In recent years, semantic segmentation with pixel-level classification has become one of the types of research focus in the field of polarimetric synthetic aperture radar (PolSAR) image interpretation. Fully convolutional network (FCN) can achieve end-to-end semantic segmentation, which provides a basic framework for subsequent improved networks. As a classic FCN-based network, U-Net has been applied to semantic segmentation of remote sensing images. Although good segmentation results have been obtained, scalar neurons have made it difficult for the network to obtain multiple properties of entities in the image. The vector neurons used in the capsule network can effectively solve this problem. In this paper, we propose a complex-valued (CV) U-Net with a CV capsule network embedded for semantic segmentation of a PolSAR image. The structure of CV U-Net is lightweight to match the small PolSAR data, and the embedded CV capsule network is designed to extract more abundant features of the PolSAR image than the CV U-Net. Furthermore, CV dynamic routing is proposed to realize the connection between capsules in two adjacent layers. Experiments on two airborne datasets and one Gaofen-3 dataset show that the proposed network is capable of distinguishing different types of land covers with a similar scattering mechanism and extracting complex boundaries between two adjacent land covers. The network achieves better segmentation performance than other state-of-art networks, especially when the training set size is small.

Three-dimensional target reconstruction from 2-D synthetic aperture radar (SAR) images can reduce the difficulty of data acquisition when compared to 3-D imaging technologies. Pixel2mesh was well designed for 3-D target reconstruction from 2-D optical images. However, when it is directly applied to small SAR datasets, overfitting is prone to occur. In this letter, we propose a lightweight Pixel2mesh for 3-D target reconstruction from a single 2-D SAR image. Based on the original Pixel2mesh, we improve two subnetworks: feature extraction and 3-D deformation. First, we lightweight the graph residual network (G-ResNet) module in the 3-D deformation subnetwork to avoid overfitting. Second, we add a decoder after the feature extraction subnetwork to reconstruct the input image and then obtain the image reconstruction loss to guide the training of the whole network. In addition, we modify the Laplace regularization term for the training of the proposed network, aiming to make the deformation more reasonable. Experiments are implemented on the Gotcha dataset, where 2-D images of seven cars are obtained by performing 2-D imaging. The 3-D labels of these cars are generated by using their CAD models downloaded from public websites. Experimental results verify that our network can achieve better 3-D reconstruction results than the original Pixel2mesh.

Deep learning approaches have significantly enhanced the classification accuracy of hyperspectral images (HSIs). However, the classification process still faces difficulties such as those posed by high data dimensions, large data volumes, and insufficient numbers of labeled samples. To enhance the classification accuracy and reduce the data dimensions and training needed for labeled samples, a 3D fully convolutional neural network (3D-FCNN) model was developed by including a bottleneck attention module. In such a model, the convolutional layer replaces the downsampling layer and the fully connected layer, and 3D full convolution is adopted to replace the commonly used 2D and 1D convolution operations. Thus, the loss of data in the dimensionality reduction process is effectively avoided. The bottleneck attention mechanism is introduced in the FCNN to reduce the redundancy of information and the number of labeled samples. The proposed method was compared to some advanced HSI classification approaches with deep networks, and five common HSI datasets were employed. The experiments showed that our network could achieve considerable classification accuracies by reducing the data dimensionality using a small number of labeled samples, thereby demonstrating its potential merits in the HSI classification process .

