CARP: cloud-adaptive robust prompting of vision-language models for ship classification under cloud occlusion

Abstract

Fine-grained few-shot ship classification under cloud occlusion is vital for maritime safety but remains challenging due to corrupted features and limited data utility. While the advent of large pre-trained vision-language models (VLMs) provides promising solutions, the lack of specialized benchmarks hinders their effective application. To address this, we introduce SeaCloud-Ship, the first benchmark dedicated to this task. It comprises 7,654 high-resolution, high-quality annotated images across 30 classes, featuring quantified cloud coverage (12.5%–75%) for standardized evaluation. We innovatively propose CARP, a cloud-aware prompting framework built upon CoOp, to combat feature corruption, semantic misalignment, and utility decay. Our core contributions include: (1) GCE Loss dynamically adjusting classification weights to suppress cloud interference based on feature degradation severity; (2) Adaptive Optimization Prompt Design (AOPD) utilizing distortion-aware vectors for effective multi-modal feature alignment and semantic deviation repair; (3) Dynamic Weight Adjustment Mechanism (DWAM) real-time balancing of multi-source feature fusion by evaluating inter-modal information gain. Extensive experiments on SeaCloud-Ship demonstrate CARP’s superior robustness and state-of-the-art performance, establishing a strong baseline for cloud-occluded ship classification.

1 Introduction

Advancements in remote sensing have led to widespread Earth observation programs, elevating Remote Sensing Fine-Grained Ship Classification (RS-FGSC) as a critical task. However, RS-FGSC faces significant challenges, particularly severe data scarcity. Acquiring sufficient high-quality, labeled multi-category satellite ship imagery is extremely costly, often resulting in few-shot or zero-shot scenarios. Under these low-data conditions, traditional deep learning models struggle to generalize effectively.

Pretrained on large image-text pairs, Vision-Language Models (VLMs) like Contrastive Language-Image Pretrainin (CLIP) (Radford et al., 2021) offer a new paradigm for RS-FGSC, leveraging rich cross-modal semantics to lead in various remote sensing tasks (Bao et al., 2022; Chen et al., 2020; Jia et al., 2021; Lee et al., 2018; Li et al., 2021; Harold Li et al., 2019; Li X. et al., 2020; Lu et al., 2019; Singh et al., 2022; Su et al., 2019; Tan et al., 2019; Wang et al., 2025a; Wang et al., 2024a; Wang et al., 2025b; Wang et al., 2024b; Wang et al., 2021; Zhang et al., 2024). Standard adaptation fine-tunes these VLMs often updating only the classification head or prompts—using annotated RS-FGSC data. Yet for cloud-occluded few-shot RS-FGSC, this strategy encounters key obstacles: (1) Absence of dedicated benchmarks for cloud robustness evaluation; (2) Vulnerability to Cloud occlusion interference distinct to satellite imagery.

To bridge the first gap, we establish SeaCloud-Ship–the first benchmark for cloud-occluded fine-grained ship classification. Comprising 7,654 high-resolution images across 30 categories with quantified cloud coverage (12.5%-75%), it enables rigorous evaluation under controlled noise conditions. We further develop CloudGEN, a physics-based cloud synthesis method, to extend benchmark versatility.

Regarding the second challenge, three critical sub-problems emerge: Feature Corruption: Clouds obscure ships’ discriminative local features, compromising reliable feature extraction from limited samples. Semantic Misalignment: Cloud distortions warp the visual feature space, causing severe mismatch between degraded features and VLMs’ textual prompts. Data Utility Degradation: Cloud contamination reduces effective training samples, intensifying few-shot learning difficulties. Conventional fine-tuning and prompt learning methods show significantly reduced robustness and generalization under these combined pressures.

