Graph convolutional networks (GCNs)-based methods for hyperspectral image (HSI) classification have received more attention due to its flexibility in information aggregation. However, most existing GCN-based methods in HSI community rely on capturing fixed K-hops neighbors for feature information aggregation, which ignores the inherent imbalance in class distributions and fails to achieve optimal feature smoothing through graph convolution operator. It is unreasonable to apply fixed K-hops strategy for feature smoothing in imbalanced classes, as class regions with rich contextual information and those with poor contextual information require to capture different hops neighbors to achieve the optimal feature smoothing. To address this issue, this article proposes a novel approach called class-imbalanced graph convolution smoothing (CIGCS) for HSI classification, which achieves adaptive feature smoothing for imbalanced class regions. Firstly, we construct a semantic block-diagonal graph structure that describes imbalanced semantic class regions by considering label connectivity and spectral Laplacian regularizer. Secondly, we develop the class-imbalanced graph convolution smoothing technique to adaptively aggregate neighbor information for imbalanced class regions based on the decreasing Euclidean distance of samples within each bock-diagonal structure from the perspective of over-smoothing. The choice of adaptive neighbors can be guaranteed by a theoretical upper bound. Finally, the obtained optimal smoothed features are fed into the logistic regression to achieve good classification results. The proposed CIGCS method is evaluated on three real HSI data sets to demonstrate its superiority compared to some popular GCN-based methods.

Deep learning-based medical image segmentation has made great progress over the past decades. Scholars have proposed many novel transformer-based segmentation networks to solve the problems of building long-range dependencies and global context connections in convolutional neural networks (CNNs). However, these methods usually replace the CNN-based blocks with improved transformer-based structures, which leads to the lack of local feature extraction ability, and these structures require a huge number of data for training. Moreover, those methods did not pay attention to edge information, which is essential in medical image segmentation. To address these problems, we proposed a new network structure, called P-TransUNet. This network structure combines the designed efficient P-Transformer and the fusion module, which extract distance-related long-range dependencies and local information respectively and produce the fused features. Besides, we introduced edge loss into training to focus the attention of the network on the edge of the lesion area to improve segmentation performance. Extensive experiments across four tasks of medical image segmentation demonstrated the effectiveness of P-TransUNet, and showed that our network outperforms other state-of-the-art methods.

Although various methods based on convolutional neural networks have improved the performance of biomedical image segmentation to meet the precision requirements of medical imaging segmentation task, medical image segmentation methods based on deep learning still need to solve the following problems: (1) Difficulty in extracting the discriminative feature of the lesion region in medical images during the encoding process due to variable sizes and shapes; (2) difficulty in fusing spatial and semantic information of the lesion region effectively during the decoding process due to redundant information and the semantic gap. In this paper, we used the attention-based Transformer during the encoder and decoder stages to improve feature discrimination at the level of spatial detail and semantic location by its multihead-based self-attention. In conclusion, we propose an architecture called EG-TransUNet, including three modules improved by a transformer: progressive enhancement module, channel spatial attention, and semantic guidance attention. The proposed EG-TransUNet architecture allowed us to capture object variabilities with improved results on different biomedical datasets. EG-TransUNet outperformed other methods on two popular colonoscopy datasets (Kvasir-SEG and CVC-ClinicDB) by achieving 93.44% and 95.26% on mDice. Extensive experiments and visualization results demonstrate that our method advances the performance on five medical segmentation datasets with better generalization ability.

Hyperspectral image classification methods based on the graph convolutional network (GCN) have received more attention because they can handle irregular regions by graph encoding techniques. However, GCN-based HSI classification methods are highly sensitive to the quality of the graph structure. Its performance degrades in the case of underdeveloped graphs because it cannot excavate the intrinsic adjacency relationships. Thus, it is necessary to improve the quality of graph structure in GCN-based methods. In this paper, a novel diversity-connected graph convolutional network (DCGCN) method is proposed to improve the quality of the graph structure for HSI classification, and its basic idea can be adopted by other GCN-based methods. First, the potential neighbors are excavated by performing topological extensions based on the given graph. The diversity of surrounding neighbors is maintained by adaptively smoothing operation via a global threshold value from Kullback-Leibler divergence to eliminate weak interclass connections caused by weakly spectral variability. Second, another key connectivity restriction is imposed on the diverse neighbors to further refine the ambiguous connections of hard samples aiming at removing strong interclass connections where the spectral information is heavily confounded. Finally, the DCGCN method is analyzed theoretically to demonstrate its low-pass filter property. The comprehensive experiments demonstrate the effectiveness of the proposed DCGCN method and the basic idea of the diversity-connected graph in terms of overall accuracy (OA), kappa coefficient (KC), average accuracy (AA) indexes.

