Synthetic aperture radar (SAR) images obtained from multi-sensor systems usually exhibit significant shift in data distribution, known as the domain shift. It is challenging to utilize the relevant knowledge of multi-sensor datasets for SAR image cross-domain classification with general supervised learning methods. While domain adaptation (DA) methods can alleviate the domain shift by aligning distributions, they are limited in handling the scale/resolution variations observed in multi-sensor SAR images. These methods mainly focus on feature representations at a single scale for distribution alignment, which may fail to fully align the distributions across different scales. To solve this problem, we propose a new multi-scale alignment domain adaptation network (MSADAN) for SAR ship cross-domain classification. MSADAN explicitly considers the scale factor, enabling us to overcome the weak generalization observed in existing DA methods when dealing with SAR images exhibiting significant scale/resolution variations. Specifically, we develop a scale-aware feature extractor to effectively capture the multi-scale information present in the datasets, facilitating comprehensive representation learning. Furthermore, we propose a multi-level bilinear fusion (MLBF)-based adversarial learning strategy to overcome the limitations of single-scale feature extraction for distribution alignment, aiming to enhance the generalization ability of the model across domains. In addition, a customized class contrastive loss is designed to improve the inter-class separability and intra-class compactness by penalizing cross-class confusion. Experimental results on datasets demonstrate the superiority of MSADAN in SAR ship cross-domain classification.Note to Practitioners—Ship classification using Synthetic aperture radar (SAR) images proves to be a promising strategy in modern maritime monitoring systems. The primary motivation of this paper is to develop a cross-domain classification system tailored for SAR ship target. In real-world scenarios, SAR data often presents significant resolution variations due to the different imaging modes and conditions across multiple sources. Existing DA methods fail to effectively address this unique challenge in the remote sensing field, resulting in poor cross-domain classification performance. The proposed method considers the scale or resolution variations of the targets during feature learning. Furthermore, explicitly incorporating resolution into the distribution alignment enhances the generalization ability of the model to target tasks when dealing with the significant scale or resolution variations across datasets. Experimental results confirm the practicality and robustness of our SAR ship cross-domain classification method in real scenarios. This is of significant importance in promoting the application of machine learning approaches in practical scenarios. In the future, we plan to integrate the physical scattering information of SAR images for the interpretable cross-domain classification for SAR targets.

Partially occluded target recognition is a pressing issue in synthetic aperture radar (SAR) target recognition. Occlusion causes the loss of crucial information, like target structure details. This paper proposes a new structural relation extraction framework to address partial occlusion. It is achieved through the tailored design of counterfactual samples synthesizing and jigsaw mutual learning (CSS-JML). The ASC model parameters have clear physical meanings, aiding in understanding local structural changes. By integrating SAR and ASC images, effective structural relationship representations are extracted, mitigating occlusion effects. The CSS module is designed to generate occluded counterfactual SAR and ASC images using pairs of target data. There is no longer a requirement for further annotation information because this new generation process is limited by recognition tasks. The JML module employs mutual learning to complete jigsaw puzzle tasks in both modalities. And in this process, we design two types of similarity constraints to facilitate the extraction of unified structural information across different modalities. The FA module interacts with recognition features, facilitating the classification and identification of partially obscured targets. Experimental results on MSTAR-based and FUSARShip-based test datasets with three occlusion patterns demonstrate the method’s superiority in most occluded conditions, confirming its effectiveness in SAR occluded target recognition. Note to Practitioners—The motivation for this paper stems from the need to design a robust and efficient SAR target recognition method under partial occlusion. The structural relationships within samples can provide valuable information for recognizing occluded targets. We designed a counterfactual generation module to generate occluded target samples using paired targets in an unsupervised manner. This module is jointly optimized with subsequent recognition tasks, without the need for additional information. Recognizing the interaction bottleneck in mutual learning tasks, we designed two similarity constraints to promote the extraction of unified representations. Finally, we designed a feature affinity module to introduce structural representation information into the recognition branch. The proposed method achieves excellent classification performance in various occlusion environments.

