Hyperspectral Image Classification Using 3D Capsule-Net Based Architecture

Abstract

Convolution neural networks have received much interest recently in the categorization of hyperspectral images (HSI). Deep learning requires a large number of labeled samples in order to optimize numerous parameters due to the expansion of architecture depth and feature aggregation. Unfortunately, only few examples with labels are accessible, and the majority of spectral images are not labeled. For HSI categorization, the difficulty is how to acquire richer features with constrained training data. In order to properly utilize HSI features at various scales, a 3D Capsule-Net based supervised architecture is presented in this paper for HSI classification. First, the input data undergo incremental principal component analysis (IPCA) for dimensionality reduction. The reduced data are then divided into windows and given to a 3D convolution layer to get the shallow features. These shallow features are then used by 3D Capsule-Net to compute high-level features for classification of HSI. Experimental investigation on three common datasets demonstrates that the categorization performance by Capsule-Net based architecture exceeds a number of other state-of-the-art approaches.