CL-ODGAN: an unpaired attention-guided GAN framework for remote sensing image dehazing Authors: Hongwei Zhang, Xin Mu, Bin Yin Published: 2026 ABSTRACT Haze in remote sensing imagery severely degrades image quality and impedes accurate interpretation. While deep learning has advanced dehazing, its reliance on large-scale paired data is impractical for remote sensing applications. To overcome this, we propose CL-ODGAN, a novel unpaired dehazing framework built upon a systematically designed generative architecture and a dual-level contrastive learning constraint. Our framework is architected around three synergistic components: (1) A robust generator that employs custom Residual blocks embedded with Omni-dimensional Dynamic Convolution (ODConv) for dynamic multi-scale perception and a hybrid normalization (BN + IN) strategy for training stability. This design enables adaptive feature extraction and detail restoration under non-uniform haze. (2) A core unsupervised supervisory signal implemented via a dual-level (feature and pixel) contrastive learning mechanism within the GAN framework. This mechanism guides the image translation process without paired data by enforcing structural and feature consistency. (3) An attention-guided adversarial training paradigm that focuses the network on critical hazy regions, ensuring efficient learning. In essence, CL-ODGAN integrates ODConv and hybrid normalization withinthe generator to enhance its representational capacity, while employing contrastive learning as the supervisory engine—all orchestrated under an unpaired, attention-aware GAN framework. Experiments demonstrate its superior performance in removing non-uniform haze and recovering fine details without requiring paired training data. Introduction Detection and interpretation of remote sensing imagery play a vital role in various fields including climate prediction, agricultural yield assessment, natural disaster monitoring, mineral resource exploration, and military surveillance1–4. However, due to interference factors such as haze and clouds in the atmosphere, captured remote sensing images often suffer from severe quality degradation manifested as reduced contrast, color distortion, and blurred details, which significantly afects subsequent image analysis and application effectiveness. Research methods in remote sensing image dehazing have primarily undergone an important evolution from physics-based model-driven to data-driven approaches. Traditional methods mainly rely on the atmospheric scattering physical model, with the most representative being the Dark Channel Prior (DCP)5 and Color Attenuation Prior6. With the rapid development of deep learning technology, data-driven methods such as AOD-Net7 and GridDehazeNet8 have achieved significant breakthroughs in dehazing performance. However, current remote sensing image dehazing techniques still face difficulties in obtaining precisely paired hazy/clear image datasets in real-world scenarios; additionally, existing methods show inadequate performance in handling non-uniform haze distributions. In recent years, unsupervised learning has demonstrated significant value in the image dehazing field. The successful application of CycleGAN9 in unpaired image translation tasks established the foundation for this direction. Subsequently, CycleDehaze10 (Engin et al., 2018) first introduced an unsupervised learning framework into image dehazing tasks, effectively alleviating the dependency on paired training data. Zheng et al.11 proposed Dehaze-AGGAN (Attention-guided GAN) which further enhanced model performance by incorporating attention mechanisms. However, these methods still suffer from insufficient feature representation capability and inefficient cross-scale feature fusion in remote sensing image dehazing tasks. To address the core challenges in remote sensing image dehazing, such as strong dependence on paired data, insufficient generalization ability, and poor detail recovery, this paper proposes CL-ODGAN, an innovative unpaired dehazing framework. The main contributions are succinctly summarized as follows:An Integrated Dynamic Generator with Hybrid Normalization: We design a novel generator that deeply integrates Omni-Dimensional Dynamic Convolution (ODConv) with a Batch-Instance Normalization scheme. This architecture enables dynamic feature adaptation to non-uniform haze while preserving scene details, effectively improving detail recovery.A Dual-Level Contrastive Learning Strategy for Unpaired Supervision: We propose a novel feature-level and pixel-level contrastive learning mechanism