Dimension Reduction And Classification Of Hyperspectral Remote Sensing Images

Hyperspectral images usually refer to the spectral image with spectral resolution smaller than10nm. After half a century of development, remote sensing techniques receive major changes in theory, technique and applications. Among all, the emergence and rapid development of hyperspectral imaging technology is a very prominent one of this change. By hyperspectral sensors mounted in different space platforms, in the electromagnetic spectrum ultraviolet, visible, near-infrared and mid-infrared region, we take images of tens to hundreds of continuous and narrow spectral bands of the target area at the same time. Containing not only surface image information but also spectral information, hyperspectral image is the first true combination of spectra and images.Though hyperspectral images have higher spectral resolution, due to its high data dimension, the classifiers which applied to multispectral images usually cannot effectively apply to hyperspectral images. When the number of training samples is limited, as the number of spectra increases, the classification accuracy of classifiers like maximum likelihood, support vector machine will decrease. Such phenomenon is called "curse of dimensionality", also known as the Hughes phenomenon. In this paper, we did experiments with real hyperspectral data, trying to explain and study the curse of dimensionality problem. The main innovations of the paper are the following aspects:1. We propose two image Euclidean distance based manifold dimension reduction algorithms:image Euclidean distance based Isometric Feature Mapping and image Euclidean distance based Locally Linear Embedding. Considering the feature of hyperspectral image data, the two algorithms introduce image-based information to classic data-driven dimensionality reduction methods. The Euclidean distances in the hyperspectral data which simply indicate distances between pixels are expanded as hyperspectral data image distance (IMage Euclidean Distance, IMED) in the two algorithms. Using IMED-ISOMAP or IMED-LLE improves both the classification accuracy and visual effect of hyperspectral images.2^We propose a Spectral Angle-based self-organizing neural network classify algorithm. The algorithm uses spectral angle distance instead of the Euclidean distance, which significantly reduces the spectral variability in hyperspectral images. The experiment results show our proposed method improves the classification accuracy without raising computational complexity.