Li Fang

Hyperspectral image classification (HSIC) can degrade under in-plane rotations because spatial structures change orientation and the learned spectral–spatial coupling in patch-based models becomes less stable. Most existing approaches improve robustness mainly through data augmentation, without explicitly linking geometric invariance with spectral calibration in an end-to-end manner. To address this limitation, we propose RIHP-Net, a dual-path framework that models rotation as a shared latent factor for coordinated spectral and spatial processing. In the spectral path, we proposed a Rotation-Conditioned Spectral Fusion (RCSF) branch that learns band interactions and predicts a compact rotation code to adaptively calibrate per-band responses. In contrast, in the spatial path, we proposed the Harmonic Fiber Spatial Pooling (HFSP) branch to rotation-consistent geometric cues via an equivariant backbone and harmonic pooling to obtain a rotation-invariant spatial descriptor. A bi-directional rotation-gating module then fuses both representations by conditioning feature aggregation on the predicted rotation code. We evaluate RIHP-Net across multiple benchmarks with different training-data regimes, including both extremely limited-label and higher-label settings, and report consistent gains in rotation robustness while maintaining competitive accuracy under standard (non-rotated) evaluation. The source code is available here https://github.com/Arafat-Aljawari/RIHP-Net.

Significant progress has been made in semi-supervised hyperspectral image (HSI) classification regarding feature extraction and classification performance. However, due to high annotation costs and limited sample availability, semi-supervised learning still faces challenges such as boundary label diffusion and pseudo-label instability. To address these issues, this paper proposes a novel semi-supervised hyperspectral classification framework integrating spatial prior information with a dynamic learning mechanism. First, we design an Edge-Aware Superpixel Label Propagation (EASLP) module. By integrating edge intensity penalty with neighborhood correction strategy, it mitigates label diffusion from superpixel segmentation while enhancing classification robustness in boundary regions. Second, we introduce a Dynamic History-Fused Prediction (DHP) method. By maintaining historical predictions and dynamically weighting them with current results, DHP smoothens pseudo-label fluctuations and improves temporal consistency and noise resistance. Concurrently, incorporating condifence and consistency measures, the Adaptive Tripartite Sample Categorization (ATSC) strategy implements hierarchical utilization of easy, ambiguous, and hard samples, leading to enhanced pseudo-label quality and learning efficiency. The Dynamic Reliability-Enhanced Pseudo-Label Framework (DREPL), composed of DHP and ATSC, strengthens pseudo-label stability across temporal and sample domains. Through synergizes operation with EASLP, it achieves spatio-temporal consistency optimization. Evaluations on four benchmark datasets demonstrate its capability to maintain superior classification performance.

Recently, domain adaptation techniques have been introduced for cross-domain few-shot hyperspectral image (HSI) classification tasks, but effectively aligning the source and target domains remains a core challenge. Current domain adaptation methods focus mainly on the instance-level, which can easily damage the category separability of the target domain features. To address this, we propose a cross-domain multilevel contrast (CDMLC) method. We implement cross-domain few-shot learning (FSL) at the instance-level, category-level, and category-distribution-level. At the instance-level, meta-learning is employed to capture meta-knowledge. At the category-level, we design an intradomain contrastive loss with domain-mixed dictionaries. The intradomain contrastive loss can ensure feature compactness within each category in both domains and similar feature representations for semantically similar categories across domains. At the category-distribution-level, we perform category distribution fitting by supervised and unsupervised Gaussian mixture models (GMMs) in the source and target domains, respectively. Interdomain contrastive loss is constructed on the basis of the category distribution to promote domain alignment. Extensive experiments are conducted on five publicly available target HSI datasets, and the results demonstrate that our method outperforms the state-of-the-art FSL methods. The code is available at https://github.com/AAAA-CS/CDMLC.

Identifying correspondences of road segments in different road networks, namely road-network matching, is an essential task for road network-centric data processing such as data integration of road networks and data quality assessment of crowd-sourced road networks. Traditional road-network matching usually relies on feature engineering and parameter selection of the geometry and topology of road networks for similarity measurement, resulting in poor performance when dealing with dense and irregular road network structures. Recent development of graph neural networks (GNNs) has demonstrated unsupervised modeling power on road network data, which learn the embedded vector representation of road networks through spatial feature induction and topology-based neighbor aggregation. However, weighting spatial information on the node feature alone fails to give full play to the expressive power of GNNs. To this end, this paper proposes a Spatial Pattern-aware Graph EMbedding learning method for road-network matching, named SP-GEM, which explores the idea of spatially-explicit modeling by identifying spatial patterns in neighbor aggregation. Firstly, a road graph is constructed from the road network data, and geometric, topological features are extracted as node features of the road graph. Then, four spatial patterns, including grid, high branching degree, irregular grid, and circuitous, are modelled in a sector-based road neighborhood for road embedding. Finally, the similarity of road embedding is used to find data correspondences between road networks. We conduct an algorithmic accuracy test to verify the effectiveness of SP-GEM on OSM and Tele Atlas data. The algorithmic accuracy experiments show that SP-GEM improves the matching accuracy and recall by at least 6.7% and 10.2% among the baselines, with high matching success rate (>70%), and improves the matching accuracy and recall by at least 17.7% and 17.0%, compared to the baseline GNNs, without spatially-explicit modeling. Further embedding analysis also verifies the effectiveness of the induction of spatial patterns. This study not only provides an effective and practical algorithm for road-network matching, but also serves as a test bed in exploring the role of spatially-explicit modeling in GNN-based road network modeling. The experimental performances of SP-GEM illuminate the path to develop GeoEmbedding services for geospatial applications.