within the GAN framework. This strategy guides the network to learn haze-invariant features and ensures structural consistency without relying on paired data, fundamentally breaking the paired-data dependency bottleneck.A Stable and Generalizable Training Framework: The synergistic design of the dual-level contrastive loss and the hybrid normalization scheme stabilizes training on unpaired data and enhances the model’s generalization capability across diverse scenarios and multi-source remote sensing datasets. An Integrated Dynamic Generator with Hybrid Normalization: We design a novel generator that deeply integrates Omni-Dimensional Dynamic Convolution (ODConv) with a Batch-Instance Normalization scheme. This architecture enables dynamic feature adaptation to non-uniform haze while preserving scene details, effectively improving detail recovery. An Integrated Dynamic Generator with Hybrid Normalization: We design a novel generator that deeply integrates Omni-Dimensional Dynamic Convolution (ODConv) with a Batch-Instance Normalization scheme. This architecture enables dynamic feature adaptation to non-uniform haze while preserving scene details, effectively improving detail recovery. A Dual-Level Contrastive Learning Strategy for Unpaired Supervision: We propose a novel feature-level and pixel-level contrastive learning mechanism within the GAN framework. This strategy guides the network to learn haze-invariant features and ensures structural consistency without relying on paired data, fundamentally breaking the paired-data dependency bottleneck. A Dual-Level Contrastive Learning Strategy for Unpaired Supervision: We propose a novel feature-level and pixel-level contrastive learning mechanism within the GAN framework. This strategy guides the network to learn haze-invariant features and ensures structural consistency without relying on paired data, fundamentally breaking the paired-data dependency bottleneck. A Stable and Generalizable Training Framework: The synergistic design of the dual-level contrastive loss and the hybrid normalization scheme stabilizes training on unpaired data and enhances the model’s generalization capability across diverse scenarios and multi-source remote sensing datasets. A Stable and Generalizable Training Framework: The synergistic design of the dual-level contrastive loss and the hybrid normalization scheme stabilizes training on unpaired data and enhances the model’s generalization capability across diverse scenarios and multi-source remote sensing datasets. Network Prior-based dehazing via the atmospheric scattering model The atmospheric scattering model serves as the theoretical foundation of the dehazing field. Its formula I(x) = J(x)t(x) + A(1—t(x)) describes the image formation process in the atmospheric environment. Based on this model, He et al.5 proposed the dark channel prior, which assumes that the minimum value in local regions of a haze-free image is close to zero, thereby recovering the haze-free image by estimating the atmospheric light value and the transmission map. However, the dark channel prior may fail in sky regions or near bright objects because the dark channel assumption does not hold in these areas. To address this issue, He et al.12 proposed a novel haze removal algorithm named HALP, based on a heterogeneous atmospheric light prior. Furthermore, Liang et al.13 developed an Atmospheric Light value Fusion Estimation (ALFE) model, which integrates a Heterogeneous Prior-guided Haze Concentration Model (HP-HCM) to derive the global atmospheric light value. Simultaneously, Golts et al.14 proposed an unsupervised training method that achieves dehazing by minimizing the Dark Channel Prior (DCP) energy function. It is noteworthy that during the development of prior-based dehazing, several foundational and complementary core works have emerged. For example, Fattal21 pioneered an image statistics-based prior, which restores images under the assumption that the transmission is uncorrelated with the scene albedo, laying the groundwork for the field. His subsequent research introduced the Color-Lines Prior22, which further leverages the geometric properties of pixel distributions in RGB space to estimate the transmission map. Tan23 adopted a different approach by employing a Markov Random Field-based contrast maximization strategy, achieving dehazing without directly relying on the physical model. The most influential Dark Channel Prior5 provides a powerful and intuitive transmission estimation scheme by statistically analyzing the low-intensity characteristics in local regions of