Unsupervised remote sensing dehazing remains a challenging and ill-posed task due to the absence of reliable supervision signals. Existing dehazing methods with unpaired data often oversimplify haze removal as style transfer, limiting generalization in complex scenarios. Moreover, current unimodal frameworks neglect cross-modal cues that could improve contextual reasoning. To address these issues, we propose a novel cross-modal guided self-supervised dehazing framework called CLIP-HNet, which achieves multi-model feature extraction, boundary-focused reconstruction and adaptive sample filtering. Specifically, to capture global-local contextual features, a hybrid feature interaction network is designed, which bridges the feature representations of multi models with global context-aware module (GCAM) and hybrid feature fusion module (HF2 M). Then, based on the hybrid features, a boundary-aware feature reconstruction (BFRec) is proposed to further refine edge details. Furthermore, a CLIP-guided progressive information distillation scheme is presented to dynamically prioritize training samples and distill useful signals, which predicts haze concentration by CLIP and progressively increases sample difficulty during the training stage. Finally, a frequency-domain texture matching (FTM) strategy refines texture and spectral details, enhancing the model's ability to recover fine details. Experiments on synthetic and real RSIs demonstrate that the proposed CLIP-HNet surpasses state-of-the-art approaches, achieving superior visual quality and quantitative performance.

Remote sensing images (RSIs) are often exposed to various degradation factors such as sensor noise and poor observation environments. These factors can result in the loss of image details, spectral distortion, and blurred scenes, which will severely hinder the performance of subsequent RSI applications. Unfortunately, current methods are only capable of addressing individual degradation factors, and lack the ability to handle multiple degradation factors within a unified framework. To mitigate this problem, we model the complex degradation process as a double-degradation model for RSIs, and propose a unified framework based on saliency-guided interaction learning (SGIL) for double-degradation RSI restoration. The proposed SGIL can simultaneously alleviate the influence of external environment degradation and internal sensor noise degradation, which comprises three parts: pseudo pixel supervision-based saliency analysis (PPS-SA), a task-aware interaction learning (TAIL) model, and a global feature enhancement module (GFEM). In PPS-SA, an explicit PPS-SA method is designed to generate saliency maps to effectively distinguish different texture complexities of RSIs, and a saliency-guided mapping selection mechanism is introduced to adapt to complex external environmental interference factor. In the TAIL model, a dehazing module and a super resolution (SR) module are specialized in alleviating the external environment interference and internal sensor noise, respectively. Instead of simple cascading, the two modules interact and collaborate with each other, which drastically improves the performance of double-degradation image restoration. To further improve the performance of the proposed SGIL, we also propose the GFEM to exploit global features and refine the restored results. Experiments on several RSI datasets demonstrate that the proposed SGIL achieves promising results on complex double-degradation RSIs.

Remote sensing images (RSIs) in real scenes may be disturbed by multiple factors, such as optical blur, undersampling, and additional noise, resulting in complex and diverse degradation models. At present, mainstream super-resolution (SR) algorithms only consider a single and fixed degradation (such as bicubic interpolation) and cannot flexibly handle complex degradations in real scenes. Therefore, designing an SR model that can deal with various degradations has gradually attracted researchers’ attention. Some early studies estimate degradation kernels and then perform degradation-adaptive SR but face the problems of estimation error amplification and insufficient high-frequency details in the results. Although blind SR algorithms based on generative adversarial networks (GANs) have greatly improved visual quality, they still suffer from pseudo-texture, mode collapse, and poor training stability. This article proposes a novel blind SR framework based on the stochastic normalizing flow (BlindSRSNF) to address the above problems. BlindSRSNF learns the conditional probability distribution over the high-resolution image space given a low-resolution (LR) image by explicitly optimizing the variational bound on the likelihood. BlindSRSNF is easy to train and can generate photorealistic SR results that outperform GAN-based models. In addition, we introduce a degradation representation strategy based on contrastive learning to avoid the error amplification problem caused by explicit degradation estimation. Comprehensive experiments show that the proposed algorithm can obtain SR results with excellent visual perception quality on both simulated LR and real-world RSIs. The code is available at https://github.com/hanlinwu/BlindSRSNF.