2 Related work

2.1 Prompt learning in vision-language models

Prompt learning has emerged as a dominant paradigm for efficiently adapting large-scale vision-language models (VLMs) to downstream vision tasks without the computational burden of full fine-tuning. Pioneered by CLIP (Radford et al., 2021), which utilized hand-crafted textual prompts like “a photo of a [class]” to align image-text representations, this approach evolved significantly with Context Optimization (CoOp) (Zhou et al., 2022b). CoOp replaced manual prompts with learnable continuous vectors optimized through gradient descent, substantially improving few-shot generalization by dynamically adapting prompts to target datasets. Subsequent research expanded this foundation along several dimensions: CoCoOp (Zhou et al., 2022a) introduced instance-conditional prompts to enhance generalization beyond base categories; VPT (Jia et al., 2022) unified visual and textual prompting within a shared optimization framework; while methods like MaPLe (Uzair Khattak et al., 2023) and ProGrad (Zhu et al., 2023) enforced hierarchical multimodal alignment through constraint-based learning. Domain-specific adaptations also emerged, such as ship-targeted prompt tuning (Lan et al., 2024) that customizes maritime semantics, and training-free variants like Tip-Adapter (Zhang et al., 2021) leveraging cached embeddings for zero-shot transfer. Collectively, these methods demonstrate robust performance across diverse vision tasks including open-vocabulary classification and object detection, provided they operate on high-resolution, unobstructed imagery.

However, these successes falter dramatically under cloud occlusion in remote sensing contexts. VLMs’ reliance on discriminative visual features renders them acutely vulnerable to cloud-induced local information loss, which corrupts feature extraction. Static prompt embeddings cannot dynamically compensate for such distortions, leading to progressive misalignment between visual and textual representations as cloud density increases. This drift is exacerbated in few-shot settings, where limited data impedes robust calibration against noise. While recent multimodal alignment techniques address generic domain shifts, they remain fundamentally unequipped to handle structured atmospheric interference (Wang et al., 2025b; Wang et al., 2024b). Consequently, existing frameworks fail to resolve the core challenges of feature corruption, semantic drift, and data scarcity in cloud-occluded fine-grained ship classification, highlighting a critical research void.

2.2 Ship classification

Ship classification in remote sensing imagery is challenged by fine-grained inter-class variations, complex maritime backgrounds, and pervasive cloud occlusion—a uniquely disruptive factor that catastrophically degrades visual features. Recent deep learning approaches have made significant strides through two primary strategies: domain adaptation for sensor invariance and multimodal learning for feature enrichment. Zheng et al. (2023) pioneered a dual-teacher framework (SCSD) that decomposes optical/SAR supervision into interactive cross-domain and semi-supervised tasks, substantially improving pseudo-label reliability on unlabeled SAR imagery. This direction was extended by the Multi-Level Alignment Network (Xu et al., 2022), which integrates pixel-, instance-, and feature-level alignment to mitigate domain shifts in detection tasks. For fine-grained discrimination, Huang et al. (2022) combined CNN-Swin hybrid architectures with multi-branch feature extraction, setting new benchmarks on FGSC-23 and military ship datasets. Concurrently, Lu et al. (2025) developed the Multi-Scale Context Aggregation Network (MSCAN), leveraging hierarchical convolution-attention fusion to suppress coastal clutter in SAR imagery and significantly enhance small-vessel localization. In multimodal learning, Li W. et al. (2020) achieved cross-sensor alignment via Cross-Modal Contrastive Learning (CMCL), constructing a shared semantic space that maximizes inter-modal consistency while preserving sensor-specific diversity.

Despite these innovations, cloud occlusion remains a critical unsolved vulnerability. Existing methods predominantly optimize for domain shifts (Xu et al., 2022; Zheng et al., 2023) or background interference (Lu et al., 2025), while cloud corruption—which intrinsically obliterates local discriminative features—demands fundamentally different mitigation mechanisms. Although unsupervised change detection (Zheng et al., 2021) offers partial relief by exploiting temporal sequences, it fails under persistent cloud cover common in maritime monitoring. Crucially, no current approach addresses the tripartite challenge of cloud-induced feature loss, semantic distortion, and data scarcity in few-shot settings. Our work bridges this gap by integrating physics-aware cloud modeling with prompt-based vision-language adaptation, establishing the first unified framework for occlusion-robust fine-grained classification.

3 Methods

3.1 Preliminary

CoOp (Zhou et al., 2022b) pioneers prompt learning for vision-language models by optimizing continuous context vectors in text prompts to adapt models like CLIP to downstream tasks. Unlike manual prompt engineering, CoOp automatically learns task-specific contextual representations through backpropagation. Given input image and class label , the text prompt is constructed as:

Here are learnable context vectors (each matching CLIP’s text embedding dimension), and is the class name embedding. During training, these context vectors are optimized to minimize prediction loss while keeping CLIP’s original image and text encoders ( and ) frozen. Classification probability is computed as:where represents cosine similarity, is the temperature parameter, and denotes the number of classes. CoOp substantially boosts few-shot performance (e.g., increasing CLIP’s accuracy by 10%–20% in 1–16 shot tasks) while demonstrating strong cross-dataset generalization and maintaining parameter efficiency.

