Zhuoqun Fang

Semantic segmentation of high-resolution remote sensing imagery frequently encounters difficulties stemming from intricate scene layouts and significant variations in object scales. To tackle these issues, we propose a Global–Local Fourier-Aware Network (GLF-Net). The core innovation exploits Fourier transforms to decouple features for separate enhancement of global consistency and local details. Specifically, the model consists of three key components. (1) The Global Fourier-Aware Module (GFAM) selectively extracts and fuses low-frequency components to suppress background interference and improve scene-level semantic consistency. (2) The Multi-Scale Local Fusion Module (MSLF) aggregates high-frequency components with multi-scale local information to strengthen boundaries and fine-grained structures, thereby alleviating intra-class variations caused by scale differences. (3) A vHeat backbone is adopted to efficiently capture long-range spatial dependencies among geographic entities. Extensive evaluations conducted on the ISPRS Vaihingen and Potsdam benchmarks indicate that our GLF-Net yields superior results compared to existing methods. These findings confirm the model’s robustness in handling precise segmentation tasks, even when faced with intricate multi-scale variations arising from object size disparities.

Unsupervised domain adaptation has become an important approach to address spectral drift in cross-scene hyperspectral image (HSI) classification. However, existing methods suffer from two key defects in learning target domain features. First, without label supervision in the target domain, the model struggles to learn discriminative features and tends to overfit to the specific statistical properties of the source domain, resulting in insufficient generalization to the target domain. Second, while existing domain alignment methods can narrow the distribution gap between the source and target domains, they lack explicit constraints on the inter-class structure of the target domain, leading to blurred boundaries between different classes in the feature space and insufficient inter-class separability. To address these issues, this letter proposes an unsupervised domain adaptation framework based on masked robustness learning and class structure regularization (MRSR-UDA). This framework uses a masked robustness learning module to randomly mask pixels to create difficult samples with incomplete information. By using a combination of strong and weak constraints to augment view prediction consistency, MRSR-UDA forces the model to learn robust discriminative features from the context of spectral-spatial neighborhoods, thereby improving generalization to the target domain. Furthermore, a class structure regularization module is introduced to constrain the inter-class correlation matrix to remain consistent before and after data augmentation from a global class relationship perspective, explicitly optimizing inter-class separability in the target domain. The two modules are collaboratively optimized at the sample-level and the class-level to improve feature discriminability and structural rationality of the feature space, respectively. Experiments on two cross-domain HSI datasets verify the effectiveness of the proposed method.

Accurate segmentation of very high-resolution remote sensing images is vital for downstream tasks. Most semantic segmentation methods fail to fully consider the inherent characteristics of the images, such as intricate backgrounds, significant intraclass variance, and spatial interdependence of geographic object distribution. To address these challenges, we propose an efficient global–local scene awareness network with rotary position embedding (RoGLSNet). Specifically, we introduce the dynamic global filter (DGF) module to adaptively select frequency components, thereby mitigating interference from background noise. For high intraclass variance, the class center aware block (CCAB) performs class-level contextual modeling with spatial information integration. Additionally, the rotary position embedding (RoPE) is incorporated into vanilla attention to indirectly model the positional and distance relationships of geographic target objects. Extensive experimental results on two widely used datasets demonstrate that RoGLSNet outperforms the state-of-the-art (SOTA) segmentation methods. The code is available at https://github.com/bai101315/RoGLSNet

Prototype-driven few-shot learning holds promise for cross-domain hyperspectral image classification but encounters two key limitations: (1) over-reliance on visual features overlooks deep semantic information, hindering the differentiation of visually similar categories; and (2) inadequate semantic connections between local and holistic features restrict the utilization of spatial-spectral characteristics. To tackle these issues, we propose a multi-modal prototype correction and multi-dimensional knowledge distillation framework for cross-domain few-shot hyperspectral image classification (MCMD-CFSL). This framework implements a multi-modal prototype correction strategy that leverages pre-trained language models to extract spatial and spectral text features from spatial and spectral text descriptions, thereby aligning image spatial prototypes with spatial text features and spectral prototypes with spectral text features to enhance the recognition of diverse ground objects. Additionally, a multi-dimensional knowledge distillation mechanism is developed to ensure semantic consistency between local and holistic features across spatial-spectral dimensions, bolstering the model’s discriminative power and generalization ability. Experimental results demonstrate that MCMD-CFSL significantly surpasses existing deep learning and few-shot learning techniques on four hyperspectral datasets. The code is available at https://github.com/Li-ZK/MCMD-CFSL-2025.

The feature-level domain alignment based on deep learning techniques has greatly improved the performance of unsupervised domain adaptation (UDA) for hyperspectral image (HSI) classification. However, most of these methods leverage convolutional neural networks to capture local features, overlooking the comparable spatial and spectral global information shared by both the source and target domains. To overcome this issue, we propose a local auxiliary spatial-spectral decoupling transformer (LASDT) network to ease the learning of global domain-invariant information. The spatial and spectral global features of HSIs are extracted through a dual-branch design, preventing the feature coupling of different dimensions. In order to compress the model's parameter search space, a local auxiliary global feature extraction strategy is devised. First, local prior constraints are introduced by extracting primitive features using convolutional intra-token embedding. Next, the extraction of global spatial and spectral information from these primitive features is effectively achieved using the self-attention mechanism. Finally, a dynamic feature fusion mechanism is devised that enables the model to focus on features more conducive to transfer while suppressing irrelevant features. By using only standard adversarial domain alignment, LASDT achieves the state-of-the-art performance, demonstrating the model's superior capability in UDA for HSI classification.