haze-free images. To address the failure of the dark channel prior in specific scenarios, subsequent studies proposed various improved priors. For instance, Berman et al.24 introduced the Haze-Line prior, which utilizes the distribution patterns of non-local pixels in color space to handle non-uniform haze. Zhu et al.6 proposed a computationally efficient Color Attenuation Prior, which rapidly estimates the transmission map by establishing a linear relationship between the difference in image saturation and brightness and the scene depth. These hand-crafted priors, based on different physical insights and statistical regularities, collectively constituted the core technical framework of single image dehazing before the deep learning era and have continued to provide theoretical inspiration for unsupervised and physics-driven dehazing methods. Learning-based remote sensing image dehazing With the advancement of deep learning, various learning-based models have been utilized for image dehazing. Cai et al.15 first proposed DehazeNet, which employs convolutional neural networks (CNNs) to estimate the transmission map and combines it with the atmospheric scattering model to achieve end-to-end dehazing. Subsequently, Li et al.7 proposed AOD-Net, which directly predicts clear images through a lightweight network, avoiding explicit estimation of the transmission map and atmospheric light. In recent years, Generative Adversarial Networks (GANs) have demonstrated significant advantages in dehazing tasks. Engin et al.10 proposed Cycle-Dehaze, based on the CycleGAN framework, enabling dehazing training without paired data and addressing the scarcity of paired data in real-world scenarios. Meanwhile, end-to-end CNN-based dehazing methods continue to evolve. Zhang et al.25 proposed a densely connected pyramid dehazing network, enhancing performance through feature reuse. Liu et al.26 designed GridDehazeNet, an attention-based multi-scale network that effectively aggregates features at different scales. Likewise dedicated to strengthening feature representation, Xie et al.27 designed the EFABA and GDAL modules, which achieve accurate edge localization and global context modeling, respectively, and enhance the feature representation capability of remote sensing images. Qin et al.28 introduced the Feature Fusion Attention Network (FFA-Net), focusing on key regions with uneven haze distribution. With the rise of Transformers in the field of computer vision, they have also been applied to dehazing tasks. Furthermore, in the context of remote sensing, Xie et al.29 designed the TMCFM and FSDC modules along with a DSC classifier, focusing on the extraction and fusion of multi-scale contextual features. The multi-feature constraints and context-aware framework proposed by Zhu et al.30 provide new ideas for dehazing remote sensing images.Song et al.31 first explored the application of Vision Transformers to single image dehazing. Guo et al.32 further proposed an Image Dehazing Transformer with transmission-aware 3D position embedding, better modeling the spatial distribution characteristics of haze. Additionally, self-supervised techniques such as contrastive learning have been successfully introduced into the dehazing field. Wu et al.33 proposed a compact single image dehazing framework based on contrastive learning, improving performance by pulling the features of clear images close and pushing those of hazy images apart. Most recently, two advanced methods have been proposed for remote sensing image dehazing. Dehazing-DiffGAN36 combines diffusion models and GANs in a two-stage pipeline to achieve high-fidelity image restoration. Meanwhile, INP-Net37 introduces an implicit neural prompting mechanism and an ASP Transformer block to enhance global modeling and feature diversity. Both works share the same goal of improving dehazing performance for remote sensing imagery. While their technical paths differ from our encoder-decoder and contrastive learning framework, there is potential synergy between their advanced generative/implicit modeling strategies and our feature discrimination approach, which could be explored in future work. To further clarify the differences between existing methods and our work, we summarize the key characteristics of mainstream dehazing methods in the following Table 1. Table 1Comparison of key characteristics of mainstream dehazing methods (× : Not supported, △: Partially supported, √: Fully supported).MethodRS-optimizedUneven hazeDCP5 × × AOD-Net7 × △GridDehazeNet8 × △CL-ODGAN √ √ Table 1Comparison of key characteristics of mainstream