In recent years, advances in deep learning have greatly promoted the development of hyperspectral image (HSI) compression algorithms. However, most existing compression approaches directly rely on rate-distortion optimization without other guidance during model learning. Therefore, this brings challenges to distinguishing similar features or objects that are widely available in HSIs, especially in remote sensing scenes, since quantification in lossy compression can cause informative attribute (e.g., category) collapse and loss problems at high compression ratios. In this paper, we propose a novel hyperspectral compression network via contrastive learning (HCCNet) to help generate discriminative representations and preserve informative attributes as much as possible. Specifically, we design a contrastive informative feature encoding (CIFE) to extract and organize discriminative attributes from the original HSIs by enlarging the discrimination over the learned latents in different channel indexes to relieve attribute collapses. In the case of attribute losses, we define a contrastive invariant feature recovery (CIFR) to discover the lost attributes via contrastive feature refinement. Experiments on five different HSI datasets illustrate that the proposed HCCNet can achieve impressive compression performance, such as improvement of the peak signal-to-noise ratio (PSNR) from 28.86 dB (at 0.2284 bpppb) to 30.30 dB (at 0.1960 bpppb) on the Chikusei dataset.

Compressing hyperspectral images (HSIs) into compact representations under the premise of ensuring high-quality reconstruction is an essential task in HSI processing. However, existing compression methods usually encode images by smoothing due to the low-frequency information occupying a prominent component in most images. Consequently, these methods fail to capture sufficient structural information, especially in low bit rates, often causing inferior reconstruction. To address this problem, we propose here an edge-guided hyperspectral compression network, called CENet, to realize high-quality reconstruction. To enhance the structural latent representation ability, the CENet model incorporates an edge extractor neural network into the compression architecture to guide compression optimization by the edge-guided loss. We propose an interactive dual attention module to selectively learn edge features, obtain the most effective edge structure, and avoid additional edge information redundancy at the same time. In the proposed CENet, the edge-guided loss and interactive dual attention module are combined to enhance the comprehensive structure of the latent representation. Concretely, interactive dual attention makes the edge extraction network focus only on moderate boundaries rather than on all edges, which enables savings on the bit rate cost and helps achieve a strong structural representation. As a result, the reconstruction quality is significantly improved. The extensive experiments on seven HSI datasets verify that our model can effectively raise the rate–distortion performance for HSIs of any type or resolution (e.g., yielding an average peak signal-to-noise ratio (PSNR) of 30.59 dB at 0.2382 bpppb, which exceeds the baseline for Chikusei by 10.99%).

Although graph convolutional network (GCN) has achieved remarkable success in hyperspectral image (HSI) classification, most existing GCN-based approaches have failed to realize a deep network structure due to the oversmoothing problem. This problem largely limits the expression ability and feature extraction ability of GCN and hampers GCN’s capacity to model long-range relationships between samples in hyperspectral (HS) scenes. Moreover, there is a lack of theoretical analysis in those works that constructed deep GCN for HSI classification to illustrate how they overcome the oversmoothing problem. Aside from this, the characteristics and complexity of HSI are often neglected when constructing deep GCN models in HSI classification. To address these problems, a novel deep graph network based on first-order smoothing is proposed for HSI classification. Specifically, a local and global topologically consistent graph is constructed to thoroughly explore the union between fine pixel information and semantic superpixel information. Subsequently, a novel propagation procedure is proposed to address the oversmoothing problem. We creatively build a residual connection to the first layer to emphasize the feature information aggregated from the first-order neighborhood, which adds node features that have not yet become indistinguishable into deep layer, and at the same time, it can be considered as a correction to the original pixels affected by spectral variation in the input graph. Finally, we demonstrate how first-order smoothing-based deep graph network (FSDGN) can slow down the convergence rate of the oversmoothing problem by analyzing the propagation of FSDGN from the standpoint of the Laplacian spectral domain. In addition, the results of experiments performed on three benchmark datasets demonstrate its superiority over other state-of-the-art methods.