In this paper, we propose a dynamic mutual enhancement network (DMENet) for haze removal in remote sensing images. It has three major advantages compared with other dehazing algorithms: 1) The proposed DMENet is based on the U-Net architecture to extract features effectively, which is composed of three components, i.e., a multi-scale encoder, a middle transmission layer (MTL), and a dynamic mutual decoder. 2) The dynamic mutual enhancement (DME) module is designed to dynamically integrate multi-level feature maps in a mutual way, which contains the low-level detail information and high-level semantic information respectively. 3) To improve the robustness and generalization performance of the DMENet, the hybrid supervision is built for network training between the restored results and their ground-truth labels, which consists of the pixel-level supervision, patch-level supervision and image-level supervision. Experimental results on both synthetic datasets and real remote sensing hazy images demonstrate that the proposed DMENet can gain significant progresses over the competing methods.

The demands for spectral and spatial quality in remote sensing (RS) images vary from region to region. Saliency detection is an effective tool to distinguish different regions with different demands. In this paper, we introduce saliency detection to satisfy these demands and propose a novel saliency-driven pan-sharpening network to further improve the fusion quality. Firstly, we combine foreground distribution with background prior to generate the initial saliency map, and implement least-square optimization to improve the detection accuracy. Then, we construct a dense spatial attention network trained through a new spatial-spectral-based loss function designed by saliency to meet diverse spectral and spatial needs of different regions. Thus, accurate fused images can be predicted. Experiments on SPOT-5 dataset indicate that our proposal has excellent properties with respect to the unified spatial-spectral quality against state-of-the-art methods.

Haze in remote sensing images (RSIs) usually causes serious radiance distortion and image quality degeneration, resulting in difficult remote sensing inversion and interpretation. Under the condition of dense haze, existing dehazing methods still experience problems to be solved: 1) the texture details and spectral characteristics in RSIs cannot be restored well; 2) small-scale objects, such as cars and ships, which often consist of only a few pixels in RSIs, cannot be effectively highlighted in dehazed results. To solve these issues, we propose a novel dynamic collaborative inference learning (DCIL) framework that can significantly restore real surface information from dense hazy RSIs. First, we design a dynamic mutual enhancement (DME) mechanism to reinforce the low-level texture features by integrating primary information and semantic information at different levels. Second, we propose a spectrum-aware aggregation (SAA) strategy to mine the spectrum features among multiscale restored results, which can fully capture spectral characteristics. Third, we build a collaborative criterion by constructing a Siamese network structure in the training stage to improve the robustness and generalization performance of DCIL considering the diversity of the scale range and view change of RSIs. Finally, we propose a phased learning strategy to deduce the implicit haze-relevant features by gradually increasing the concentration of haze which can effectively address small-scale objects obscured by dense haze. To this end, we develop two synthetic remote sensing dehazing datasets to train our model, which can also alleviate the dilemma of hazy RSI datasets shortages. Experimental results on both synthetic datasets and real remote sensing hazy images demonstrate that the proposed DCIL can attain significant progress compared to competing methods. The two synthetic hazy datasets are available at https://github.com/Shan-rs/DCI-Net.

The ambient light and transmission of light source regions and non-light source regions have different physical properties in nighttime scenes. However, existing nighttime dehazing algorithms ignored this phenomenon and got oversaturated results looked like in the daytime. In this paper, we propose a nighttime dehazing method using fused ambient light and transmission based on saliency detection. We first propose a multi-scale saliency detection algorithm to differentiate light source regions and non-light source regions. Then we estimate ambient light and transmission of the two regions based on different methods to acquire the fused ambient light and fused transmission. Finally, we get the haze-free results based on the nighttime imaging model. Experimental results show that our method outperforms other state-of-the-art methods in color recovery and our haze-free results still look like in the nighttime.

