Scene classification of remote sensing images plays a vital role in Earth observation applications. Among various challenges, occlusion is a prevalent and critical issue in practical applications, particularly when dealing with large-area occlusions caused by clouds, shadows, and man-made structures. Current methods, whether based on occlusion recovery or occlusion-robust feature extraction, generally show limited performance when processing extensive occluded regions due to ignoring the inconsistency in feature representation caused by multiscale occlusions. To address the occlusion challenge, this letter proposes a novel contrastive learning-based multiscale occlusion framework with three key components: 1) a pretext task module that distinguishes between small and large occlusions to enable occlusion-invariant feature learning; 2) a multibranch feature extraction network based on ResNet-50’s shared-weight convolutional layers for consistent feature extraction across occlusion levels; and 3) a joint loss function that adaptively balances contrastive feature learning and classification. Extensive experiments were evaluated on the DIOR-Occ and LEVIR-Occ benchmark datasets, demonstrating significant improvements in classification accuracy across different occlusion scenarios. Compared with existing approaches, the proposed framework achieves superior robustness and generalization capabilities, with notable advantages in the analysis of highly occluded data. Future research will explore the adaptation of this framework to detection and segmentation tasks.

Hyperspectral unmixing is a technique in hyperspectral image processing that decomposes the spectra of mixed pixels into pure spectral components (endmembers) and their corresponding contributions (abundances). When dealing with complex mixed-terrain scenes, such as urban areas, significant challenges arise due to the complexity of the environment. Urban areas feature intricate geometric structures in individual pixels, including diverse 2-D and 3-D structures and the composite use of various building materials, resulting in highly complex scenarios. To address these challenges, this work exploits urban auxiliary information in the framework of an adaptive multitask autoencoder (AE) unmixing model, utilizing graph associations. The framework enhances the information in hyperspectral images by utilizing urban auxiliary data. Specifically, it performs superpixel segmentation to subdivide complex urban environments into simpler units. Subsequently, different AE-based unmixing methods are applied to these segmented results. Graph associations are employed to identify similar blocks in the image, incorporating this additional information into the unmixing process. In the experiments conducted for this work, two hyperspectral unmixing datasets were prepared, along with their corresponding urban auxiliary data. The results demonstrate that the proposed method achieves robust performance, even in complex urban environments.

Presents corrections to the paper, (Corrections to “Adaptive Multitask Autoencoder-Based Hyperspectral Unmixing Exploiting Auxiliary Data via Graph Associations”).

Implicit neural representations are increasingly recognized as a transformative approach for arbitrary-scale image Super-Resolution (SR), demonstrating capabilities that transcend traditional upscaling methods. However, the high complexity of urban structures poses significant challenges for existing methods, which mainly operate in the spatial domain and often fail to recover fine-grained textures and high-frequency details. To overcome this limitation, in this work we consider the frequency domain as a complementary representation, enabling the extraction of high-frequency information that enhances fine-grained detail reconstruction. Specifically, this study proposes a new Dual-domain Implicit Super-Resolution (DISR) framework that simultaneously extracts and fuses features from both the spatial and frequency domains, enabling continuous representation learning for remote sensing images. The DISR architecture comprises three key components: (1) The Dual-domain Feature Extraction (DFE) module, which extracts and enhances complementary features encompassing both local details and global structural information; (2) The Dual-domain Cross Attention Fusion (DCAF) mechanism, which dynamically integrates frequency-domain cues with spatial features to improve texture fidelity and structural coherence; and (3) The Dual-path Adaptive Implicit Parser (DAIP), which establishes a nonlinear mapping from coordinate positions to pixel values through parallel global-local processing pathways. Experimental results demonstrate that our new DISR framework not only improves other methods in terms of overall reconstruction quality metrics, but also excels at restoring high-frequency textures and fine structural details. The code is available at https://github.com/Liuyx-max/DISR.

Hyperspectral unmixing (HU) is a fundamental task in hyperspectral image analysis, aiming to decompose mixed pixels into endmember spectra and their corresponding fractional abundances. With the development of deep learning (DL), numerous DL-based unmixing methods have been proposed. However, most of these methods fail to integrate geometric priors (e.g., minimum volume constraints), limiting their performance in non-pure pixel scenarios (i.e., when no pure pixels are present in the scene). Current minimum volume constraint strategies based on the distance between endmembers and the data centroid may yield biased estimations under non-uniform data distributions. To overcome these limitations, we propose a new blind unmixing network with minimum volume constraint, named MVCUN. By jointly constraining the simplex volume through a total variation (TV) term related with the endmembers and a volume upper bound term, MVCUN achieves improved unmixing performance in non-pure pixel scenarios and exhibits robustness against non-uniform data distributions. Furthermore, we design a multi-scale global context attention (MSGCA) module that fuses local and global channel information through multi-scale convolutions and fully connected layers, thereby improving the network’s ability to estimate abundances. Experimental results on one simulated dataset and three real datasets demonstrate that our MVCUN significantly outperforms existing approaches in terms of both endmember extraction and abundance estimation accuracy.

Urban hyperspectral image super-resolution can rapidly acquire high-quality data with rich spatial details and spectral fidelity through technical means for urban development. However, conventional super-resolution methods for multi-spectral and natural images struggle to meet the aforementioned data requirements simultaneously. this paper proposes an urban hyperspectral super-resolution method that combines subpixel mapping and interpolation. This method aims to ensure spectral information and minimize time consumption through a high-precision surface interpolation method based on curve theory. Additionally, by utilizing subpixel mapping to introduce urban unmixing information. Finally, by employing wavelet transformation, the method integrates the effective information from both approaches, obtaining urban hyperspectral images with enhanced spatial detail and spectral fidelity. This method has been subjected to comprehensive experimentation, affirming that our proposed method surpasses the current state-of-the-art super-resolution in terms of performance and effectiveness.

