A hybrid capsule network for hyperspectral image classification

Limited training data, high dimensionality, image complexity, and similarity between classes are challenges confronting hyperspectral image (HSI) classification often resulting in suboptimal classification performance. The capsule network (CapsNet) preserves the hierarchy between different parts of the entity in an image by replacing scalar representations with vectors and can address these aforementioned issues. Motivated by CapsNet, this article presents a novel end-to-end deep learning (DL) architecture, the hybrid capsule network (HCapsNet), for HSI classification. HCapsNet employs 2-D and 3-D convolutional neural networks (CNNs) to extract higher level spatial and spectral features. In order to establish a route between capsules in the lower layers to the most-related capsule in the higher layer, dynamic routing (DR) is used to identify several overlapped objects during training sessions. Hyperparameter optimization is performed using nested cross-validation (nested-CV) to ensure thorough generalization evaluation. The proposed HCapsNet significantly outperformed the state-of-the-art methods in terms of overall classification accuracy on three widely used hyperspectral datasets, Indian Pines dataset achieving ($>3\%$, $p< {1}\times 1\times 10^{-11}$), the University of Pavia dataset ($>4\%$, $p< {1}\times 1\times 10^{-9}$), the Salinas Valley dataset ($>3\%$, $p< {1}\times 1\times 10^{-10}$) when using only 1% of the data for training. The performance of all CNN-based approaches degraded significantly with smaller training sample sizes. HCapsNet, therefore, is demonstrated to offer significant advantages in HSI classification problems with low sample sizes.