dehazing methods (× : Not supported, △: Partially supported, √: Fully supported).MethodRS-optimizedUneven hazeDCP5 × × AOD-Net7 × △GridDehazeNet8 × △CL-ODGAN √ √ Comparison of key characteristics of mainstream dehazing methods (× : Not supported, △: Partially supported, √: Fully supported). Method Overall network architecture This paper proposes CL-ODGAN, a network module designed for unpaired remote sensing image dehazing . The framework comprises three core modules: generator, discriminator, and self-supervised feature constraint module (Dual-Level Contrastive Learning Strategy). The generator, built upon a ResNet-50 backbone, incorporates an enhanced multi-scale hybrid attention module within its encoder. This module integrates spatial and channel attention to enable adaptive screening of multi-scale features. A cross-scale feature fusion mechanism in the deeper network layers aggregates shallow detail information with deep semantic information. The decoder employs a progressive upsampling strategy alongside a reconstruction module that combines residual connections and attention gating, thereby restoring image resolution while maximally preserving structural and textural details. The discriminator employs a 70 × 70 PatchGAN to ensure training stability with unpaired data, and complements the generator’s global feature learning, jointly enhancing dehazing precision while avoiding local artifacts. The self-supervised feature constraint module, optimized via contrastive learning, establishes a dual constraint mechanism at the feature and pixel levels. At the feature level, it constructs positive and negative sample pairs to align the generated dehazed features with real scene features in the high-level semantic space while distinguishing them from hazy features. At the pixel level, a local patch contrast strategy improves the fidelity and naturalness of local details in the output image. Additionally, the method adopts a multi-loss fusion strategy, combining adversarial, perceptual, structural similarity, and contrastive learning losses into a comprehensive objective to guide the network training collaboratively. The core innovation of this framework is the organic integration of the improved multi-scale attention mechanism with multi-level contrastive learning, coupled with the PatchGAN’s local discrimination architecture. This synergy enables effective removal of non-uniform haze in complex remote sensing scenes without paired training data and significantly enhances the detail preservation in the dehazed results. Improved multi-scale attention encoder The encoder of the generator is constructed upon a ResNet-50 backbone, with its core innovations strategically designed to address the unique challenges inherent in unpaired remote sensing dehazing tasks, such as non-uniform haze distribution and the critical requirement for preserving ground-object details. The principal enhancements are concentrated within three key aspects: adaptive multi-scale feature extraction, stable deep feature refinement, and effective cross-level information fusion. A innovation involves the integration of Omni-Dimensional Dynamic Convolution (ODConv) coupled with a parallel standard convolutional branch at the feature extraction stage. This design tackles the challenge of highly variable haze densities and scales across remote sensing images. The ODConv module adaptively adjusts its kernel parameters to capture these irregular patterns, while the parallel branch ensures the stability of fundamental local feature extraction. The outputs from both pathways are subsequently fused, enabling the network to respond dynamically to complex haze conditions without compromising structural integrity. To enhance training stability and feature preservation in deep layers, a hybrid normalization strategy is implemented within the nine improved residual blocks. Diverging from approaches reliant on a single normalization technique, this strategy employs Instance Normalization in shallow layers to adapt to individual image styles, and utilizes Batch Normalization in deeper layers to stabilize training dynamics via batch statistics. This combination effectively balances style adaptability with training robustness, thereby mitigating the issue of detail over smoothing commonly encountered in deep networks. Furthermore, a cross-level feature fusion module is devised to resolve the inconsistency between shallow spatial details and deep semantic information. By concatenating and compressing features from different hierarchical levels, this module selectively enhances