Recently, Transformer has achieved significant success in hyperspectral image (HSI) classification task. However, most Transformer and its variants focus more on spatial domain global feature learning, ignoring the complementary characteristics provided by frequency domain features. The Fast Fourier Transform (FFT), due to its sensitivity for frequency domain information, has become a primary tool for frequency domain analysis. However, different frequency bands are often assigned the same attention values, and the differences between different frequency bands are not considered. To fully explore and fusion spatial and frequency domain features, we propose a multi-scale spatial-frequency domain cross transformer (SFDCT-Former) network. We design a two-branch structure for spatial domain and frequency domain feature learning: one branch utilizes the multi-head self-attention (MHSA) module for spatial domain feature learning, while the other incorporates a Multi-Frequency Domain Transformer (MFre-Former) encoder for frequency domain feature learning. The MFre-Former encoder divides the frequency domain into non-overlapping frequency bands and assigns distinct attention to each frequency band, therefore, different frequency domain information can be captured more precisely. Furthermore, to fuse the spatial and frequency domain features, we design a Multi-Level Cross Attention (MLCA) fusion module. The MLCA module effectively combines spatial and frequency domain features at different levels to better capture their complementary characteristics. Extensive experiments conduct on four publicly available HSI datasets demonstrate that the proposed method outperforms nine state-of-the-art methods in classification performance. The code is available at https://github.com/AAAA-CS/SFDCT-Former.

Self-supervised contrastive learning is a powerful pre-training framework for learning the invariant features from the different views of remote sensing images, therefore, the performance of contrastive learning heavily depends on the generation of views. Current view generation is primarily accomplished through different transformations, and the types and parameters of the transformations are require hand-crafted. Hence, the diversity and discriminability of generated views cannot be guaranteed. To address this, we propose a multi-type views optimization method to optimize these transformations. We formulate contrastive learning as a min-max optimization problem, and transformation parameters are optimized by maximizing the contrastive loss. The optimized transformations encourage the negative sample pairs to be close and the positive sample pairs to be far apart. Different from the current adversarial view generation methods, our method can optimize both photometric transformations and geometric transformations. For remote sensing images, the geometric transformation is more critical for view generation, while the existing view optimization methods fail to achieve this. We consider the hue, saturation, brightness, contrast, and geometric rotation transformations in contrastive learning, and evaluate the optimized views on the downstream remote sensing images semantic segmentation task. Extensive experiments are carried on the three remote sensing image segmentation datasets, including ISPRS Potsdam dataset, ISPRS Vaihingen dataset, and LoveDA dataset. Results show that the learned views obtain highly advantages compared to the hand-crafted views and other optimized views. The code associated with this paper has been released and can be accessed at https://github.com/AAAA-CS/AMView.

In this article, a spectral method for multiview point cloud registration is presented. The primary challenge of existing multiview registration methods is to accurately estimate the weights of the pose graph. Previous multiview registration methods rely on predicting the overlap between pairs of point clouds to assign weights for pairwise transformation matrices. This global way may result in higher weights being assigned to pairs of point clouds that are structurally similar but actually have no or low overlap. In addition, the learning-based approach is biased for training scenarios. However, we obtain the weights by measuring the confidence of each correspondence used for the estimation of the pairwise transformation matrix. Specifically, we construct spatial compatibility matrices for the correspondences and compute the edge weights of the pose graphs using spectral decomposition (SD). Considering the compatibility between correspondences, the proposed method is robust to outliers. Simultaneously, to mitigate the disruption of outliers to the global pose estimation and enhance the computational efficiency, we propose an overlap score estimation method based on the algebraic connectivity of the graph for pruning of fully connected pose graphs. For the initial sparse pose graph, we apply iteratively reweighted least-squares (IRLS) to refine the global transformation and design a new reweighting function based on the historical weighted average (HWA). The nonlearning framework empowers the proposed method to have better generalization to unknown scenarios. The experimental results further validate the performance of our method which achieves about 12.9% and 21.4% lower rotation and translation errors on the ScanNet dataset compared to the state-of-the-art method. The source code is available at https://github.com/swccj/SMVR .

Point clouds are widely used in remote sensing applications, e.g., 3D object classification, semantic segmentation, and building reconstruction. Generating dense and uniformly distributed point clouds from low-density ones is beneficial to 3D point cloud applications. The traditional methods mainly focus on the global shape of 3D point clouds, thus ignoring detailed representations. The enhancement of detailed features is conducive to generating dense and uniform point clouds. In this paper, we propose a point cloud upsampling network to improve the detail construction ability, named PU-CTG. The proposed method is implemented by a cross-transformer-fused module and a GRU-corrected module. The aim of the cross-transformer module is to enable the interaction and effective fusion between different scales of features so that the network can capture finer features. The purpose of the gated recurrent unit (GRU) is to reconstruct fine-grained features by rectifying the feedback error. The experimental results demonstrate the effectiveness of our method. Furthermore, the ModelNet40 dataset is upsampled by PU-CTG, and the classification experiment is applied to PointNet to verify the promotion ability of this network.