Task-Guided Prompting for Unified Remote Sensing Image Restoration PubID: pubid: 0000–0000/00$00.00 © 2025 IEEE
Wenli Huang*, Yang Wu*, Xiaomeng Xin, Zhihong Liu, Jinjun Wang, Ye Deng · 2026
Read Task-Guided Prompting for Unified Remote Sensing Image Restoration PubID: pubid: 0000–0000/00$00.00 © 2025 IEEE by Wenli Huang*, Yang Wu*, Xiaomeng Xin, Zhihong Liu, Jinjun Wang, Ye Deng (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-02742.txt.
Introduction
Huang and five colleagues propose TGPNet, a unified framework for restoring remote sensing images degraded in many ways at once. Most methods handle one degradation type in uniform data, but real scenes mix denoising, cloud and shadow removal, deblurring, and radar despeckling across spectral bands. Task-guided prompting uses learnable task embeddings to steer the decoder per degradation, keeping one shared weight set across RGB, multispectral, SAR, and thermal inputs.
Catalog record
| CREATOR | Wenli Huang*, Yang Wu*, Xiaomeng Xin, Zhihong Liu, Jinjun Wang, Ye Deng |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74820 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.02742
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Task-Guided Prompting for Unified Remote Sensing Image Restoration PubID: pubid: 0000–0000/00$00.00 © 2025 IEEE Wenli Huang*, Yang Wu*, Xiaomeng Xin, Zhihong Liu, Jinjun Wang, Ye Deng Abstract 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 . I Introduction 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., E_{N},D_{N} for denoising) for each degradation type. (c) Our unified approach employs a single, prompt-guided network ( E_{\text{all}},D_{\text{all}} ) to handle diverse tasks, including denoising, deshadowing, declouding, despeckling, and deblurring. (d) The corresponding high-quality restored images. Remote sensing (RS) images are indispensable to modern Earth observation, underpinning critical applications in environmental monitoring [ 1 ] , resource management [ 2 ] , and sustainable development [ 3 ] . However, the quality of raw imagery is frequently compromised by a variety of degradations. These range from atmospheric interference, such as clouds and shadows, to sensor-specific artifacts like noise and Synthetic Aperture Radar (SAR) speckle, which severely impair the accuracy of derived geospatial information (see Fig. 1 (a)). Consequently, remote sensing image restoration (RSIR) has emerged as a crucial post-processing step aimed at reconstructing high-fidelity imagery from corrupted observations. Effective restoration not only improves visual quality but also significantly enhances the performance of downstream tasks, including object detection [ 4 ] and land-cover segmentation [ 5 ] . Recent advances in deep learning have led to a proliferation of specialized models for single-task RSIR, which achieve impressive performance on isolated degradations such as denoising [ 6 ] , declouding [ 7 , 8 ] , shadow removal [ 9 ] , deblurring [ 10 ] , and SAR despeckling [ 11 , 12 ] . While powerful, these task-specific approaches operate as “expert” models, each meticulously engineered for a single purpose (Fig. 1 (b)). Their practical applicability is severely constrained by the complex and dynamic nature of real-world remote sensing data, where multiple degradation types often vary spatiotemporally. This inherent variability makes it infeasible to select and deploy an appropriate specialized model for every scenario, creating a significant bottleneck for operational RSIR pipelines. The limitations of single-task models have motivated the development of unified restoration frameworks, primarily within the natural image domain [ 13 , 14 , 15 ] . However, these “all-in-one” models are not directly transferable to the remote sensing context due to the pronounced domain gap. Existing methods often treat restoration tasks in isolation or strictly within the optical domain. However, real-world remote sensing faces dual challenges: diverse degradation types (e.g., noise, clouds, shadows, blur, speckle) and heterogeneous imaging domains (e.g., optical RGB, SAR microwave, TIR thermal). These domains differ fundamentally in imaging physics (active vs. passive sensing) and degradation characteristics (additive vs. multiplicative noise), making unified restoration a non-trivial problem. Therefore, our work focuses on Unified Cross-Domain Multi-Degradation Restoration . We aim to design a single framework adaptable to both varied degradation objectives and distinct imaging domains. We validate TGPNet through a progressive experimental hierarchy: starting with two and three optical-domain tasks (intra-domain baseline), extending to include SAR despeckling (dual-domain generalization), and finally incorporating TIR deblurring (tri-domain comprehensive unification). This design systematically verifies the framework’s scalability across fundamental sensing modalities. To address this critical need, we propose TGPNet, a unified framework designed to handle heterogeneous remote sensing image restoration (RSIR) tasks within a single model. Crucially, rather than embedding rigid, hand-crafted physical equations—which inherently conflict when unifying incompatible modalities like SAR speckle and optical degradations—TGPNet relies on data-driven implicit priors . At the heart of TGPNet is our novel Task-Guided Prompting (TGP) strategy, which enables the network to perform diverse restoration tasks through a unified architecture (Fig. 1 (c)). TGP consists of two key components: a Learnable Task-Specific Embedding (LTSE) module that encodes the restoration task into a compact vector, and a Hierarchical Feature Modulation (HFM) module. Functioning as a dynamic physical context switch, the HFM leverages this embedding to generate multi-scale affine parameters that modulate features across the decoder stages. This mechanism allows the network to adaptively reconfigure feature statistics to match specific signal distributions, aligning its behavior with distinct degradation characteristics without altering its core weights. This effectively enables a single model to function as multiple experts. Furthermore, to facilitate rigorous evaluation and advance future research, we construct a comprehensive multi-degradation benchmark for RSIR. This benchmark encompasses RGB, multispectral (MS), SAR, and thermal infrared (TIR) data across five key restoration tasks: denoising, cloud removal, shadow removal, deblurring, and SAR despeckling. Extensive evaluations demonstrate that TGPNet achieves state-of-the-art performance; it not only outperforms existing unified frameworks but also exhibits robust generalization to unseen composite degradations (e.g., simultaneous denoising and cloud removal) and surpasses specialized single-task models in their respective domains. This work advances remote sensing image restoration (RSIR) through three key innovations: • We propose TGPNet, the first framework to unify the restoration of heterogeneous degradations across fundamentally distinct sensor modalities (optical, SAR, and TIR) within a single, prompt-guided network. • We devise the Task-Guided Prompting (TGP) strategy, where a learnable task embedding guides a hierarchical modulation module to achieve dynamic, task-adaptive feature processing in the decoder. • We construct and will publicly release a multi-modal benchmark for multi-degradation RSIR, on which TGPNet achieves state-of-the-art performance and demonstrates superior cross-task generalization. II Related Works This section reviews the literature relevant to Remote Sensing Image Restoration (RSIR). We first survey methods developed specifically for RS data, noting a prevailing trend toward specialized models that lack a unified framework. We then analyze unified restoration models from the natural image domain to highlight their limitations for RSIR scenarios. II-A Remote Sensing Image Restoration Remote sensing imagery is inherently susceptible to various degradations that hinder downstream analysis tasks like classification and change detection. These are commonly categorized into four types [ 16 ] : atmospheric interference (e.g., clouds), illumination effects (e.g., shadows),