information critical to the dehazing task, ultimately producing a balanced feature map that provides a solid foundation for the decoder to reconstruct a clear, detailed image. Decoder and image reconstruction The decoder is responsible for reconstructing the dehazed image from the encoder features. Its structure is symmetrical to the encoder, and at each level, it concatenates with the corresponding encoder features through skip connections, thereby integrating low-level details with high-level semantic information. The concatenated features are optimized through Instance Normalization and the ODBottleneck module, followed by progressive up sampling via transposed convolution to restore spatial details and structural integrity. The final output preliminary dehazed image exhibits high textural clarity and color consistency. Discriminator For the discriminator design, this study adopts the classic PatchGAN architecture. In contrast to discriminators with a single global decision output, the PatchGAN divides the input image into multiple local patches, performs an independent judgment for each patch, and finally aggregates these results. This mechanism involves fewer parameters and higher computational efficiency, which is particularly conducive to stabilizing the training process in unpaired data scenarios. Furthermore, its emphasis on local textures enables effective preservation of critical ground feature details in remote sensing images. In detail, the discriminator takes as input an image pair composed of the generated dehazed image and the original hazy image, which are concatenated along the channel axis. As a fully convolutional network (FCN), this discriminator uses 4 × 4 convolutional kernels with a stride of 2 for progressive down sampling to output a discrimination map. For the discriminator’s final output layer, each element in this map corresponds to a 70 × 70-pixel receptive field at the input image scale, representing the probability that the corresponding local regions are real. Contrastive learning process Inspired by the Contrastive Unpaired Translation (CUT) framework , this study develops an innovative dual-path contrastive learning framework for remote sensing image dehazing, effectively replacing the cycle-consistency constraint commonly used in unsupervised methods like CycleGAN. This framework establishes dense semantic correspondences between hazy inputs and clear targets through two synergistic components: feature-level and pixel-level contrastive learning. Feature-level contrastive learning: Specifically, we extract a feature vector from the hazy input image as the anchor feature (denoted as X). Meanwhile, a 8 × 8 block is cropped as the anchor block with the anchor feature point as the center(denoted as x). And the clear image block of the same size at the identical spatial location is selected as the positive sample (denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}^{+}$$\end{document}). Image blocks at other non-corresponding positions of the same clear target image are taken as negative samples (denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum}_{\mathrm{k}=1}^{\mathrm{N}}{x}^{-}$$\end{document}). Based on the constructed positive and negative sample pairs, we employ the Noise Contrastive Estimation (NCE) loss to optimize feature matching. Its mathematical expression is as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{L}}_{NCE}=-{\mathrm{E}}_{\mathrm{x},{\mathrm{x}}^{+}}\mathrm{log}\left[\frac{\mathrm{exp}\left(\uptau \cdot \mathrm{f}\left({\mathrm{x}}^{+}\right)\cdot \mathrm{f}\left(\mathrm{x}\right)\right)}{\mathrm{exp}\left(\uptau \cdot \mathrm{f}\left({\mathrm{x}}^{+}\right)\cdot \mathrm{f}\left(\mathrm{x}\right)\right)+{\sum}_{\mathrm{k}=1}^{\mathrm{N}}\mathrm{exp}\left(\uptau \cdot \mathrm{f}\left(\mathrm{x}\right)\cdot \mathrm{f}\left({\mathrm{x}}_{\mathrm{k}}^{-}\right)\right)}\right]$$\end{document} Here, τ is the temperature hyperparameter (default 0.7), f(·) denotes the encoder network. In order to achieve effective reconstruction of blurred images to clear images, on the basis of the structural consistency of feature-level contrastive learning, further introduces a pixel-level contrast learning strategy to improve the detail clarity and visual fidelity of the reconstruction results. To achieve this goal, we leverage the Pixel-wise Contrastive Loss function proposed by Xie et al.19 (2021). Specific steps: First, the dehazed output image and the clear reference image are input into the pre-trained VGG-16 network to extract shallow feature maps, and anchor sample pairs M are