Although CoOp shows strong capabilities in zero-shot recognition tasks for vision-language models, it faces significant challenges in fine-grained ship recognition under cloud occlusion in remote sensing imagery. Standard prompt learning struggles to overcome visual feature degradation due to heavy cloud obstruction, causing misalignment between image semantics and textual class prompts while hindering the capture of critical fine-grained visual features (e.g., distinguishing warships from cargo vessels). Furthermore, cloud corruption degrades visual features, increasing model sensitivity to label noise. Heavily occluded samples exhibit ambiguous representations, which amplify the impact of mislabeled data during training.

We therefore propose the cloud-adaptive robust prompt (CARP)—a novel prompt learning paradigm for cloud-occluded scenarios. Building upon the CoOp framework, CARP systematically addresses cloud interference through two strategies: 1) Introducing Adaptive Optimization Prompt Design (AOPD) and Dynamic Weight Adjustment Mechanism (DWAM) modules to resolve visual-semantic misalignment and fine-grained feature extraction; 2) Replacing traditional cross-entropy (CE) loss with generalized cross entropy (GCE) to mitigate label noise sensitivity. By incorporating a tunable parameter , GCE adaptively balances attention to hard samples (potentially noisy) while maintaining high-confidence predictions for clear samples, thereby reducing negative impacts from noisy labels. CARP provides the first end-to-end robust prompt learning solution for fine-grained ship recognition under cloud occlusion in remote sensing (see structure in Figure 1).

FIGURE 1

3.2 Adaptive optimization prompt design (AOPD)

To enhance model robustness against cloud occlusion, we propose an Adaptive Optimization Prompt Design (structure shown in Figure 2). This integrates a compensation vector dynamically into the prompt learning process, forming a dual-stream prompt architecture: one stream employs standard learnable context vectors to capture task-related general semantics, while the other stream utilizes a specifically learned occlusion compensation vector to adaptively adjust and enhance semantic prompts in response to potential occlusion patterns in input images.

FIGURE 2

The Occlusion Compensation Vector —a learnable parameter with the same dimensionality as —strengthens model resilience to visual occlusions. This mechanism optimizes prompt embedding generation by dynamically compensating for representation deviations in occluded regions. Its core operational principle involves:

Here, denotes the original context vector, and —a compensation parameter optimized via backpropagation and initialized from a zero-mean Gaussian distribution—is jointly optimized with during training. This produces the final prompt embedding:which dynamically handles visual occlusion scenarios, significantly enhancing representation capability with missing visual information. This design allows prompts to perceive and compensate for occlusion interference, improving text-visual feature correlation under occlusions.

3.3 Dynamic weight adjustment mechanism (DWAM)

When input images suffer severe cloud occlusion, degrading visual feature reliability and impeding text-visual alignment, we propose the Dynamic Weight Adjustment Mechanism (DWAM). This strategy adaptively reduces visual modality weighting while enhancing reliance on language inputs (e.g., reliable categorical text descriptions), maintaining robust semantic alignment under heavy occlusion. Crucially in few-shot scenarios, this adaptive balancing efficiently utilizes limited information and prevents overfitting to noisy data.

DWAM captures real-time multimodal gradients via backpropagation, generating dynamic weight coefficient from gradient -norms:where and denote image/text encoder gradient norms, and is the Sigmoid function. The coefficient re-weights features:

Normalized features compute cosine similarity:

This dynamically allocates representation weights based on per-modality optimization difficulty, where is a learnable log scale, and , are original image/text features.

3.4 Generalized cross entropy (GCE)

To address label noise sensitivity from cloud occlusion, we replace conventional cross-entropy (CE) with generalized cross entropy (GCE). This module dynamically balances noise robustness and training efficiency via tunable parameter : degenerating to standard CE when and equaling MAE loss at . It is defined as:where is sample ’s predicted probability for its true class, and is batch size. Experiments confirm optimally adapts to remote sensing cloud occlusion: retaining CE’s fast convergence while suppressing cloud-induced label noise overfitting through moderate loss curvature adjustment. For enhanced robustness, gradient backpropagation halts when prediction confidence :

Combined with AOPD’s compensation vector, this dual robust optimization significantly boosts generalization under degraded visual features. When approaches 0, GCE degenerates into the standard cross-entropy (CE), which is sensitive to noisy samples. When , it is equivalent to the mean absolute error (MAE), which enhances noise tolerance but impairs the efficiency of gradient updates. Experiments show that achieves the optimal balance in the cloud occlusion noise environment—it not only suppresses the interference of label noise by adjusting the loss curvature but also retains the discriminative ability for key ship features. Algorithm 1 provides complete CARP pseudocode.

