Spectral–spatial classification of hyperspectral remote sensing image based on capsule network

Abstract

Hyperspectral image (HSI) classification is a hot topic in remote sensing community; many researchers have made a great deal of effort in this domain. Recently, deep learning-based manner paves a new way to better classification accuracy. However, the flow of information between layers and layers (e.g. max-pooling) in traditional deep architecture turns out to be ineffective. In this study, a novel spectral–spatial classification framework for HSI based on Capsule Network (CapsNet) and dynamic routing algorithm is introduced. The proposed architecture is composed of a hybrid of 1D and 2D convolutional layers and two capsule layers for better and effective mining and combining features. Consequently, experiments on two popular dataset indicate that CapsNet-based framework outperforms traditional CNN-based counterparts. Moreover, this study reveals great potential for CapsNet model in the field of HSI classification.

Keywords

• hyperspectral imaging
• learning (artificial intelligence)
• feature extraction
• image classification
• remote sensing
• geophysical image processing
• convolutional neural nets
• hyperspectral remote sensing image
• capsule network
• hyperspectral image classification
• deep learning-based manner
• dynamic routing algorithm
• 2d convolutional layers
• capsule layers
• capsnet-based framework
• hsi classification
• deep architecture
• spectral–spatial classification framework
• cnn-based counterparts
• 1d convolutional layers