constructed. For the i feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left({\widehat{y}}_{i}\right)$$\end{document} in the feature map of the generated image, it is taken as the anchor; the feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left({y}_{i}^{+}\right)$$\end{document} extracted from the corresponding position in the feature map of the clear reference image is taken as its positive sample, and K feature vectors sampled from other noncorresponding regions,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum}_{j=1}^{K}f\left({y}_{ij}^{-}\right)$$\end{document} ,are defined as negative samples. Based on this, the pixel-level contrastive loss function is intended to narrow the distance between each anchor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left({\widehat{y}}_{i}\right)$$\end{document} and its positive sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left({y}_{i}^{+}\right)$$\end{document} in the feature space, while pushing the anchor away from all its negative samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left({y}_{ij}^{-}\right)$$\end{document} . Its specific formulation is given as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{\mathrm{P}ixCon}=-\frac{1}{M}{\sum}_{i=1}^{M}\mathrm{log}\frac{\mathrm{exp}\left(sim\left(f\left({\widehat{y}}_{i}\right),f\left({y}_{i}^{+}\right)\right)/\uptau \right)}{\mathrm{exp}\left(sim\left(f\left({\widehat{y}}_{i}\right),f\left({y}_{i}^{+}\right)\right)/\uptau \right)+{\sum}_{j=1}^{K}\mathrm{exp}\left(sim\left(f\left({\widehat{y}}_{i}\right),f\left({y}_{ij}^{-}\right)\right)/\uptau \right)}$$\end{document} By jointly optimizing complementary, our framework provides a more flexible and effective constraint mechanism than Cycle Consistency Loss, promoting high-frequency detail restoration while preserving critical visual content for clearer, more natural dehazing results in unpaired scenarios. Optimization and evaluation of image dehazing In the described optimization process, a comprehensive fused loss function is constructed to achieve efficient training and performance optimization of the dehazing network. This total loss function is defined as a weighted linear combination of several component losses, with the specific formulation as follows: The optimization process is achieved by constructing a combined loss function to update the network parameters. The total loss function is defined as follows:: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{total}=10{L}_{rec}+1{L}_{adv}+5{L}_{perc}+0.5{L}_{\mathrm{P}ixCon}+2{L}_{NCE}$$\end{document}, This combined loss function integrates the above components in a weighted manner. It drives the parameter updates of the dehazing network through backpropagation, ultimately achieving high-fidelity dehazing results with strong detail preservation capabilities. Experiments Experimental setup This section provides a comprehensive experimental evaluation of the dehazing performance. All experiments were implemented based on the PyTorch 1.12.1 framework and executed in parallel on 4 × NVIDIA RTX 3080 GPUs. The training process utilized the Adam optimizer 20 (Kingma and Ba, 2015) with an initial learning rate of 0.0001 (β₁ = 0.9, β₂ = 0.999), complemented by weight decay (1e-4) and gradient clipping (threshold 0.01) to stabilize training. The training strategy comprised a complete cycle of 200 epochs, which included a fixed learning rate for the first 100 epochs, followed by linear decay over the subsequent 100 epochs, and incorporated a 10-epoch warm-up phase. The experiments utilized the public SateHaze1k (Haze1k) remote sensing dehazing dataset 34. The dataset consists of three subsets: SateHaze1k-thin, SateHaze1k-moderate, and SateHaze1k-thick. Each subset contains 400 pairs of synthetic remote sensing hazy images. The data was partitioned such that 320 image pairs were used for training, 35 pairs for validation, and 45 pairs for performance evaluation. Since the StateHaze1K dataset is inherently paired, we intentionally employed a training procedure with shuffled data order to simulate a unpaired training scenario. This design is motivated by our goal to validate the robustness of our pixel-level contrastive learning framework in scenarios where one-to-one paired labels are unavailable, which aligns more closely with real-world dehazing tasks. Specifically, after shuffling, the clear reference images are no longer strictly paired with their original hazy counterparts. This method combines feature-level contrastive learning and pixel-level contrast learning to achieve effective dehazing and performance