A Novel Classification Framework for Hyperspectral Image Classification Based on Multi-Scale Dense Network

Abstract

The combined use of spatial information and spectral information has been widely applied to hyperspectral image (HSI) classification. In recent years, multiscale spatial-spectral convolutional neural networks (CNN) have been introduced for hyperspectral image classification (HSIC). However, most of HSIC methods based on CNN mainly use patches as input for classifier. This may cause a lot of redundancy in the training and testing process, and reduce the efficiency of the model. In order to address this problem, we design a novel image-based classification framework. Based on this framework, we propose a multi-scale dense network for HSIs, called HyMSDN. This network merges features from different scales through a feature pyramid structure. Experimental results on real hyperspectral dataset verify the efficiency and effectiveness of the proposed framework, with superior performances compared with other related methods.