Multi-expert networks have shown great superiority for imbalanced data classification tasks due to their complementary and diverse. We have summarized two aspects for further explorations:(1)uncontrollable results, arising from the performance differences of individual experts and variations in sample difficulty;(2)insufficient exploration of the internal data structure. These factors result in inconsistent model performance across different data distributions, thereby impact the model’s generalization ability. To address the above issues, we propose a Collaborative Global-Local Structure Network (CGL-Net) with knowledge distillation for imbalanced data classification. Firstly, CGL-Net, as a new framework, decouples the representation learning of imbalanced data into global and local structure, enhancing the controllability of integration model in a hierarchical manner. Secondly, CGL-Net innovatively combines knowledge distillation, data augmentation, and multiple expert networks, efficiently extracting the internal structure of the data and improving robust recognition on imbalanced data. In particular, the global structure learning introduces an independent student network that integrates knowledge from diverse experts, enabling the model to achieve comprehensive and balanced performance across categories in imbalanced data. The local structure learning incorporates augmented data, allowing the model to focus on discriminative regional learning of individual objects, thereby enhances the robust representation for imbalanced data. After completing these two sequential learning stages, the model hierarchically integrates knowledge to achieve robust recognition performance on imbalanced data. Extensive experiments on six benchmark datasets demonstrate that the proposed CGL-Net significantly outperforms recent state-of-the-art methods.

Fuzzy clustering is still a hot topic because it can calculate the support degrees of an object belonging to different clusters to characterize uncertainty. However, it remains a challenge to detect clusters of arbitrary shapes, sizes, and dimensionality. What's worse, some objects are indistinguishable (imprecise) when they are in the overlapping regions of different clusters. To address such issues, this paper investigates a belief-based fuzzy and imprecise clustering (BFI) method, which can detect arbitrary clusters and provide the behavior (support) of objects to these clusters. Moreover, BFI can assign each imprecise object to a meta-cluster, defined as the union of specific clusters, to characterize (partial) imprecision. The proposed BFI can significantly reduce the risk of misclassification, and the effectiveness is validated in image processing (e.g., image segmentation and classification) and several benchmark datasets by comparing it with some typical methods.

The distribution of SAR targets generally conforms to a long-tailed distribution. Due to the existence of sample distribution bias and sample selection bias, training classifiers on this distribution of data often introduces spurious correlations between samples and classes. To address this issue, we propose a two-stage causal intervention framework. The core is that structural causality allows for independent interventions on multiple biases, thereby ensuring high-quality tail class predictions while maintaining unbiased performance for head classes. Firstly, we construct a structural causal graph for the long-tailed recognition task from causal perspective. Based on this graph, the causal paths underlying the two types of biases are identified. Secondly, we design a data augmentation method named DiagPatch-M, which identifies causal features within samples. In this process, these generated patches randomly integrate causal and non-causal features from two different samples, disrupting the original recognition process and effectively eliminating biases induced by sample selection. Thirdly, we design an unbiased structural risk minimization (USRM) optimization strategy, which eliminates the “head preference” of conventional models and the “tail preference” of modified models. This strategy reduces the bias introduced by the model’s dependence on the original sample distribution, and achieves stable recognition under different sample distributions. Experimental results on two long-tailed and two balanced datasets demonstrate that the effectiveness of our model surpasses the state-of-the-art (SOTA) methods, indicating the efficacy of our proposed framework in tackling the challenges posed by the long-tailed distribution in SAR target recognition.

The multisource universal domain adaptation (MSUDA) relaxes the constraints between the source and target domains, enabling the transfer of knowledge between domains without any restrictions on the number of source domains and the existence of unknown (private) categories. However, identifying the unknown samples in the target domain is extremely challenging since there are no available samples with the same label in source domains. Another immense challenge lies in extracting domain-invariant features for knowledge transfer since there are distribution discrepancies between each source and target domain. In this article, we propose the multirepresentation DA network (MRDAN) to classify the unlabeled targets by harnessing multiple source domains with nonidentical label sets. First, we propose a threshold-free conflict-based predictions with uncertainty (CPU) module, which comprehensively mines the complementary knowledge from different source domains to identify both known and unknown samples simultaneously. To accurately extract the domain-invariant features for recognizing known and unknown samples, a multilevel distribution alignment (MLDA) strategy is introduced to decrease the distribution discrepancy between multiple domains with nonidentical category spaces progressively. Finally, comprehensive experiments conducted on three commonly used datasets demonstrate the effectiveness of the proposed MRDAN in recognizing both known and unknown samples.

