Hyperspectral Image Classification Using 2D-Convolutional Capsule Network with EM Routing

Convolutional neural networks (CNNs) have lately proved exceptional efficacy in hyperspectral image (HSI) classification. However, the inherent complexity of CNN structures and the high dimensionality of HSIs often lead to significant training efforts. Additionally, the accuracy of CNNs decreases for the lack of labeled samples which poses a great challenge to overcome. To mitigate these challenges, a novel approach is proposed employing a 2D convolutional capsule network (2DConvCapsNet) with the Expectation-Maximization (EM) routing algorithm. Unlike traditional CNNs, the 2D-ConvCapsNet architecture is tailored to efficiently handle the spatial and spectral intricacies of HSIs. The capsule networks encapsulate spatial and spectral information separately, thus reducing the computational complexity compared to conventional CNNs and enhancing efficiency without compromising accuracy. The capsule units provide the precise identification of spectral signatures and part-whole relationships in the data, while EM routing guarantees viewpoint consistency. The capsule units enable effective identification of spectral signatures and part-whole relationships in the data while EM routing ensures viewpoint-invariance. The efficacy of the proposed framework is validated across the Indian pine dataset, provided an accuracy of 96.41 % whereas the framework of the CNNs model gained an accuracy of $\mathbf{9 9. 9 7 \%}$. The outcomes demonstrate superior performance achieving higher accuracy while significantly reducing training efforts and offering a balance between accuracy and computational efficiency in real world applications.