TAttMSRecNet:Triplet-attention and multiscale reconstruction network for band selection in hyperspectral images. (February 2023)

• Record Type: Journal Article
• Authors: Nandi, Utpal; Roy, Swalpa Kumar; Hong, Danfeng; Wu, Xin; Chanussot, Jocelyn
• Is Part Of: Expert systems with applications. Volume 212(2023)
• Publication Date: 2023-02
• Subjects: Hyperspectral images (HSIs) -- Band selection (BS) -- Triplet attention -- Multiscale reconstruction network; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118797
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: The band selection (BS) is an essential task in hyperspectral images (HSIs), consisting of huge spectral bands with noises and redundancies. The attention mechanism can be used for BS. However, the existing attention-based BS schemes are failed to capture the cross-dimension interaction between the input spectral and spatial dimensions during computation of attention weights and may produce poor feature representations. Again, the used reconstruction network in the existing BS methods unable to detect the HSIs features in multiple scales by blindly increasing the depth of the network. To deal with these problems, a novel end-to-end unsupervised triplet-attention multiscale reconstruction network for BS (TAttMSRecNet ) has been proposed. The proposed network utilizes a triplet-attention mechanism having three parallel branches responsible to aggregate interactive cross-dimensional features between the spatial and spectral dimensions. After that, the network restores the original HSIs by using a 3D multiscale reconstruction network that applies multiple size convolution kernels to capture the discriminative HSIs features over the multiple scales where these features also communicate themselves to find the most efficacious HSIs information. In this way, the rich features are captured at a little computation cost, and the most informative bands can be effectively chosen for classification. Three standard data sets — Indian Pines (IP), Salinas (SA), and University of Pavia (UP) have been taken to conduct the experiments. The presented TAttMSRecNet can efficiently suppress the redundant or useless bands and selects more informative bands for better classification performance and also outperforms the other existing BS methods. Highlights: We propose an end-to-end network for band selection in hyperspectral images. It applies a triplet-attention with a multiscale reconstruction network. Captures the robust feature representations at a low computation overhead. Finds the most informative subset of spectral bands. Yields significant classification performance improvements.