To efficiently remove haze in unmanned aerial vehicle (UAV) remote sensing images, a novel attention-based feedback dehazing network (AFDN) is proposed, which is constructed by feedback connections and attention-based feedback blocks (AFBs). It has three major advantages compared with other dehazing algorithms: 1) The feedback connections, which allow network to use previous state to improve current performance, can effectively help the proposed AFDN generate clear remote sensing scenes progressively. 2) The AFBs are specially designed to extract global residual features, in which the dual attention block can usefully reduce redundant information and improve the fitting ability of network. 3) To obtain abundant texture information from UAV remote sensing images and restore real ground surfaces, an energy loss is employed for texture features learning. Experiments on synthetic datasets and real UAV remote sensing images verify the superiority of AFDN over several state-of-the-art methods in terms of qualitative and quantitative analysis.

Abstract Haze is a common atmospheric phenomenon that causes poor visibility in outdoor images, which greatly limits image application in later stages. Therefore, haze removal has become the first and most indispensable step when dealing with degraded images. In this paper, we propose a novel bright channel prior (BCP) model and a saliency analysis strategy for haze removal. First, we obtain a more robust and accurate atmospheric light by a superpixel‐based dark channel method. Second, we utilize the dark channel prior (DCP) to handle dark regions in hazy images. However, the DCP often mistakes white regions for opaque haze and thus causes serious colour distortion and halo effects. To solve this problem, a new BCP is proposed to accurately estimate the transmission of bright regions in hazy images. Third, we fuse the DCP and BCP using a multiscale fusion strategy with Laplacian pyramid representation to gain the correct transmission information for both bright and dark regions. Finally, a novel saliency analysis strategy for transmission refinement is proposed, so that the texture details can remain present to the greatest extent in the restored images. The experimental results illustrate that our proposed method performs well in restoring images containing bright objects.

Transmission similarity is an important feature which can greatly increase the capability of convolutional neural network (CNN) to fit transmission map. However, it is not sufficiently utilized in existing algorithms. In this paper, we propose a novel light-weight similarity driven transmission computing network called SDTCN that is guided by the attributes of transmission similarity. First, we adopt a non-data-driven image segmentation method to acquire the transmission similarity. Compared with CNN based segmentation approaches, our method can not only greatly save computing resources, but also separate the objects and background precisely. Second, a full convolutional network is introduced to reduce blocky effects in SDTCN. Finally, unlike previous "first airlight then transmission" mode, a dependable airlight estimation approach is designed drawing on the transmission map generated by SDTCN, which can improve the accuracy of airlight effectively. Extensive experiments demonstrate that the proposed algorithm outperforms the state-of-the-art methods on synthetic and real-world images.

Saliency detection is a newly emerging tool to extract target in image processing. However, due to the loss of color in synthetic aperture radar (SAR) images, the detection result using the traditional saliency analysis is not satisfying. Therefore, a new saliency-driven target detection model based on common visual feature clustering is introduced for multiple SAR images. Firstly, Markov Random Field is applied to extract intra-image saliency map. Secondly, intensity, texture and curve features are extracted from multiple SAR images as common visual features, which can effectively compensate for the lack of color information. And then fuzzy c-means is employed to construct inter-image saliency map. Finally, an effective fusion strategy is used to combine the intra-image saliency map with the inter-image saliency map to obtain the final common saliency map. The experimental results demonstrate that the proposed model outperforms most the state-of-the-art saliency detection models.

Saliency analysis is a hot topic in target detection for synthetic aperture radar (SAR) image. In this paper, we come up with a novel target detection model based on statistical saliency analysis and saliency-oriented geodesic active contour model for SAR imagery. Firstly, the contrast and homogeneity features are extracted from the gray-level co-occurrence matrix to make up the texture saliency map. Then we obtain the prior graph and likelihood map using superpixels segmentation and Otsu, respectively, which comprise the Bayesian saliency map. The two saliency maps subsequently merge into the final statistical saliency map. Finally, we present a saliency-oriented geodesic active contour model, in which a saliency orientation map is embedded into the level set-based energy functional. Based on the final saliency map, the saliency-oriented geodesic active contour model can acquire the specified and accurate target contour. In qualitative and quantitative experiments, the comprehensive performance of the proposed method outperforms the competing models in maintaining complete targets and accurate boundaries.