi-task dataset spanning eleven restoration and enhancement tasks, integrating real paired observations and controlled synthetic degradations with diverse natural language prompts. Experiments show LLaRS consistently outperforms seven competitive models, and parameter-efficient finetuning experiments demonstrate strong transfer capability and adaptation efficiency on unseen data. Repo: https://github.com/yc-cui/LLaRS
Yongchuan Cui, Peng Liu
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China · School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing, China
Unified image restoration using a single model often faces task interference due to diverse degradations. To address this, we propose DACG-IR (Degradation-Aware Adaptive Context Gating), which enables explicit perception of degradation characteristics to dynamically modulate feature representations. Our method constructs degradation-aware contextual representations from the input to modulate attention distribution, frequency-domain features, and feature aggregation. Specifically, a lightweight multi-scale degradation-aware module extracts coarse degradation information and generates layer-wise prompts. These prompts guide attention temperature and output gating in encoder and decoder blocks for adaptive feature extraction. Additionally, a spatial-channel dual-gated adaptive fusion mechanism refines encoder features, suppressing noise propagation from shallow to deep layers. This design effectively suppresses degradation-induced noise while preserving informative structures. Experiments show DACG-IR outperforms state-of-the-art methods in single-task, all-in-one, adverse weather removal, and composite degradation settings. Code: https://github.com/HlHomes/DACG-IR-code
Lei He, Jielei Chu, Fengmao Lv +4
the School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China · the Engineering Research Center of Sustainable Urban Intelligent Transportation, Ministry of Education, Chengdu, Sichuan 611756, China · School of Computing and Artificial Intelligence, Jiangxi University of Finance and Economics, Nanchang, 330013, China +3