improvement without relying on paired data. Comparison with SOTA methods As shown in Table 2, to comprehensively evaluate the performance of the proposed unpaired dehazing method, we compare it with a variety of representative approaches, including the prior-based traditional method (DCP5), supervised CNN methods (AOD-Net7, GridDehazeNet8), and state-of-the-art methods based on Transformer and diffusion models in recent years (DehazeFormer31, WeatherDiffusion35). All comparative methods are tested with their original designs or publicly available pre-trained models, while our method strictly adheres to the unpaired training setting and only leverages unpaired hazy and clear images for learning. Table 2Quantitative comparison of dehazing methods on the Stathaze1K dataset under different haze levels (thin, moderate, thick). Metrics include PSNR (↑) and SSIM (↑), where higher values indicate better performance. The proposed CL-ODGAN (Ours) achieves the best results across all conditions.MethodStathaze1K -thinStathaze1K -moderateStathaze1K -thickPSNRSSIMPSNRSSIMPSNRSSIMDCP513.150.72469.7800.573510.250.5850AOD-Net719.540.854320.100.885415.920.7313GridDehazeNet824.340.875523.050.927420.240.8312DehazeFormer3120.540.867417.950.845616.710.8124WeatherDiffusion 3510.250.50149.500.44869.310.4156Ours28.520.899728.330.826628.150.8525 Table 2Quantitative comparison of dehazing methods on the Stathaze1K dataset under different haze levels (thin, moderate, thick). Metrics include PSNR (↑) and SSIM (↑), where higher values indicate better performance. The proposed CL-ODGAN (Ours) achieves the best results across all conditions.MethodStathaze1K -thinStathaze1K -moderateStathaze1K -thickPSNRSSIMPSNRSSIMPSNRSSIMDCP513.150.72469.7800.573510.250.5850AOD-Net719.540.854320.100.885415.920.7313GridDehazeNet824.340.875523.050.927420.240.8312DehazeFormer3120.540.867417.950.845616.710.8124WeatherDiffusion 3510.250.50149.500.44869.310.4156Ours28.520.899728.330.826628.150.8525 Quantitative comparison of dehazing methods on the Stathaze1K dataset under different haze levels (thin, moderate, thick). Metrics include PSNR (↑) and SSIM (↑), where higher values indicate better performance. The proposed CL-ODGAN (Ours) achieves the best results across all conditions. The quantitative results (Table 2) demonstrate that our method achieves the best or competitive performance on all key metrics under the unsupervised learning framework. Notably, several advanced methods that exhibit excellent performance in natural image dehazing (e.g., DehazeFormer, WeatherDiffusion) fail to deliver expected results in the remote sensing scenario of this experiment. We analyze that this may be attributed to the significant differences between natural and remote sensing images in terms of scene distribution, physical haze models and scale variations. Since these methods are not specifically optimized or trained for remote sensing data, their generalization ability is limited. This result further highlights the importance of designing dedicated dehazing models tailored to the characteristics of remote sensing scenarios, and also indirectly validates the robustness of our method in learning remote sensing domain-specific representations under unpaired conditions (Figs. 1, 2, 3 and 4). Fig. 1Overview of the CL-ODGAN network architecture for unpaired remote sensing image dehazing. Overview of the CL-ODGAN network architecture for unpaired remote sensing image dehazing. Fig. 2Generator architecture of the proposed CL-ODGAN framework, featuring ODConv, mixed normalization, and skip connections for adaptive feature extraction. Generator architecture of the proposed CL-ODGAN framework, featuring ODConv, mixed normalization, and skip connections for adaptive feature extraction. Fig. 3Qualitative evaluation of different dehazing methods on real-world hazy remote sensing images. The proposed CL-ODGAN (Ours) produces clearer images with better color fidelity and detail recovery. Qualitative evaluation of different dehazing methods on real-world hazy remote sensing images. The proposed CL-ODGAN (Ours) produces clearer images with better color fidelity and detail recovery. Fig. 4Visual comparison of ablation results on real remote sensing images. Visual comparison of ablation results on real remote sensing images. While quantitative and visual evaluations substantiate the advantages of CL‑ODGAN, the method remains challenged under complex atmospheric scenarios. As illustrated in Fig. 3, in regions where dense and light haze interlace, the model suffers from localized color distortion and luminance