4 Experiments

In this section, we first introduce SeaCloud-Ship, a dataset specifically constructed for few-shot ship fine-grained classification under cloud occlusion conditions. Based on this dataset, we conduct comprehensive experiments to evaluate the performance of the proposed method under varying levels of cloud occlusion and validate the effectiveness of the cloud-adaptive robust prompt (CARP). The experiments cover dataset construction details, baseline comparisons, implementation parameters, and multi-dimensional result analysis.

4.1 Datasets

With the growing demand for cloud detection technology, an increasing number of open-source datasets (Aybar et al., 2022; Foga et al., 2017; Mohajerani et al., 2019; Shendryk et al., 2019) have become available for training purposes. However, the absence of fine-grained ship recognition datasets under cloud occlusion conditions hinders validation of model robustness in such scenarios. To address this, we developed SeaCloud-Ship—a publicly accessible benchmark dataset specifically designed for fine-grained ship classification under cloud interference, with Figure 3 illustrating the dataset construction workflow.

FIGURE 3

4.1.1 Source screening

Due to the lack of open-source fine-grained ship datasets under cloud occlusion, the challenges in collecting such data, the inability to cover all common categories, and the difficulty in controlling cloud coverage ratios to accurately validate model performance, our data primarily derives from three public fine-grained ship classification datasets: FGSC-23 covers 23 categories (22 ship types and 1 non-ship category) with 4,080 samples. Partially sourced from the HRSC2016 dataset, its main imagery comes from Google Earth and GF-1 satellite, covering vessels under various lighting conditions, land/sea backgrounds, and arbitrary orientations. FGSCR-42 comprises 9,320 images across 42 common ship categories (including naval and civilian vessels), with approximately 200 images per class. Image resolutions range from 50 50 to 1,500 1,500 pixels, reflecting multi-scale ship features. Sources include Google Earth and mainstream remote sensing datasets like DOTA, HRSC 2016, and NWPU VHR-10. FGSCM-52 expands FGSCR-42 by adding 10 categories, forming 52 fine-grained ship classes (e.g., warships, civilian vessels). It incorporates multi-scale optical remote sensing imagery (50 50 to 1,600 1700 pixels, 0.4–2 m resolution from GF-2/Sentinel-2), integrating multi-source data from Google Earth, DOTA, and HRSC 2016. This dataset enhances annotations for rare hull types (e.g., auxiliary vessel subtypes) and adds complex scenarios (ports, offshore areas), emphasizing feature variations under diverse lighting, angles, and backgrounds.

4.1.2 Selection rule

To construct a cloud-interference-aware fine-grained ship recognition benchmark with scenario generalization capabilities, we selected source dataset samples using a three-tiered criterion: First, First, ensure the completeness of category coverage, and prioritize the selection of multi-view and multi-scale samples to retain fine-grained discriminative features; Second, guaranteeing scene diversity by mandating coverage across four typical scenarios: open waters, coastal zones, complex channels, and extreme illumination conditions; Finally, applying quality constraints to retain relatively high-resolution images inherently free from cloud occlusion, thereby preventing pre-existing interference. Through this process, 7,654 high-quality images were selected to form the foundational dataset. Controllable occlusion was applied via the automated cloud generation system (CloudGEN), while strictly preserving the original directory structure and class balance. Figure 4 demonstrates visual comparisons under varying occlusion ratios, validating the effectiveness of data simulation.