Quantification and reduction of uncertainty in deep-learning techniques have received much attention but ignored how to characterize the imprecision caused by such uncertainty. In some tasks, we prefer to obtain an imprecise result rather than being willing or unable to bear the cost of an error. For this purpose, we investigate the representation of imprecision in deep-learning (RIDL) techniques based on the theory of belief functions (TBF). First, the labels of some training images are reconstructed using the learning mechanism of neural networks to characterize the imprecision in the training set. In the process, a label assignment rule is proposed to reassign one or more labels to each training image. Once an image is assigned with multiple labels, it indicates that the image may be in an overlapping region of different categories from the feature perspective or the original label is wrong. Second, those images with multiple labels are rechecked. As a result, the imprecision (multiple labels) caused by the original labeling errors will be corrected, while the imprecision caused by insufficient knowledge is retained. Images with multiple labels are called imprecise ones, and they are considered to belong to meta-categories, the union of some specific categories. Third, the deep network model is retrained based on the reconstructed training set, and the test images are then classified. Finally, some test images that specific categories cannot distinguish will be assigned to meta-categories to characterize the imprecision in the results. Experiments based on some remarkable networks have shown that RIDL can improve accuracy (AC) and reasonably represent imprecision both in the training and testing sets.

Deep learning has been widely used in the field of radar image classification because of its powerful feature extraction capabilities. In the land-sea clutter classification of sky-wave over-the-horizon radar (OTHR), deep learning methods perform poorly due to the radar receiver noise and the ionosphere. Addressing this challenge, a dual attention aided residual neural networks (DAAResNet) is proposed for OTHR land-sea classification. Leveraging prior knowledge that landsea clutter features predominantly cluster around the 0 Hz frequency, two attention mechanisms are introduced. Firstly, a channel attention module (CAM) is proposed, which directs the network’s focus towards critical channels. Secondly, a frequency attention module (FAM) is proposed, which directs attention towards pivotal frequencies. The classification performance of DAAResNet is validated on the original dataset and the scarce dataset. Experimental results show that DAAResNet outperforms state-of-the-art methods.

Cross-scene classification aims to transfer knowledge acquired from a label-rich source domain to an unlabeled target domain with a distribution shift. With the large amount of available remote sensing data from diverse satellite platforms, it prompts us to utilize the knowledge from extensive datasets to address target tasks in small-scale domains, known as partial domain adaptation (PDA). However, the PDA setting poses significant challenges for remote sensing scene images. Existing methods often fail to sufficiently explore both task-specific and transferable knowledge across domains based on the representations of entire samples, potentially resulting in the amplification of negative transfer brought by irrelevant knowledge. To address this, we propose a new selective alignment transformer (SAT) designed to distinguish transferable and untransferable knowledge across domains for cross-scene classification in RSIs under the PDA scenario. Specifically, a new bi-level reweighting strategy that incorporates transferability-aware patch selection and class-wise reweighting is developed to emphasize the transferable image patches and classes. Based on the aforementioned reweighting strategy, we further introduce a patch-weighted maximum mean discrepancy (PMMD) loss, which selectively aligns the distributions from the patch-level perspective, facilitating the learning of transferable domain-invariant representations. The experimental results of SAT demonstrate its effectiveness and superiority in addressing this practical domain adaptation (DA) task, outperforming state-of-the-art methods in PDA tasks on four datasets.

Multisource unsupervised domain adaptation (MUDA) is an important and challenging topic for target classification with the assistance of labeled data in source domains. When we have several labeled source domains, it is difficult to map all source domains and target domain into a common feature space for classifying the targets well. In this article, a new progressive multisource domain adaptation network (PMSDAN) is proposed to further improve the classification performance. PMSDAN mainly consists of two steps for distribution alignment. First, the multiple source domains are integrated as one auxiliary domain to match the distribution with the target domain. By doing this, we can generally reduce the distribution discrepancy between each source and target domains, as well as the discrepancy between different source domains. It can efficiently explore useful knowledge from the integrated source domain. Second, to mine assistance knowledge from each source domain as much as possible, the distribution of the target domain is separately aligned with that of each source domain. A weighted fusion method is employed to combine the multiple classification results for making the final decision. In the optimization of domain adaption, weighted hybrid maximum mean discrepancy (WHMMD) is proposed, and it considers both the interclass and intraclass discrepancies. The effectiveness of the proposed PMSDAN is demonstrated in the experiments comparing with some state-of-the-art methods.

Classification of missing data based on estimation is still challenging since existing methods relying on one imputation strategy fail to consider the diversity of different attribute distributions. In this case, there are inevitably some “bad” estimations at the attribute level, reducing the performance of classification. This article proposes a mixed-type imputation method (MTI) to classify missing data under the theory of belief functions (TBF) via two quality matrices to address this problem. The proposed MTI method has the advantages of making estimations as close to the truth as possible at the attribute level while reducing the negative impact of possible bad estimations on the classification. Specifically, the first matrix used to impute missing values can characterize the different supports of multiple imputation methods for estimating various attributes. The other matrix used to perform the classification task can extract the reliabilities of estimations on the different classes. The validity has been demonstrated in the final decision support based on the TBF, famous for characterizing uncertainty and imprecision, for example, caused by missing values.