In hyperspectral image (HSI) classification task, semisupervised graph convolutional network (GCN)-based methods have received increasing attention. However, two problems still need to be addressed. The first is that the initial graph structure in the GCN-based methods is not sufficiently flexible to encode the homogenous structure similarity of HSI pixels when facing the complex scenarios induced by the spatial variability. Another problem is that the input (graph structure) and output (output features) of the GCN-based methods are separated with a "single pass" procedure, which is a suboptimal problem for HSI classification because it does not flexibly optimize the graph construction with a feedback method via output features. In this article, a novel spatial-spectral unified adaptive probability GCN (SSAPGCN) method is proposed for HSI classification. First, considering the homogeneous structural similarity of the pairwise relationships of HSI pixels, this article combines the inherent spectral information and spatial coordinates to obtain the spatial-spectral adaptive probability graph (SSAPG) structure, which can capture the probabilistic connectivity between each pair of the homogeneous HSI pixels. Second, the SSAPG structure and GCN model are combined into a unified framework to a daptively learn both the graph structure and the output features simultaneously with feedback. Finally, the proposed SSAPGCN method with two layers is evaluated on four public HSI datasets to demonstrate its superiority over different classification methods in terms of two evaluation metrics, the overall accuracy (OA) and kappa coefficient (KC), especially with small training sample sizes.

Cellular traffic prediction will play a key role in the deployment of future smart cities. Although the current traffic prediction methods based on deep learning show better performance than traditional prediction methods, they still have the following problems: (1) In spatial domain, the correlations between cellular traffic features cannot be captured accurately in non-local (including “geographic adjacency” and long-distance) spatial areas. (2) In temporal domain, the correlation of different time-grained features is failed to consider. To address these problems, a deep learning method considering dynamic non-local spatial correlation, self-attention, and correlation of spatio-temporal feature fusion is proposed. In spatial domain, our method can accurately capture the spatial correlation and highlight the contribution of more relevant traffic in the non-local area by designing a NLG-NLAM model. In temporal domain, the correlations of time-periodic features with different granularities are considered to clarify the key roles of different periodic features and eliminate the influence of irrelevant cellular traffic features on the prediction by designing a calibration layer. Experimental results indicate that the proposed method shows better performance than other mainstream prediction methods on three real-world cellular traffic datasets.

Most of the person re-identification (re-ID) algorithms based on deep learning mainly learn the global feature representation of pedestrians, while ignoring the important role of fine-grained pedestrian attribute features on re-ID tasks. Pedestrian attributes are middle-level semantic features, which have invariance in different poses, camera views, and illumination conditions. Considering the robustness and promotion of pedestrian attributes for person re-ID task, we propose an Attribute-guided Global and Part-level identity Network (AGPNet), which consists of a global identity task, a part-level identity task, and a pedestrian attributes learning task. AGPNet takes advantage of perceived semantic information of pedestrian attributes and deploys them as guidance to attend to human body regions and learn robust feature representation in the feature representation construction stage. Extensive experiments on two large-scale person re-ID datasets (Market-1501 and DukeMTMC-reID) show the effectiveness of our method, which is competitive with the state-of-the-art algorithms.

Semantic segmentation of small-scale objects in very high resolution (VHR) remote sensing images plays an important role in some special tasks, such as change detection and mapping of land cover. However, due to small size, small-scale objects are more likely to be completely obscured by shadows than large-scale objects, which make it difficult for the traditional convolutional neural network (CNN) to distinguish small-scale objects from shadows. Furthermore, even if small-scale objects are distinguished, their boundaries are still difficult to refine. To solve the above problems, a novel context union edge network (CEN) for small-scale objects semantic segmentation is proposed by comprehensively considering both the contextual and edge information. In CEN, a plug-and-play context-based feature enhancement module (CFEM) is designed to enhance the ability of CNNs to distinguish small-scale objects. Then, an information exchange mechanism (IEM) is proposed based on the dual-stream (semantic and edge stream) network to refine the boundaries of small-scale objects. Finally, some experiments based on the ISPRS Vaihingen data set are conducted in terms of both overall accuracy (OA) and F1-score. The proposed CEN achieves 89.9% of F1-score for small-scale objects (cars) and 90.9% of OA, harvesting new state-of-the-art results.