FIGURE 4

4.1.3 Cloud generation method

Based on the rigorously curated high-quality ship image dataset, we developed the automated cloud generation system CloudGEN, designed to controllably add simulated cloud layers with varying occlusion ratios. While existing cloud generation methods approach realistic cloud quality, genuine cloud occlusion data presents challenges in controlling obstruction ratios and collecting sufficient category coverage. Consequently, we employ an automated approach for generating cloud occlusion ratios. To authentically replicate real cloud interference characteristics in remote sensing imagery, we utilize a Perlin noise-based physical simulation method. By modeling optical properties under diverse cloud thicknesses, density distributions, and lighting conditions, we dynamically generate realistic cloud noise matching actual satellite observations directly onto raw ship images. First, we dynamically create naturalistic cloud textures using a fractal noise algorithm, adjusting octaves and persistence parameters to control cloud complexity and distribution. Second, a dichotomy-based threshold calibration technique precisely governs cloud coverage, ensuring accurate occlusion ratios per image. Finally, Gaussian blurring achieves optical transitions at cloud edges, while overlaying spectrally specific cloud colors simulates authentic atmospheric scattering. This process batch-generates cloud-occluded samples conforming to remote sensing imaging principles while preserving the original dataset’s directory structure. Figure 4 illustrates comparative results under varying cloud occlusion ratios.

4.1.4 Statistic

The SeaCloud-Ship dataset comprises 30 fine-grained ship categories (detailed in Table 1), spanning warships (e.g., destroyers, aircraft carriers, frigates), civilian vessels (e.g., yachts, cargo ships), and specialized ships (e.g., hospital ships, tugboats). It contains 7,654 high-quality optical remote sensing images with a mean resolution of 512 512 pixels, ranging from low-scale (92 92) to high-scale (1024 1024). The dataset integrates multi-source heterogeneous data, including Google Earth imagery, China’s Gaofen satellites (GF-1/GF-2), the European Sentinel-2 satellite, and aerial platforms like HRSC2016 and DOTA, primarily using high-resolution optical sensors (0.4–2 m). Samples originate from global maritime zones—Pacific, Atlantic, and Indian Oceans—covering diverse scenarios such as open seas, coastal waters, complex channels, and extreme lighting. The largest category, Towing_vessel, includes 778 images, averaging 255 per class with balanced military-civilian ratios. High-resolution classes focus on critical military targets: aircraft carriers (Nimitz-class averaging 812 812 pixels) and destroyers (Arleigh Burke-class averaging 540 540 pixels). Civilian classes like container ships (mean 442 442 pixels) and sand carriers (Sand_carrier averaging 230 230 pixels) emphasize low-resolution robustness testing. This multi-scale, multi-source, multi-scenario framework establishes SeaCloud-Ship as the first benchmark dataset supporting fine-grained vessel recognition under cloud-occluded conditions.

4.2 Datasets and baselines

Datasets Building upon (Di et al., 2021; Lan et al., 2024; Zhang et al., 2020), we construct the first cloud occlusion benchmark dataset SeaCloud-Ship. Using CloudGEN4.1.3, we synthesize progressive cloud occlusion (coverage levels: 12.5%, 25%, 37.5%, 50%, 62.5%, 75%), resulting in 7,654 images.

Baselines Four state-of-the-art prompt learning methods are selected for comparison: (1) CoOp (Zhou et al., 2022b): classical context optimization; (2) CoOp + MAE: incorporates masked autoencoder pretraining on CoOp for enhanced robustness; (3) CoCoOp (Zhou et al., 2022a): conditionally generated instance-level prompts; (4) NLPrompt (Pan et al., 2024): language-vision alignment method specifically designed for noisy scenarios, serving as the strongest relevant baseline.

4.3 Implementation details

Experiments utilize the CLIP-ViT/B-16 architecture with uniform training settings: SGD optimizer (initial learning rate , momentum 0.9, weight decay ) coupled with cosine learning rate decay, batch size of 32 (16 samples per visual/text modality) across 200 epochs. Generalized cross entropy (GCE) employs noise tolerance parameter . The Dynamic Weight Adjustment Mechanism (DWAM) adopts a GRU structure (hidden dimension 64), maintains a gradient history buffer of length 10, and dynamically weights features by constraining .

Context learning configuration includes 16-dimensional context vectors with class tokens at text suffix positions; compensation vectors follow initialization. Full training employs FP16 mixed precision. Hardware: NVIDIA GeForce RTX 3080 GPU and Intel Core i9-10920X CPU. Software: PyTorch 2.4.1/CUDA 12.1/cuDNN 8.0.5. Data augmentation includes random cropping (scale = [0.08, 1.0]), horizontal flipping, and ImageNet normalization. Reported results average three independent runs.