Imputing the incomplete patterns in clustering tasks is a common but risky procedure, because the estimated values may affect the real distribution of the data and deteriorate the results. To address this problem, a new belief-based incomplete pattern unsupervised classification method (BPC) with uncertainty and imprecision reasoning is proposed in this paper. First, the complete patterns are grouped into a few clusters to obtain the corresponding reliable centers, and thereby are divided into reliable patterns and unreliable ones by an optimization method. Second, a basic classifier trained by reliable patterns is employed to classify unreliable patterns and incomplete patterns edited by the neighbors. Finally, some imprecise patterns are carefully reassigned again by a new distance-based rule depending on the obtained reliable centers and belief functions theory. The simulation results show that the BPC has the potential to deal with real datasets.

Open set recognition (OSR) aims to correctly recognize the known classes and reject the unknown classes for increasing the reliability of the recognition system. The distance-based loss is often employed in deep neural networks-based OSR methods to constrain the latent representation of known classes. However, the optimization is usually conducted using the nondirectional euclidean distance in a single feature space without considering the potential impact of spatial distribution. To address this problem, we propose orientational distribution learning (ODL) with hierarchical spatial attention for OSR. In ODL, the spatial distribution of feature representation is optimized orientationally to increase the discriminability of decision boundaries for open set recognition. Then, a hierarchical spatial attention mechanism is proposed to assist ODL to capture the global distribution dependencies in the feature space based on spatial relationships. Moreover, a composite feature space is constructed to integrate the features from different layers and different mapping approaches, and it can well enrich the representation information. Finally, a decision-level fusion method is developed to combine the composite feature space and the naive feature space for producing a more comprehensive classification result. The effectiveness of ODL has been demonstrated on various benchmark datasets, and ODL achieves state-of-the-art performance.

Open set recognition (OSR) aims to classify known classes and recognize unknown classes simultaneously. Existing OSR methods have primarily focused on learning decision boundaries based on overall feature representations, and have achieved good performance on various coarse-grained image datasets. However, the overall feature representations of objects in fine-grained image datasets are highly similar, making it difficult to distinguish between known and unknown classes by overall feature-based decision boundaries. To address this problem, we propose a progressive learning vision transformer (PLViT) with a coarse-to-fine optimization strategy. In PLViT, the overall feature representations are first optimized in the distance space to learn the initial decision boundaries. Then, a context-aware patch selection module is designed to locate the discriminative part regions. Afterwards, the multi-layer representations of each selected patch are aggregated according to the self-attention weights, and input into the last transformer layer to extract local feature representations. Finally, overall and local feature representations are adaptively fused and optimized in the angular space to further refine the decision boundaries. Experimental results on four fine-grained remote sensing object recognition datasets show that PLViT outperforms state-of-the-art methods.

With the rapid development of supervised learning-based SAR target recognition technology, it is easy to find that the recognition performance is proportional to the amount of training samples. However, the biased data distribution and under-representation of the model caused by incomplete data within categories exacerbate the challenge of SAR interpretation. In this paper, we propose a new view-semantic transformer network (VSTNet) that generates synthesized samples to complete the statistical distribution of training data and improve the discriminative representation of the model. First, SAR images from different views are encoded into a disentangled latent space, which allows us to synthesize data with more diverse views by manipulating view-semantic features. Second, the synthesized data as a complement effectively expands the training set and alleviates the overfitting problem of limited data in sparse views. Third, the proposed method unifies SAR image synthesis and SAR target recognition into an end-to-end framework to boost their performance against each other. Experiments conducted on moving and stationary target acquisition and recognition (MSTAR) data demonstrate the robustness and effectiveness of the proposed method.

Diagnosis of skin lesions based on imaging techniques remains a challenging task because data (knowledge) uncertainty may reduce accuracy and lead to imprecise results. This paper investigates a new deep hyperspherical clustering (DHC) method for skin lesion medical image segmentation by combining deep convolutional neural networks and the theory of belief functions (TBF). The proposed DHC aims to eliminate the dependence on labeled data, improve segmentation performance, and characterize the imprecision caused by data (knowledge) uncertainty. First, the SLIC superpixel algorithm is employed to group the image into multiple meaningful superpixels, aiming to maximize the use of context without destroying the boundary information. Second, an autoencoder network is designed to transform the superpixels' information into potential features. Third, a hypersphere loss is developed to train the autoencoder network. The loss is defined to map the input to a pair of hyperspheres so that the network can perceive tiny differences. Finally, the result is redistributed to characterize the imprecision caused by data (knowledge) uncertainty based on the TBF. The proposed DHC method can well characterize the imprecision between skin lesions and non-lesions, which is particularly important for the medical procedures. A series of experiments on four dermoscopic benchmark datasets demonstrate that the proposed DHC yields better segmentation performance, increasing the accuracy of the predictions while can perceive imprecise regions compared to other typical methods.

