Hyperspectral image classification based on ConvGRU and spectral–spatial joint attention

Abstract

In hyperspectral image classification, methods based on spectral–spatial joint attention mechanisms have demonstrated the ability to effectively enhance feature extraction. However, existing approaches still face limitations: spectral attention mechanisms often lack local–global feature interaction, spatial attention fails to fully exploit multi-scale information, and the joint modeling of spectral and spatial features remains insufficiently explored. To address these issues, this paper proposes a spectral–spatial joint attention network based on Convolutional Gated Recurrent Units (ConvGRU). First, a Local-Global Spectral Attention (LGSA) mechanism is designed, where one-dimensional convolution extracts local spectral features and fully connected layers enable global feature interaction. Second, a Multi-Scale Spatial Attention (MSSA) mechanism is introduced, employing three convolutional branches with different receptive fields to capture spatial features, followed by hierarchical feature fusion via 1 × 1 convolution. Finally, a channel-level feature fusion strategy based on ConvGRU is proposed, leveraging sequence modeling to achieve channel-wise joint enhancement of LGSA and MSSA, thereby enabling deep coupling of spectral and spatial features. Comparative experiments on three public datasets demonstrate that the proposed method outperforms seven state-of-the-art algorithms in terms of classification performance.