4.4 Main results

As shown in Table 2, our proposed method demonstrates significant advantages in remote sensing image recognition under cloud occlusion. It achieves optimal performance across all 30 comparative experiments, particularly excelling under high occlusion (75.00%) and few-shot (1-shot) scenarios. Compared to the best-performing baseline (CoOp + MAE), our method improves average accuracy by 3.6%, with the maximum gain in the 16-shot/12.50% scenario (61.03% vs. 56.56%). These results validate the effectiveness of our cloud-feature adaptive learning mechanism.

Our method exhibits exceptional robustness across varying occlusion levels. When occlusion increases from 12.50% to 75.00%, the best baseline (CoOp + MAE) shows 46.16% performance degradation under 1-shot conditions, while our method maintains degradation within 51.97%. Notably at the critical 50.00% occlusion threshold, our 1-shot accuracy reaches 23.63%, significantly exceeding CoOp + MAE (18.22%) and NLPrompt (16.55%). Even under extreme 75.00% occlusion, our 16-shot result (36.66%) outperforms the best baseline by 1.31 percentage points.

Training sample size analysis reveals pronounced few-shot advantages. Under 1-shot conditions, our method improves average accuracy by 4.27% across occlusion levels; this advantage expands to 4.82% at 16-shot. Specifically in 4-shot scenarios at 62.50% occlusion, our method achieves 28.33% accuracy—a 3.00 percentage point improvement (11.84% relative gain) over CoCoOp (25.33%). These findings indicate that increased training samples enhance our method’s ability to learn occlusion-invariant features.

4.5 Ablation experiments

As shown in Table 3, we systematically evaluate the synergistic effects of generalized cross entropy (GCE), Adaptive Optimization Prompt Design (AOPD), and Dynamic Weight Adjustment Mechanism (DWAM). Experiments demonstrate that individually introducing any module improves baseline performance, with DWAM yielding a 3.1% gain under 75% occlusion, highlighting its adaptability to extreme occlusion. Further analysis reveals functional complementarity: GCE enhances noise robustness, AOPD optimizes occlusion feature reconstruction, and DWAM achieves cross-modal dynamic balancing. When integrating all modules, the model reaches 50.30% accuracy under 50% occlusion—a 0.6 percentage point improvement over baseline, validating the efficacy of collaborative module design.

Loss function analysis (Table 4) uncovers key mechanisms: GCE with significantly outperforms standard CE and MAE losses, showing a 2.10 percentage point improvement over CE at 25% occlusion. This stems from GCE’s noise tolerance, dynamically weighting hard samples via to suppress false signals in cloud-occluded regions. Notably, exhibits sensitivity thresholds: at , diminished gradient smoothing causes noticeably lower performance gains under 75% occlusion compared to the configuration.

5 Conclusion

This paper addresses the critical challenge of few-shot fine-grained ship classification in cloud-occluded remote sensing by proposing CARP (cloud-adaptive robust prompt)—an innovative framework that systematically tackles feature corruption, semantic misalignment, and data utility degradation. By designing a generalized cross entropy (GCE) loss for noise immunity, developing an Adaptive Optimization Prompt Design (AOPD) to repair semantic mismatches, and introducing a Dynamic Weight Adjustment Mechanism (DWAM) for cross-modal alignment, our method achieves breakthrough performance on the SeaCloud-Ship dataset.

As shown in Table 2, CARP outperforms all existing methods across 30 experimental settings, notably achieving 23.6% accuracy under 1-shot/50% occlusion (5.38% higher than CoOp + MAE) and setting a new state-of-the-art of 36.66% at 16-shot/75% occlusion. Experiments confirm CARP’s superior robustness: when occlusion increases from 12.5% to 75.0%, its performance degradation (52.2%) is significantly lower than baselines (46.2%). The newly constructed SeaCloud-Ship dataset addresses a critical domain gap, providing reliable solutions for satellite imaging systems. Future work will explore multi-modal cloud occlusion modeling and zero-shot generalization enhancement.

References

• 1

  AybarC.YsuhuaylasL.LojaJ.GonzalesK.HerreraF.BautistaL.et al (2022). CloudSEN12, a global dataset for semantic understanding of cloud and cloud shadow in Sentinel-2. Sci. Data9, 782. 10.1038/s41597-022-01878-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  BaoH.WangW.DongL.LiuQ.MohammedO. K.AggarwalK.et al (2022). Vlmo: unified vision-language pre-training with mixture-of-modality-experts. Adv. Neural Inf. Process. Syst.35, 32897–32912.

