Hyperspectral Image Classification Combining Spatial Information

Hyperspectral remote sensing is a frontier field in the development of remote sensing technology, the remote sensors capture useful information in hundreds of narrow spectral bands spanning from the visible to infrared spectrum. Pixels in hyperspectral image(HSI)are represented by vectors whose entries correspond to the spectral bands. Different materials usually reflect electromagnetic energy differently at certain wavelengths. This enables discrimination of materials based on spectral characteristics. HSI has found many applications in various fields, Such as agriculture, military and mineralogy. Compared with multi-spectral remote sensing, HSI has the advantage of high spectral resolution and contains large amounts of information. The ability of describing and discriminating the land-cover categories can thus be greatly improved, which makes it possible to analyse and process the spectral information precisely. Due to the large number of the hyperspectral image’s band, the data’s dimension is high and there inevitably exists a lot of noise in the image. How to mining useful information precisely and rapidly from large amount of HSI data for achieving high-precision classification is still a knotty problem.Most of the hyperspectral image classification model only considers spectral information while ignores the spatial information. The main research of this article is aiming to improve the classification accuracy based on the analysis of the existing algorithms and the characteristic of HSI data.In this article, a hyperspectral image classification method based on contextual information and joint sparse model is proposed. As the pixels in an image path usually have high correlation among them. If we expand the image path into a matrix, then there is a possible hidden low-rank structure under it. We propose to use a low-rank model to fit the expanded matrix, then the matrix can be decomposed into the sum of a low-rank matrix,a sparse matrix and a noise matrix. We can find the similar pixels to the central one by analyzing the obtained low-rank matrix. These pixels can be gathered to form a set of similar pixels, Then a joint sparse problem which assumes these similar pixels share the same sparsity pattern can be formed. The pixels can be represented by the same training samples of a certain dictionary with different weights, then we can obtain a row-sparse solution to this joint sparse model by enforcing a sparsity constraint on it. Experimental results on two widely used real hyperspectral images demonstrate the efficiency of the proposed methods, which has a stable and highest accuracy performance than other sparse representation-based approaches.