Remarkable improvements have been seen in the semantic segmentation of remote-sensing images. As an effective structure to aggregate shallow information and deep information, encoder–decoder structure has been widely used in many state-of-the-art models, but it possesses two drawbacks that have not been fully addressed. On the one hand, encoder–decoder structure fuses the features obtained from shallow and deep layers directly; despite harvesting some detailed information, it also brings in noisy features owing to the poor discriminant ability of the shallow layers. On the other hand, existing encoder–decoder structure merely fuses the high-level information generated by the last layer of encoder once, which neglects its guidance ability to the feature aggregation process in the decoder. In this letter, we first propose an edge perception module (EPM) to eliminate the noisy features in the shallow information, as well as enhance features’ structural information. And then, we generate the most suitable guidance information adaptively for different stages in the decoder through high-level information module (HIM). Finally, we apply the guidance information to achieve feature aggregation in the feature aggregation module (FAM). Combined with EPM, HIM, and FAM, our proposed model achieves 89.5% overall accuracy (OA) on the challenging ISPRS Vaihingen test set, which is the new state-of-the-art in the semantic segmentation of remote-sensing images.

By reason of factors such as terrains, weather conditions, sensor imaging methods and cultural and economic development, there is a large shift between the remote sensing imagery collected from different geographic locations and different sensors, which makes the state-of-the-art semantic segmentation models trained on source domain (a image set gathered from specific geographic locations and sensors) difficult to generalize to target domain (another image set collected from other geographic locations and sensors). Currently, unsupervised domain adaptation using adversarial training whose purpose is to align the marginal distribution in the output space between source and target domain, is the most explored and practical approach to address this issue. However, this global alignment approach does not take into account diversities of different regions in a specific image nor the category-level distribution, which leads to the consequence that some regions and categories which are already well aligned between the source and target domain may be incorrectly remapped. Therefore, we propose a region and category adaptive domain discriminator, aiming to emphasize the differences in regions and categories during the process of alignment. Specifically, on the one hand, we propose an entropy-based regional attention module in domain discriminator to emphasize the importance of difficult-to-align regions. On the other hand, we propose a class-clear module to update only the distribution of existing categories in one iteration without affecting all categories. Finally, a lot of experiments are introduced to indicate that the proposed method can obtain better results when compared with other state-of-the-art unsupervised domain adaptation methods using adversarial training.

Content-weighted compression scheme for high-resolution remote-sensing (RS) images can be well modeled by Markov random field (MRF)-oriented attention. This article addresses high-resolution RS image compression by incorporating MRF into attention mechanism. To this end, we reformulate the attention mechanism with MRF-based probabilistic graph modeling implicitly and combine the target of image compression and parameter learning of MRF in a unified framework, namely high-order MRF-oriented attention (HMA) network. Specifically, HMA extends key-value query (KVQ) pairwise terms of the vanilla attention to high-order terms, by which the prior information could be expressed effectively to boost performance of high-resolution RS image compression. It is noted that several superiorities of HMA are listed. First, unlike the vanilla attention network that apt to yield coarse features, HMA is capable of output more pleasing decoding results. Second, HMA can accelerate the convergence in the training of the deep neural networks (DNNs), thus facilitating deploying it on resource-limited IOT devices. Third, HMA demonstrates its potential of processing semantic joint task. Moreover, We thoroughly evaluate our approach on standard data sets of varying resolutions, the proposed framework performs favorably against most image coding standards and DNN-based codecs on the ISPRS Vaihingen data set and the USC-SIPI data set especially at low bit rates.