discontinuity—indicative of inconsistent transmittance estimation. This limitation stems from the training data’s inability to encompass the joint distribution of extreme haze concentrations and complex surface reflectances, which impedes the model from disentangling physically consistent radiance and transmission fields. Consequently, despite its overall metric superiority, CL‑ODGAN still requires improvements in physical interpretability and local robustness. To address these issues, future research could incorporate lightweight physics‑informed modules or weakly‑supervised atmospheric scattering priors, thereby enhancing generalization to complex real‑world conditions without sacrificing the flexibility of unpaired learning. To validate the rationale and synergistic effect of our proposed framework, we conducted a systematic ablation study that progressively builds from the baseline towards our full model. We evaluate four configurations: the baseline CycleGAN, our enhanced generator (with ODConv and hybrid normalization), and the full CL-ODGAN (enhanced generator + dual-level contrastive learning), against clear ground-truth using PSNR and SSIM. The results (Table 3) clearly demonstrate the effectiveness of our hierarchical design. The baseline CycleGAN achieves a PSNR of 15.76 and SSIM of 0.8432. Replacing its residual blocks with our custom ODBottleneck units—integrating ODConv for dynamic perception and hybrid normalization for stability—yields a noticeable improvement (PSNR: 15.92, SSIM: 0.8455). This confirms that our architectural modifications provide a stronger foundation. Crucially, incorporating the dual-level contrastive learning mechanism atop this enhanced generator (i.e., the full CL-ODGAN) achieves a substantial performance leap (PSNR: 28.15, SSIM: 0.8525). This dramatic gain underscores the synergistic relationship: the enhanced generator provides the capacity for high-fidelity restoration, while contrastive learning provides the essential guidance to steer this capacity without paired data. It is this combination that breaks the dependency on paired supervision (Table 3). Table 3Quantitative results of ablation experiments on the proposed CL-ODGAN components.MethodPSNRSSIMCyclegan15.760.8432Cyclegan + OD conv15.920.8455CL-ODGAN28.150.8525 Table 3Quantitative results of ablation experiments on the proposed CL-ODGAN components.MethodPSNRSSIMCyclegan15.760.8432Cyclegan + OD conv15.920.8455CL-ODGAN28.150.8525 Quantitative results of ablation experiments on the proposed CL-ODGAN components. Conclusion This study aims to address the core challenge of scarce paired real training data in remote sensing image dehazing tasks by innovatively proposing an unsupervised dehazing network framework that integrates contrastive learning with Omni-Dimensional Dynamic Convolution (ODCONV). The key contributions of this framework are twofold: First, we introduced an ODCONV module to replace traditional static convolution. This module dynamically adjusts convolutional kernel weights across spatial, channel, input, and output dimensions based on input features, enabling precise extraction and adaptive fusion of multi-scale contextual features. This significantly enhances the model’s generalization capability under complex haze distributions and achieves breakthrough progress in edge preservation and detail recovery. Second, we designed an efficient contrastive learning constraint mechanism. By constructing positive and negative sample pairs, it pulls the features of clear images and dehazed results closer while pushing them away from haze features in the feature space, thereby enhancing the model’s robustness in extracting haze-invariant features. In summary, the proposed method not only effectively circumvents the reliance on paired data but also, through the synergistic optimization of dynamic convolution and contrastive learning, achieves advanced dehazing performance in both subjective visual quality and objective evaluation metrics, offering a novel solution for unsupervised remote sensing image restoration. Although the proposed CL-ODGAN achieves promising performance on most remote sensing dehazing scenarios, it still has certain limitations. For example, the model may struggle with extremely low-light conditions or scenes with heavy, non-uniform haze, and it requires further optimization for edge cases. In future work, we plan to improve the model’s generalization to more diverse haze distributions and explore lightweight variants for real-time applications. All figures were generated by the authors using Python 3.x and draw.io.