Driven by deep learning, three-dimensional (3-D) target reconstruction from two-dimensional (2-D) synthetic aperture radar (SAR) images has been developed. However, there is still room for improvement in the reconstruction quality. In this paper, we propose a structurally flexible occupancy network (SFONet) to achieve high-quality reconstruction of a 3-D target using one or more 2-D SAR images. The SFONet consists of a basic network and a pluggable module that allows it to switch between two input modes: one azimuthal image and multiple azimuthal images. Furthermore, the pluggable module is designed to include a complex-valued (CV) long short-term memory (LSTM) submodule and a CV attention submodule, where the former extracts structural features of the target from multiple azimuthal SAR images, and the latter fuses these features. When two input modes coexist, we also propose a two-stage training strategy. The basic network is trained in the first stage using one azimuthal SAR image as the input. In the second stage, the basic network trained in the first stage is fixed, and only the pluggable module is trained using multiple azimuthal SAR images as the input. Finally, we construct an experimental dataset containing 2-D SAR images and 3-D ground truth by utilizing the publicly available Gotcha echo dataset. Experimental results show that once the SFONet is trained, a 3-D target can be reconstructed using one or more azimuthal images, exhibiting higher quality than other deep learning-based 3-D reconstruction methods. Moreover, when the composition of a training sample is reasonable, the number of samples required for the SFONet training can be reduced.

Polarimetric synthetic aperture radar (PolSAR) semantic segmentation remains challenging due to large intra‑class scattering variability, high inter‑class scattering similarity, and boundary ambiguity caused by speckle. Real‑valued networks discard polarimetric phase information, while existing complex‑valued methods often directly extend real‑valued architectures without coordinated design. To address these issues, this paper proposes Complex‑Valued HRU‑Net, a lightweight encoder‑decoder network. A three‑stage complex‑valued HRNet encoder preserves fine spatial details through parallel multi‑resolution streams, and a complex‑valued U‑Net decoder progressively restores high‑resolution semantic feature maps. At the second skip connection, a Complex‑Valued Cross‑Gated Attention Enhancement (CV‑CGAE) module suppresses speckle‑contaminated responses and selectively transfers boundary‑ and texture‑related cues via sequential enhancement, cross‑attention interaction, and gated recalibration. Experiments on three airborne and spaceborne PolSAR datasets—AIRSAR San Francisco, AIRSAR Flevoland, and RADARSAT‑2 San Francisco—show that the proposed method achieves mean intersection‑over‑union (mIoU) values of 97.21%, 98.25%, and 96.38%, respectively. In the 15‑class Flevoland scene, all per‑class IoU values exceed 94%. Parameter‑fair comparisons and phase randomization experiments confirm that the performance gains stem from explicit use of phase information, not merely increased model capacity. The network provides improved boundary delineation and post‑hoc physical interpretability for complex‑valued PolSAR segmentation.

Improving detection performance across diverse synthetic aperture radar (SAR) scenes remains challenging because different datasets exhibit different levels of target–background separability. To address this issue, we propose a constant false alarm rate (CFAR)-guided mode-selective network (CGMSN), which selects an appropriate feature-fusion mode according to the CFAR target–background separation margin. Specifically, CFAR is used as an interpretable statistical tool to construct an anomaly response map. The separation margin is then calculated by comparing the average CFAR anomaly responses of annotated target regions and their surrounding contextual backgrounds. Based on this indicator, a You Only Look Once version 8 (YOLOv8)-based mode-selective detector is constructed with three key components. First, a lightweight representation-enhanced backbone that integrates ResNet18 and a dilated convolutional spatial pyramid (DCSP) module is adopted to improve contextual representation while maintaining moderate model complexity. Second, a mode-selective neck (MSN) is designed with three predefined fusion modes, where the appropriate fusion depth is selected according to the CFAR-guided target–background separation margin of each dataset. Third, a complete intersection over the union modulated head (CMH) is developed to enhance classification-regression alignment and suppress clutter-induced responses. Experiments on SAR-Aircraft-1.0, High-Resolution SAR Images Dataset (HRSID), and SAR Ship Detection Dataset (SSDD) indicate that datasets with smaller CFAR target–background separation margins benefit from deeper fusion, while datasets with larger separation margins can adopt shallower fusion. Moreover, the proposed CGMSN achieves superior performance over representative detectors, demonstrating its effectiveness on the evaluated SAR datasets with diverse scene characteristics.