GM2FFNet: Grouped Multiscale Multiangle Feature Fusion Network With Center Attention for Hyperspectral Image Classification

Convolutional neural networks and transformers have been extensively utilized in hyperspectral image classification due to their exceptional feature learning capabilities. However, many existing patch-based classification methods often neglect the fusion of multiscale and multiangle features and cannot fully capture the relationships between the central pixel and its neighboring pixels, which is likely to compromise the classification performance. To address these challenges, this article proposes a grouped multiscale multiangle feature fusion network with center attention (GM2FFNet). The proposed network unfolds in three key stages. First, a grouped multiscale extraction module captures and fuses spectral features at different scales using various kernels. Second, a grouped multiangle convolution module extracts features from multiple directions, with an adaptive fusion module further integrating this information. Finally, a spatial-spectral attention transformer module captures the correlations between the central pixel and its surrounding pixels. Experimental results on the Indian Pines, Pavia University, and Hi-LongKou datasets show that GM2FFNet achieves overall accuracies of 98.17%, 98.45%, and 99.03%, respectively, using only 7%, 0.7%, and 0.2% of the labeled samples, with a reduced number of parameters. Notably, the model significantly outperforms existing methods in several challenging categories across all three datasets. These results highlight both the effectiveness and robustness of the proposed network.