LKVHAN: Multiscale Large Kernel Vertical-Horizontal Attention Network for Hyperspectral Image Classification

Among deep learning-based hyperspectral image (HSI) classification models, convolutional neural networks (CNNs), transformers, Mamba, and large kernel CNNs (LKCNNs) models have been widely explored for HSI classification (HSIC). Nonetheless, these models suffer from several challenges: for example, 1) CNNs have a weak learning ability in capturing global information between land covers, due to their limited receptive field derived from small kernel convolutions; 2) transformers face quadratic computational complexity introduced by their self-attention mechanisms; and 3) LKCNNs require further enhancement in extracting global features, owing to the insufficient size of their receptive fields. To tackle these limitations, we propose a novel multiscale large kernel vertical–horizontal attention network (LKVHAN) for HSIC. The proposed LKVHAN consists of a 1 × 1 convolution module and a multiscale large kernel vertical–horizontal attention-based convolution (MSLKVHAC). The 1 × 1 convolution module is designed to facilitate band reduction, noise suppression, and spectral feature learning. Furthermore, the MSLKVHAC, leveraging a large vertical kernel size of 17 × 1 and a large horizontal kernel size of 1 × 17, extracts both local and global spatial features by incorporating a vertical attention-based convolution module and a horizontal attention-based convolution module. Extensive experimental results demonstrate that the proposed LKVHAN significantly outperforms ten state-of-the-art approaches across four widely used HSI datasets.