Hyperspectral sensors can rapidly acquire high-quality spectral data, very useful for urban monitoring applications. Unfortunately, their spatial detail is not fine enough, and methods to enhance this resolution are required. However, conventional super-resolution (SR) methods for multispectral data do not match the requirements needed to maintain high spectral fidelity. Therefore, this article proposes a hyperspectral SR method that combines subpixel mapping and interpolation, and whose main aim is to enhance urban monitoring. This method aims to guarantee spectral quality and minimize computational time through a high-precision surface interpolation method based on curve theory. Moreover, unmixing-based subpixel mapping is exploited to introduce unmixing information. Finally, using wavelet transforms, the method integrates the effective information from the two previous approaches, obtaining urban hyperspectral images with enhanced spatial details and spectral fidelity. This method has been subjected to a comprehensive experimentation, affirming that the proposed method surpasses the current state-of-the-art SR results in terms of performance and effectiveness.

Hyperspectral unmixing is a crucial task in hyperspectral image processing and analysis. It aims to decompose mixed pixels into pure spectral signatures and their associated abundances. However, most current unmixing methods ignore the reality that the same pixel of a hyperspectral image has many different reflections simultaneously. To address this issue, we propose a multi-task autoencoding model for multiple reflections, which can improve the algorithm’s robustness in complex environments. Our proposed framework uses 3D-CNN-based networks to jointly learn spectral-spatial priors and adapt to different pixels by complementing the advantages of other unmixing methods. The proposed method can quantitatively evaluate each area of data, which helps improve the algorithm’s interpretability. This paper presents MAHUM (Multi-tasks Autoencoder Hyperspectral Unmixing Model), which stacks multiple models to deal with various reflections of complex terrain. We also perform sensitivity analysis on some parameters and show experimental results demonstrating our method’s ability to express the adaptability of different materials in different methods quantitatively.

Due to the trade-off of temporal resolution and spatial resolution, spatiotemporal image-fusion uses existing high-spatial-low-temporal (HSLT) and high-temporal-low-spatial (HTLS) images as prior knowledge to reconstruct high-temporal-high-spatial (HTHS) images. However, some existing spatiotemporal image-fusion algorithms ignore the issue that the spatial information of HTLS images is insufficient to support the acquisition of spatial information, which leads to the unsatisfactory accuracy of the fusion result. To introduce more spatial information, the algorithm in this article uses Cycle-generative adversarial networks (GANs) to simulate the change process of two HSLT images at k-1 and k+1, and to generate some simulated images between k-1 and k+1. Then, the generated images are selected under the help of HTLS images, and the selected ones are then enhanced with wavelet transform. Finally, the image with spatial information is introduced into the Flexible Spatiotemporal DAta Fusion (FSDAF) framework to improve the performance of spatiotemporal image-fusion. Extensive experiments on two real data sets demonstrate that our proposed method outperforms current state-of-the-art spatiotemporal image-fusion methods.

Spatiotemporal fusion utilizes the complementarity of high-temporal-low-spatial (HTLS) and high-spatial-low-temporal (HSLT) resolution data to obtain high temporal and spatial (HTHS) resolution fusion data, which can effectively satisfy the demand for HTHS data. However, due to the difference of spatial resolution, it is difficult to obtain precise spatial information in spatiotemporal fusion. To solve this problem, a multi-level strategy-based spatial domain prediction algorithm is proposed to enhance the spatial information extraction in spatiotemporal remote sensing imagery fusion, which can reduce the noise superposition in the process of multiple reconstruction. By learning-based first and then interpolation-based Super resolution reconstruction, the proposed method can obtain better prediction of spatial information and improve the accuracy of spatiotemporal fusion.

High-resolution remote sensing (HRRS) image scene classification takes an important role in many applications and has attracted much attention. Recently, notable efforts have been made to present massive methods for HRRS scene classification, wherein deep-learning-based methods demonstrate remarkable performance compared with state-of-the-art methods. However, HRRS images contain complex contextual relationships and large differences of object scale, which are significantly different from natural images. The existing deep-learning-based scene classification methods are originally designed for natural image processing and have not been optimized to adapt to the characteristics of HRRS images, which significantly affects the efficiency of the feature extraction and recognition accuracy. In addition, when designing a model for remote sensing tasks, the pretraining of the model is time-consuming. The enormous amount of pretraining time and computation resources necessarily increase the difficulty of producing an excellent model. In this letter, focusing on the problems above, we proposed a new convolutional neural network (CNN)-based scene classification method. The CNN-based scene classification method is constructed by spatial-scale-aware blocks and is efficient in extracting the abundant spatial features, but can also adaptively adjust feature responses to maximize the function of informative features in the classification results. In addition, an HRRS imagery-based learning strategy is utilized to obtain an initial model for fine-tuning the model parameters, which drastically reduces the pretraining time. The proposed method has been demonstrated using two HRRS data sets, and experimental results have proven the superiority of the proposed method.

High-resolution remote sensing (HRRS) image scene classification has been a critical task and greatly important for many applications, wherein convolutional neural network (CNN)-based methods have achieved considerable improvements. However, the CNN-based methods have countered a severe problem that massive annotation samples are required to obtain ideal model for scene classification. There is no dataset with a comparative scale to ImageNet to meet the sample requirement and labelling samples is labor-intensive and time-consuming. To solve the problem of insufficient annotation samples, a new generative adversarial network (GAN)-based sample generation method for scene classification is implemented. The proposed method is able to generate HRRS images with specific label and improve scene classification performance for the CNN-based methods.