Spectral and Multi-Spatial-Feature Based Deep Learning for Hyperspectral Remote Sensing Image Classification

摘要

Hyperspectral data has a strong ability in information expression. In this paper, we will extract a variety of spectral features and Multi-spatial-dominated features. In order to make better use of the relationship between spatial neighborhood pixels, we introduce spatial features with two different window scales, which can be give us more abundant spatial information, and then we used a novel framework to merge this extracted features. This deep learning framework is made of sparse component analysis (SPCA), deep learning architecture, and logistic regression. For hyperspectral image classification, stacked autoencoders is an efficient deep learning framework. In detail, compared with principle component analysis (PCA), SPCA has a better effect on dimensionality reduction of nonlinear data, especially for hyperspectral data. The public data set Pavia Centre scene and Pavia University scene are used to test our proposed algorithm. Experimental results demonstrate that the proposed approach outperforms the compared. It also shows that the hyperspectral data classification based on deep learning has an excellent application prospect.