We present a hyperspectral image (HSI) classification method based on spatial reconstruction to alleviate the influences of view changes to HSI encoding, and it mainly contains a spatial reconstruction mechanism, a feature representation network (FRN), and an auxiliary branch. The spatial reconstruction mechanism based on spectral metric unifies image patches with the same entities and different neighbor distributions to an identical cube, while the FRN based on soft band selection adaptively emphasizes informative spectral bands and suppresses redundant ones in the coding phase, and these two modules can form spatial distribution-insensitive data space and noise-robust discriminative feature vector and further improve the classification performance. Besides, the auxiliary branch based on the decoupling strategy ensures the latent relationships among neighbor pixels of the original patch, and it also highlights the relative importance of the center pixel. In addition, the experimental results on three public datasets demonstrate the superiority of the proposed method.

We present a spatial-spectral decoupling framework (SDF) to improve the performance of hyperspectral image classification, it mainly contains three modules, including data preprocessing, feature representation, and collaborative decision making. Specifically, the data preprocessing module based on band selection (BS) network can effectively emphasize useful spectral bands while suppressing redundant ones. Besides, the feature representation module is based on spatial-spectral decoupling (SD) network to avoid information confusion between the spatial and the spectral domains. In addition, the collaborative decision making mechanism based on joint optimization can maintain the discriminative properties of different branches and enhance mutual facilitation among them. Finally, the experimental results validate the effectiveness and superiority of our SDF.

This work presents a multidimensional attention learning-based lightweight convolutional neural network for very-high-resolution (VHR) remote sensing imagery recognition, which incorporates channel attention, spatial attention, and saliency sampler attention into the backbone to improve its recognition performance. Specifically, channel attention can alleviate the interference of original feature cube by adaptively giving different coefficient weights to different feature channels. Spatial attention can emphasize the discriminative regions by using the sum of activation value in different locations of the image to weight the original feature cube. Saliency sampler attention can increase the influence of interesting regions on the final representation according to the saliency priors. In general, different from the existing methods, this work utilizes the constraints rather than the model scale to improve the recognition performances of the network. In addition, the novel flooding loss is used to optimize the network, which can improve the performance of the framework by alleviating the severe overfitting problem.

We present a spatial–spectral decoupling interaction network for multispectral imagery change detection, which can exploit the underlying information of the multispectral imagery adequately through simultaneously considering the discriminative attribute of each pixel and robust spatial structure of the corresponding patch. Specifically, a 1-D convolutional neural network (1D-CNN) is applied to the spectral vector of each pixel to extract its discriminative feature, while a 2D-CNN is applied to the patch centering on the corresponding pixel to explore the spatial structure information. In addition, an interaction mechanism is incorporated into the feature fusion module to enhance the spatial–spectral consistency.

We consider the detection problem of maritime radar targets in the training-sample-starved and non-Gaussian sea clutter environment. The performance of conventional detectors for radar targets is seriously degraded due to both the starvation of training samples for estimating the clutter covariance matrix and the non-Gaussianity of sea clutter. In this letter, we adopt the inverse Gaussian distribution and the inverse complex Wishart distribution to model the texture and speckle covariance matrix of sea clutter, respectively. Then an adaptive Bayesian detector is developed based on the two-step generalized likelihood ratio test and the maximum posterior estimates of clutter parameters. Finally, the experimental results on simulated and measured data demonstrate the performance superiority of the proposed detector over its competitors, especially when the training samples are starved.

Non-Gaussian sea clutter causes conventional detectors designed in Gaussian clutter to suffer detection performance degradation, and some nuisance parameters cause the uniformly most powerful test to be unavailable. To improve the detection performance of maritime radar targets, we investigate the design of adaptive detectors in correlated non-Gaussian sea clutter via using suboptimal tests. The non-Gaussian sea clutter is modelled as a product of lognormal-distributed texture and complex Gaussian speckle. Two adaptive radar target detectors are developed by using the suboptimal two-step Wald and Rao tests. Specifically, a non-adaptive detector is derived by the Wald or Rao test when the clutter texture and speckle covariance matrix are assumed to be known in the first step; then the clutter parameters known in the first step are estimated, and the true parameters of the detector obtained in the first step are replaced with the estimated values. Theoretical proof and experimental verification indicate that the two proposed detectors have the constant false alarm property with regard to the clutter speckle covariance matrix and the clutter average power. Numerical results on simulated and measured radar data show that the proposed Rao-based detector outperforms its competitors, and has the stronger robustness to the signal mismatch compared to the proposed Wald-based detector.

