RESUMO
Cross-scene hyperspectral image classification (HSIC) poses a significant challenge in recognizing hyperspectral images (HSIs) from different domains. The current mainstream approaches based on domain adaptation (DA) methods need to access target data when aligning distributions between domains, limiting the applicability of the model. In contrast, recent domain generalization (DG) methods aim to directly generalize to unseen domains, eliminating the requirements for target data during training. Nonetheless, most DG-based methods overly focus on randomizing sample styles, leading to semantically compromised samples. In addition, broadening the source distribution without ensuring reasonable support may result in undesired extended distributions. To address these issues, we propose a novel DG network with frequency disentanglement and data geometry (FDGNet) for cross-scene HSIC. Specifically, we first develop a spectral-spatial encoder based on frequency disentanglement (FDSS encoder), which facilitates synthesized domains to preserve their semantic consistency while simulating interdomain gaps with the source domain. Second, to avoid the generation of unrealistic samples, we incorporate data geometry into adversarial training. This helps diversify new domains while keeping the data geometry of extended domains in an explainable support. To improve the learning of domain-invariant representation, we propose an intermediate domain sampling strategy based on the class-wise perceptual manifold. This strategy synthesizes reliable intermediate domains by sampling from class-wise manifold flows estimated over the source and extended domains. Extensive experiments and analysis on three public HSI datasets yield the superiority of our proposed FDGNet. The codes will be available from the website: https://github.com/Qba-heu/FDGNet.
RESUMO
Deep learning (DL) based methods represented by convolutional neural networks (CNNs) are widely used in hyperspectral image classification (HSIC). Some of these methods have strong ability to extract local information, but the extraction of long-range features is slightly inefficient, while others are just the opposite. For example, limited by the receptive fields, CNN is difficult to capture the contextual spectral-spatial features from a long-range spectral-spatial relationship. Besides, the success of DL-based methods is greatly attributed to numerous labeled samples, whose acquisition are time-consuming and cost-consuming. To resolve these problems, a hyperspectral classification framework based on multi-attention Transformer (MAT) and adaptive superpixel segmentation-based active learning (MAT-ASSAL) is proposed, which successfully achieves excellent classification performance, especially under the condition of small-size samples. Firstly, a multi-attention Transformer network is built for HSIC. Specifically, the self-attention module of Transformer is applied to model long-range contextual dependency between spectral-spatial embedding. Moreover, in order to capture local features, an outlook-attention module which can efficiently encode fine-level features and contexts into tokens is utilized to improve the correlation between the center spectral-spatial embedding and its surroundings. Secondly, aiming to train a excellent MAT model through limited labeled samples, a novel active learning (AL) based on superpixel segmentation is proposed to select important samples for MAT. Finally, to better integrate local spatial similarity into active learning, an adaptive superpixel (SP) segmentation algorithm, which can save SPs in uninformative regions and preserve edge details in complex regions, is employed to generate better local spatial constraints for AL. Quantitative and qualitative results indicate that the MAT-ASSAL outperforms seven state-of-the-art methods on three HSI datasets.