Hyperspectral Image Classification Based on 3D-2DCNN-ViT-KAN Architecture

This paper presents the 3D-2DCNN-ViT-KAN architecture to address challenges in hyperspectral image classification, including limited labeled samples and complexity. The architecture integrates 3D and 2DCNNs to alternately extract features in spatial and spectral dimensions. These features are converted into embedding vectors and input to the Vision Transformer (ViT) to capture global dependencies. To enhance feature interaction and capture nonlinear relationships, the MLP in ViT is replaced with a Kernel Attention Network (KAN). A Squeeze-and-Excitation Network (SENet) modules also added to improve channel feature processing and examine its insertion effect. Experimental results show that the 3D-2DCNN-SEblock2-ViT model outperforms traditional methods, achieving a 0.64% accuracy improvement on the Houston datasets. This research offers an efficient and accurate classification framework for hyperspectral image analysis.