A Hyperspectral Band Selection Network Combining Siamese Network and Local-Global Attention

Hyperspectral images can provide dozens to hundreds of continuous spectral bands, greatly enhancing the richness of information. However, the redundancy among adjacent bands leads to increased data processing complexity. Despite the recent introduction of numerous band selection methods, there has been limited focus on incorporating context information from the entire spectral range into this task. Furthermore, researchers have primarily concentrated on band informativeness and sparse representation for reconstructing all bands, often overlooking the separability of classes in downstream tasks. To address these challenges, we propose a hyperspectral band selection network combining Siamese Network Local-Global Attention (SLGA). This approach first segments the hyperspectral image into homogeneous regions and constructs sample pairs based on a random elimination strategy. Next, it utilizes the Local-Global Attention (L-GA) mechanism to obtain band weights that capture both local and global spectral structures. These reweighted bands are then fed into a twin network to obtain their high-dimensional representations, compute loss values, and update network parameters. Finally, extensive classification experiments using SVM, KNN, and LDA classifiers are conducted on the Indian Pines and Botswana hyperspectral image datasets. The results from these experiments on benchmark datasets demonstrate that the proposed SLGA method performs exceptionally well, outperforming state-of-the-art algorithms.