Unsupervised domain adaptation methods can effectively mitigate the spectral drift in cross-scene hyperspectral image classification. Among them, adversarial training methods are particularly noteworthy due to their outstanding performance. However, due to the inherent mechanism of adversarial training, these methods suffer from continuing training instability and limited classifier generalizability. To overcome these limitations, this letter proposes a two-stage domain adaptation (TSDA) framework that incorporates self-distillation and test-time adaptation. The self-distillation strategy promotes stability during adversarial training by improving training consistency across iterations. Specifically, each batch includes a subset of data from the previous iteration, and self-distillation ensures that the output of this subset in the current iteration is consistent with the previous one. This mechanism stabilizes gradient computations during the training process, facilitating more robust parameter updates. Subsequently, the test-time adaptation module utilizes a limited set of unlabeled target domain samples to refine the classifier. During this stage, a confident learning module identifies and selects high-confidence pseudo-labels to optimize the classifier, enhancing its generalizability in the target domain. Thus, TSDA facilitates domain adaptation during both the training and testing stages. Experimental results on two cross-domain datasets demonstrate the effectiveness of the proposed method. The code is available at https://github.com/Li-ZK/TSDA-2025.

Enhancing target domain discriminability is a key focus in Unsupervised Domain Adaptation (UDA) for HyperSpectral Image (HSI) classification. However, existing methods overlook bringing similar cross-domain samples closer together in the feature space to achieve the indirect transfer of source domain classification knowledge. To overcome this issue, we propose a Multi-Task Learning-based Domain Adaptation (MTLDA) method. MTLDA incorporates an inductive transfer mechanism into adversarial training, transferring the source classification knowledge to the target representation learning during the process of domain alignment. To enhance the target feature discriminability, we propose utilizing dual-domain contrastive learning to construct related tasks. A shared mapping network is employed to simultaneously perform Source domain supervised Contrastive Learning (SCL) and Target domain unsupervised Contrastive Learning (TCL), ensuring that similar samples across domains are positioned closely in the feature space, thereby improving the cross-scene HSI classification accuracy. Furthermore, we design a feature-level data augmentation method based on feature masking to assist contrastive learning tasks and generate more varied training data. Experimental results obtained from testing on three prominent HSI datasets demonstrate the MTLDA method’s superior efficacy in the realm of cross-scene HSI classification.

Recent studies have shown that deep domain adaptation (DA) technique has achieved remarkable results in crossdomain hyperspectral image (HSI) classification task. However, these DA methods assume that the source and target domains share the same classes, which may not hold true in real-world applications. Under open-set conditions, since the target domain may contain classes unseen in the source domain, direct domain alignment can lead to negative transfer phenomena. Moreover, the presence of multiple unknown classes in the target domain makes it difficult to learn more discriminative classification boundaries between known and unknown classes. To address these issues, we propose an open-set domain adaptation method for HSI classification based on weighted generative adversarial networks and dynamic thresholding. First, we introduce a class anchor strategy to learn the metric space of known classes in the source domain. By calculating the similarity between the target domain samples and the class anchor, we compute the reliability weights of the samples belonging to known classes. Then, based on these weights, we design an instance-level weighted domain adversarial learning strategy to better align samples that are more likely to belong to known classes, avoiding negative transfer phenomena. Finally, we propose a dynamic thresholding method to learn the classification boundaries between known and unknown classes in the feature space and reject unknown class samples, thereby separating known class samples in the target domain. The experimental results on four cross-scene HSI classification tasks demonstrate that our proposed method outperforms some existing methods. The code is available at https://github.com/Li-ZK/WGDT.

Unsupervised domain adaptation (UDA) is an effective approach for cross-scene hyperspectral image (HSI) classification. Notably, deep learning–based UDA methods have shown potential by leveraging their powerful capability of feature extraction. Most of these methods rely on traditional convolutional neural networks (CNNs), which typically emphasize local feature extraction but lack the capacity to capture global features, thereby limiting their ability to process complex HSI data. Given the ability of capturing long-range dependencies, the transformer demonstrates superior performance than CNNs in many fields, making its employment in UDA a promising approach. However, the original transformer has only spatial attention, which is inadequate for HSI data with spatial and spectral dimensions. Moreover, the limited available labeled HSI data cannot offer the optimization of the transformer model and constrains its capacity. To address these problems, a regularized spatial–spectral transformer for domain adaptation (RSTDA) is proposed. To effectively extract features from HSI data, a spatial–spectral transformer network is designed for HSI data specifically, which contains spatial–spectral attention modules to facilitate the extraction of spatial–spectral features. Also, convolutional layers are introduced at different stages of the transformer as a regulation technique, making it possible to train the transformer model on limited HSI data. Finally, a smooth adversarial training strategy is adopted to decrease the domain discrepancy and improve the generalization of the transformer, thus enhancing the accuracy of the cross-scene HSI classification. Experimental results on three datasets demonstrate that RSTDA surpasses the existing state-of-the-art UDA methods for HSI classification.