Multi-scale Convolutional Capsule Network for Hyperspectral Image Classification

Dongyue Wang, Qin Xu, Yong Xiao, Jin Tang, Bin Luo

Abstract: The conventional CNNs-based hyperspectral image classification faces the challenges of quite limited training samples which lead to over fitting and dissatisfied ability to describe the correlation between features. Recent Capsules network can deal with the data of limited training samples and capture the correlation between the features, but the low-level features extraction is used by a single-scale CNN whose feature representation capability is limited. In this paper, we propose a multi-scale convolutional capsule network for hyperspectral image classification, which is composed of a multi-scale convolutional layer, a single-scale convolutional layer, a PrimaryCaps layer, a DigitCaps layer and a fully connected layer. The proposed network can learn high-level spectral-spatial features with limited training data and is robust to rotation and affine transformation. The comparison experiments with five state-of-the-arts on two well-known datasets demonstrate the efficiency of the proposed method.