It is still a challenging problem to characterize uncertainty and imprecision between specific (singleton) clusters with arbitrary shapes and sizes. In order to solve such a problem, we propose a belief shift clustering (BSC) method for dealing with object data. The BSC method is considered as the evidential version of mean shift or mode seeking under the theory of belief functions. First, a new notion, called belief shift, is provided to preliminarily assign each query object as the noise, precise, or imprecise one. Second, a new evidential clustering rule is designed to partial credal redistribution for each imprecise object. To avoid the “uniform effect” and useless calculations, a specific dynamic framework with simulated cluster centers is established to reassign each imprecise object to a singleton cluster or related meta-cluster. Once an object is assigned to a meta-cluster, this object may be in the overlapping or intermediate areas of different singleton clusters. Consequently, the BSC can reasonably characterize the uncertainty and imprecision between singleton clusters. The effectiveness has been verified on several artificial, natural, and image segmentation/classification datasets by comparison with other related methods.

It is quite an important and challenging problem for change detection (CD) from heterogeneous remote sensing images. The images obtained from different sensors (i.e., synthetic aperture radar (SAR) & optical camera) characterize the distinct properties of objects. Thus, it is impossible to detect changes by direct comparison of heterogeneous images. In this article, a new unsupervised change detection (USCD) method is proposed based on image translation. The cycle-consistent adversarial networks (CycleGANs) are employed to learn the subimage to subimage mapping relation using the given pair (i.e., before and after the event) of heterogeneous images from which the changes will be detected. Then, we can translate one image (e.g., SAR) from its original feature space (e.g., SAR) to another space (e.g., optical). By doing this, the pair of images can be represented in a common feature space (e.g., optical). The pixels with close pattern values in the before-event image may have quite different values in the after-event image if the change happens on some ones. Thus, we can generate the difference map between the translated before-event image and the original after-event image. Then, the difference map is divided into changed and unchanged parts. However, these detection results are not very reliable. We will select some significantly changed and unchanged pixel pairs from the two parts with the clustering technique (i.e.,$K$-means). These selected pixel pairs are used to learn a binary classifier, and the other pixel pairs will be classified by this classifier to obtain the final CD results. Experimental results on different real datasets demonstrate the effectiveness of the proposed USCD method compared with several other related methods.

The clustering of incomplete patterns is a very challenging task because the estimations may negatively affect the distribution of real centers and thus cause uncertainty and imprecision in the results. To address this problem, a new belief-based incomplete pattern unsupervised classification method (BPC) is proposed in this paper. First, the complete patterns are grouped into a few clusters by a classical soft method like fuzzy$c$-means to obtain the corresponding reliable centers and thereby are partitioned into reliable patterns and unreliable ones by an optimization method. Second, a basic classifier trained by reliable patterns is employed to classifies unreliable patterns and the incomplete patterns edited by the neighbors. In this way, most of the edited incomplete patterns can be submitted to specific clusters. Finally, some ambiguous patterns will be carefully repartitioned again by a new distance-based rule depending on the obtained reliable centers and belief functions theory. By doing this, a few patterns that are very difficult to classify between different specific clusters will be reasonably submitted to meta-cluster which can characterize the uncertainty and imprecision of the clusters due to missing values. The simulation results show that the BPC has the potential to deal with real datasets.

The classification analysis of missing data is still a challenging task since the training patterns may be insufficient and incomplete in many fields. To train a high-performance classifier and pursue high accuracy, we learn a credal classifier based on an optimized and adaptive multiestimation (OAME) method for missing data imputation on training and test sets. In OAME, some incomplete training patterns are estimated as multiple versions by a global optimization method thereby expanding the training set. On the other hand, the test pattern is adaptively estimated as one or multiple versions depending on the neighbors. For the test pattern with multiple versions, the corresponding outputs with different discounting factors (weights), represented by the basic belief assignments (BBAs), are fused for final credal classification based on evidence theory. The discounting factor contains two aspects: the importance and reliability factors that are used, respectively, to quantify the importance of the edited version itself and to represent the reliability of the classification result of the version. The effectiveness of OAME is widely validated on several real datasets and critically compared to other related methods.