Crack detection is an indispensable premise of road maintenance, which can provide early warning information for many road damages and save repair costs to a large extent. Because of the security and convenience, many image processing technique (IPT) based crack detection methods have been proposed, but their performances often cannot meet the requirements of practical applications because of the complex texture structure and seriously imbalanced categories. To address the aforementioned problem, we present an external attention based TransUNet for crack detection. Specifically, we tackle the TransUNet as the backbone of our detection framework, which can propagate the detailed texture information from shallow layers to corresponding deep layers through skip connections. Besides, the Transformer Block equipped in the second last convolution layer of the encoding component can explicitly model the long-range dependency of different regions in an image, which improves the structural representation ability of the framework and hence alleviates the interference from shadow, noise, and other negative factors. In addition, the External Attention Block equipped in the last convolution layer of the encoding component can effectively exploit the dependency of crack regions among different images, and further enhance the robustness of the framework. Finally, combined with the Focal Loss, the proposed label expansion strategy can further alleviate the category imbalance problem through transforming semantic categories of non-crack pixels distributed in the neighbors of corresponding crack pixels.

Remote sensing image scene classification is a fundamental problem, which aims to label an image with a specific semantic category automatically. Recent progress on remote sensing image scene classification is substantial, benefitting mostly from the powerful feature extraction capability of convolutional neural networks (CNNs). Even though these CNN-based methods have achieved competitive performances, they only construct the representation of the image in location-sensitive space-domain. As a result, their representations are not robust to rotation-variant remote sensing images, which influence the classification accuracy. In this paper, we propose a novel feature representation method by introducing a frequency-domain branch to the traditional only-space-domain architecture. Our framework takes full advantages of discriminative features from space domain and location-robust features from the frequency domain, providing more advanced representations through an additional joint learning module, a property that is critically needed to perform remote sensing image scene classification. Additionally, our method produces satisfactory performances on four public and challenging remote sensing image scene data sets, Sydney, UC-Merced, WHU-RS19, and AID.

Remote sensing image scene classification is a fundamental problem, which aims to label an image with a specific semantic category automatically. Recently, deep learning methods have achieved competitive performance for remote sensing image scene classification, especially the methods based on a convolutional neural network (CNN). However, most of the existing CNN methods only use feature vectors of the last fully connected layer. They give more importance to global information and ignore local information of images. It is common that some images belong to different categories, although they own similar global features. The reason is that the category of an image may be highly related to local features, other than the global feature. To address this problem, a method based on rearranged local features is proposed in this paper. First, outputs of the last convolutional layer and the last fully connected layer are employed to depict the local and global information, respectively. After that, the remote sensing images are clustered to several collections using their global features. For each collection, local features of an image are rearranged according to their similarities with local features of the cluster center. In addition, a fusion strategy is proposed to combine global and local features for enhancing the image representation. The proposed method surpasses the state of the arts on four public and challenging data sets: UC-Merced, WHU-RS19, Sydney, and AID.

Recently, progress on semantic image segmentation is substantial, benefiting from the rapid development of Convolutional Neural Networks. Semantic image segmentation approaches proposed lately have been mostly based on Fully convolutional Networks (FCNs). However, these FCN-based methods use large receptive fields and too many pooling layers to depict the discriminative semantic information of the images. Specifically, on one hand, convolutional kernel with large receptive field smooth the detailed edges, since too much contexture information is used to depict the “center pixel.” However, the pooling layer increases the receptive field through zooming out the latest feature maps, which loses many detailed information of the image, especially in the deeper layers of the network. These operations often cause low spatial resolution inside deep layers, which leads to spatially fragmented prediction. To address this problem, we exploit the inherent multi-scale and pyramidal hierarchy of deep convolutional networks to extract the feature maps with different resolutions and take full advantages of these feature maps via a gradually stacked fusing way. Specifically, for two adjacent convolutional layers, we upsample the features from deeper layer with stride of 2 and then stack them on the features from shallower layer. Then, a convolutional layer with kernels of 1× 1 is followed to fuse these stacked features. The fused feature preserves the spatial structure information of the image; meanwhile, it owns strong discriminative capability for pixel classification. Additionally, to further preserve the spatial structure information and regional connectivity of the predicted category label map, we propose a novel loss term for the network. In detail, two graph model-based spatial affinity matrixes are proposed, which are used to depict the pixel-level relationships in the input image and predicted category label map respectively, and then their cosine distance is backward propagated to the network. The proposed architecture, called spatial structure preserving feature pyramid network, significantly improves the spatial resolution of the predicted category label map for semantic image segmentation. The proposed method achieves state-of-the-art results on three public and challenging datasets for semantic image segmentation.

Aerial scene classification, which is a fundamental problem for remote sensing imagery, can automatically label an aerial image with a specific semantic category. Although deep learning has achieved competitive performance for aerial scene classification, training the conventional neural networks with aerial datasets will easily stick in overtting and local minimum. Because the aerial datasets only contain a few hundreds or thousands images, meanwhile the conventional networks usually contain millions of parameters to be trained. To address the problem, a novel convolutional neural network named JM-Net is proposed in this paper, which has different size of convolution kernels in same layer and ignores the fully convolytion layer, so it has fewer parameters and can be trained well on aerial datasets. Additionally, Cluster-SVM, a strategy to improve the accuracy and speed up the classification is used in the specific task. Finally, our method suparssed the state-of-art result on the challenging AID dataset while cost shorter time and used smaller storage space.