  • Google Scholar

• 3

  ChenY.-C.LiL.YuL.El KholyA.AhmedF.GanZ.et al (2020). “Uniter: universal image-text representation learning,” in European conference on computer vision (Springer), 104–120.

  • Google Scholar

• 4

  DiY.JiangZ.ZhangH. (2021). A public dataset for fine-grained ship classification in optical remote sensing images. Remote Sens.13 (4), 747. 10.3390/rs13040747

  • CrossRef
  • Google Scholar

• 5

  FogaS.ScaramuzzaP. L.GuoS.ZhuZ.DilleyR. D.JrBeckmannT.et al (2017). Cloud detection algorithm comparison and validation for operational Landsat data products. Remote Sens. Environ.194, 379–390. 10.1016/j.rse.2017.03.026

  • CrossRef
  • Google Scholar

• 6

  Harold LiL.YatskarM.YinD.HsiehC. J.ChangK. W. (2019). Visualbert: a simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557.

  • Google Scholar

• 7

  HuangL.WangF.ZhangY.XuQ. (2022). Fine-grained ship classification by combining CNN and swin transformer. Remote Sens.14, 3087. 10.3390/rs14133087

  • CrossRef
  • Google Scholar

• 8

  JiaC.YangY.XiaY.ChenY.-T.ParekhZ.PhamH.et al (2021). “Scaling up visual and vision-language representation learning with noisy text supervision,” in International conference on machine learning (PMLR), 4904–4916.

  • Google Scholar

• 9

  JiaM.TangL.ChenB.-C.CardieM.BelongieS.HariharanB.et al (2022). “Visual prompt tuning,” in European conference on computer vision (Springer), 709–727.

  • Google Scholar

• 10

  LanL.WangF.ZhengX.WangZ.LiuX. (2024). Efficient prompt tuning of large vision-language model for fine-grained ship classification. IEEE Trans. Geoscience Remote Sens.63, 1–10. 10.1109/tgrs.2024.3509721

  • CrossRef
  • Google Scholar

• 11

  LeeK.-H.ChenX.HuaG.HuH.HeX. (2018). “Stacked cross attention for image-text matching,” in Proceedings of the European conference on computer vision (ECCV), 201–216.

  • Google Scholar

• 12

  LiW.GaoC.NiuG.XiaoX.LiuH.LiuJ.et al (2020a). Unimo: towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409.

  • Google Scholar

• 13

  LiX.LiC.ZhangP.HuX.ZhangL.WangL.et al (2020b). “Oscar: objectsemantics aligned pre-training for vision-language tasks,” in Computer Vision–ECCV 2020: 16th European Conference Proceedings, Glasgow, United Kingdom, August 23–28, 2020 (Springer), 121–137.

  • Google Scholar

• 14

  LiJ.SelvarajuR. R.GotmareA. D.JotyS.XiongC.HoiS. (2021). Align before fuse: vision and language representation learning with momentum distillation. Adv. Neural Inf. Process. Syst.34, 9694–9705.

  • Google Scholar

• 15

  LuJ.BatraD.ParikhD.LeeS. (2019). “ViLBERT: pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,” in Advances in neural information processing systems.

  • Google Scholar

• 16

  LuX.YueT.CaiJ.ChenY.LvC.ChuS. (2025). MSCA-net: multi-scale context aggregation Network for infrared small target detection. arXiv preprint arXiv:2503.17193.

  • Google Scholar

• 17

  MohajeraniS.SaeediP. (2019). “Cloud-Net: an end-to-end cloud detection algorithm for Landsat 8 imagery,” in IGARSS 2019-2019 IEEE international geoscience and remote sensing symposium, Yokohama, Japan, 28 July 2019 - 02 August 2019 (IEEE), 1029–1032.

  • Google Scholar

• 18

  PanB.LiQ.TangX.HuangW.FangZ.LiuF.et al (2024). NLPrompt: noise-label prompt learning for Vision-Language Models. arXiv preprint arXiv:2412.01256.

  • Google Scholar

• 19

  RadfordA.KimJ. W.HallacyC.RameshA.GohG.AgarwalS.et al (2021). “Learning transferable visual models from natural language supervision,” in International conference on machine learning (PMLR), 8748–8763.