Graph convolutional networks (GCNs) have been shown to be effective for hyperspectral image (HSI) classification due to their capacity to learn representations of spatial–spectral features. However, the existing GCN-based models heavily rely on predefined receptive fields to capture and aggregate neighbor information for each node, which limits the ability to adaptively selecting the most significant receptive field from graph data. To address the aforementioned problem, in this article, we propose a novel dynamic adaptive sampling GCN (DAS-GCN) algorithm that captures neighbor information through adaptive sampling to allow the receptive field to be dynamically obtained. The basic underlying idea is that the most meaningful receptive field for each target node can be adaptively discovered, and the edge adjacency weights can be adjusted simultaneously after each adaptive sampling operation. Thus, we enable the graph to be dynamically updated and refined. Specifically, the adaptive sampling operation consists of two complementary components; in the first step, the importance of different remote nodes in a large-scale neighborhood is learned, while in the second step, rich underlying spatial–spectral information is extracted from local neighbors and filtered. The proposed model has the ability to learn how to extensively exploit spectral–spatial correlations from both local and remote nodes. Moreover, the proposed DAS-GCN model has a superior ability to leverage node feature information to naturally generalize and efficiently generate node embeddings for unseen data. The experimental results with overall accuracy on four real HSI datasets, i.e., Indian Pines, Pavia university, Houston 2013, and Salinas are 95.63%, 96.40%, 94.70%, and 99.08%, respectively, which clearly demonstrate the advantages of the proposed method compared with other state-of-the-art approaches.

Owing to the rapid development of deep neural networks, prominent advances have been recently achieved in the semantic segmentation of remote sensing images. As the vital components of computer vision, semantic segmentation, and edge detection have strong correlation whether in the extracted features or task objective. Prior studies treated edge detection as a postprocessing operation to semantic segmentation, or they implicitly combined the two tasks. We consider that pixels around the edges are easy to be misdivided because of the prevalence of intraclass inconsistencies and interclass indistinctions, which reflect the discriminative ability of models to distinguish different classes. In this letter, we propose a multipath atrous module to first enrich the deep semantic information. Then, we combine the enhanced deep semantic information and dilated edge information generated by canny and morphological operations to obtain edge-region maps via edge-region detection module, which identifies pixels around the edges. Then, we relearn these error-prone pixels using a guidance module for the segmentation branch in a progressive guided manner. Combined with edge and segmentation branches, our progressive edge guidance network achieves an overall accuracy of 91.0% on the ISPRS Vaihingen test set, which is the new state-of-the-art result.

Generally, the apparent spectral information of hyperspectral images (HSIs) is directly used to measure the similarity among HSI pixels in the feature space, but this process cannot reveal the inherent characteristics of HSI pixels. Moreover, constructing spatial-spectral block-diagonal subspace structure representations of intraclass land-cover samples remains a challenge for low-rank representation (LRR) in HSI classification. In this article, we propose two methods to reveal the complex intrinsic spatial-spectral features of HSIs using block-diagonal subspace structures, namely, the spatial-spectral block-diagonal structure representation with class probability (SSBDCP) and spatial-spectral block-diagonal structure representation with fuzzy class probability (SSBDFCP) methods, for HSI classification. First, the SSBDFCP and SSBDCP methods explore the structure similarity characteristics of the latent subspace to form a block-diagonal LRR (BDLRR) of intraclass pixels with class probability and fuzzy class probability (FCP) and suppress the interclass pixels' representations. Then, the spatial information is considered in the proposed methods to enhance spatial-spectral graph expression and capture more comprehensive information. Note that SSBDFCP can perform better than SSBDCP because the FCP considers the “typicalness” that a sample belongs to a specific category and utilizes complex intrinsic discriminative information based on the feedback of “weakly” supervised information. Moreover, the feedback information can remove the noise features around the pixels and take advantage of the benefits of true neighbors. The experimental results for the Indian Pines and Pavia University data sets show that the SSBDFCP and SSBDCP methods achieve better HSI classification results than other popular graph construction methods.

Caching can prepare data for computational tasks in advance by tracking the requirements and behaviors of distributed geographical information systems to reduce network latency and improve computational performance. This paper presents an enhanced method to actively cache data for data-intensive computations that considers both data relationships and the timeliness of those relationships. First, the access correlations, the correlation steps and the times of the correlations are computed based on the behaviors of the computational tasks. Because the influence of historically accessed records will decrease gradually over time, only recently accessed records are used. To track changes in the relationships and prevent cache waste problems, each record is given a different age-based weight. A conditional caching probability can then be computed based on the timeliness relationships, which can be used to find the appropriate data to compute simultaneously. Finally, we present several experiments that compare the proposed method with techniques that use other data placement strategies, active caching strategies and passive caching algorithms. The results show that the proposed model has better performance than other algorithms in all respects. In addition, the proposed model results in a lower cache replacement ratio. The experiments with different data sets on different data scales indicate that the proposed algorithm can also be used in large-scale distributed environments.