Cross-Scale Window Attention for Cross-Domain Hyperspectral Image Classification

Hyperspectral image classification (HSIC) is crucial in several fields, but relies on a large number of labeled samples. Due to the high cost of manual annotation and the scarcity of samples, how to achieve high-precision classification with limited samples has become a major challenge. To cope with this challenge, a novel framework named Cross-Scale Window Attention Cross-Domain Few-Shot Learning (CSWA) is proposed in this paper. Overall, CSWA interactively utilizes training samples from source and target domains to learn their global class representations separately, and adopts an inter-domain information interaction strategy to alleviate the domain bias problem in cross-domain learning. For feature extraction, the local window attention mechanism is applied, and a local window partitioning strategy with a pyramidal hierarchical downsampling structure is introduced. Through intra-window self-attention computation and cross-window feature restructuring, the hierarchical characterization from local spectral–spatial features to global contextual semantics is gradually realized, which effectively takes into account the fine-grained local discriminative properties and global scene consistency in hyperspectral image classification tasks. In addition, a cross-scale convolutional network (XPConvNet) is designed to enable the model to capture spectral–spatial features of different dimensions and scales, which acts at the initial stage of feature extraction to help the network learn more robust and discriminative feature representations from the rich hyperspectral information, and lays the foundation for subsequent spectral–spatial fine-grained feature extraction. Extensive experimental results from three benchmark datasets show that CSWA can outperform the state-of-the-art methods with a small number of labeled samples.