  • Google Scholar

• 20

  ShendrykY.RistY.TicehurstC.ThorburnP. (2019). Deep learning for multi-modal classification of cloud, shadow and land cover scenes in PlanetScope and Sentinel-2 imagery. ISPRS J. Photogrammetry Remote Sens.157, 124–136. 10.1016/j.isprsjprs.2019.08.018

  • CrossRef
  • Google Scholar

• 21

  SinghA.HuR.GowsamiV.GuillaumeG.GalubaW.RohrbachM.et al (2022). “Flava: a foundational language and vision alignment model,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, New Orleans, LA, USA, 18-24 June 2022 (IEEE), 15638–15650.

  • Google Scholar

• 22

  SuW.ZhuX.CaoY.LiB.LuL.WeiF.et al (2019). Vl-bert: pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530.

  • Google Scholar

• 23

  TanH.BansalM. (2019). Lxmert: learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490.

  • Google Scholar

• 24

  Uzair KhattakM.RasheedH.MaazM.KhanS.KhanF. S. (2023). “MaPLe: multi-modal prompt learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Vancouver, BC, Canada, 17-24 June 2023 (IEEE), 19113–19122.

  • Google Scholar

• 25

  WangZ.YuJ.YuA. W.DaiZ.TsvetkovY.CaoY. (2021). Simvlm: simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904.

  • Google Scholar

• 26

  WangF.HuangW.YangS.FanQ.LanL. (2024a). Learning to learn better visual prompts, 5354–5363.

  • Google Scholar

• 27

  WangF.WangH.WangD.GuoZ.ZhongZ. (2024b). Scaling efficient masked autoencoder learning on large remote sensing dataset. arXiv preprint arXiv:2406.11933.

  • Google Scholar

• 28

  WangF.ChenM.LiY.WangD.WangH.GuoZ.et al (2025a). GeoLLaVA-8K: scaling remote-sensing multimodal large Language Models to 8K resolution. arXiv preprint arXiv:2505.21375.

  • Google Scholar

• 29

  WangF.WangH.WangY.WangD.ChenM.ZhaoH.et al (2025b). RoMA: scaling up mamba-based foundation models for remote sensing. arXiv preprint arXiv:2503.10392.

  • Google Scholar

• 30

  XuC.ZhengX.LuX. (2022). Multi-level alignment network for cross-domain ship detection. Remote Sens.14, 2389. 10.3390/rs14102389

  • CrossRef
  • Google Scholar

• 31

  ZhangX.LvY.YaoL.XiongW.FuC. (2020). A new benchmark and an attribute-guided multilevel feature representation network for fine-grained ship classification in optical remote sensing images. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens.13, 1271–1285. 10.1109/jstars.2020.2981686

  • CrossRef
  • Google Scholar

• 32

  ZhangR.FangR.ZhangW.GaoP.LiK.DaiJ.et al (2021). Tip-adapter: training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930.

  • Google Scholar

• 33

  ZhangJ.HuangJ.JinS.LuS. (2024). Vision-language models for vision tasks: a survey. IEEE Trans. Pattern Analysis Mach. Intell.46, 5625–5644. 10.1109/tpami.2024.3369699

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  ZhengX.ChenX.LuX.SunB. (2021). Unsupervised change detection by cross-resolution difference learning. IEEE Trans. Geoscience Remote Sens.60, 1–16. 10.1109/tgrs.2021.3079907

  • CrossRef
  • Google Scholar

• 35

  ZhengX.CuiH.XuC.LuX. (2023). Dual teacher: a semisupervised cotraining framework for cross-domain ship detection. IEEE Trans. Geoscience Remote Sens.61, 1–12. 10.1109/tgrs.2023.3287863

  • CrossRef
  • Google Scholar

• 36

  ZhouK.YangJ.LoyC. C.LiuZ. (2022a). “Conditional prompt learning for vision-language models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, New Orleans, LA, USA, 18-24 June 2022 (IEEE), 16816–16825.

  • Google Scholar

• 37

  ZhouK.YangJ.LoyC. C.LiuZ. (2022b). Learning to prompt for vision-language models. Int. J. Comput. Vis.130 (9), 2337–2348. 10.1007/s11263-022-01653-1

  • CrossRef
  • Google Scholar

• 38

  ZhuB.NiuY.HanY.WuY.ZhangH. (2023). “Prompt-aligned gradient for prompt tuning,” in Proceedings of the IEEE/CVF international conference on computer vision, Paris, France, 01-06 October 2023 (IEEE), 15659–15669.

  • Google Scholar