Hyperspectral Image Classification Techniques Based On Statistical Research

Hyperspectral imaging technique is a great leap in the remote sensing techniques which has been useful to many fields such as earth resources management, environmental monitoring, military reconnaissance, object tracking, object recognition and so on. It is well known that image classification is a base for many applications. However, classifying hyperspectral images only with traditional classification algorithms will result in low classification precision, data redundancy and great waste of resource. Considering the characteristic of hyperspectral images and the limitation of traditional classification techniques, effective classification algorithms using abundant information from hyperspectral images are studied in this paper.The main results are as follows:1. The non-Gaussian statistics of hyperspectral images and its difference images has been studied and the non-Gaussian models have been established appropriately for correlative processing techniques.2. A method about surface feature supervised classification for hyperspectral image is developed based on Markov random fields (MRF) and Gaussian mixture models. The dimension of the hyperspectral image is reduced by PCA, and the non-Gaussian stochastic model is built on prior of the dimension-reduced images and their difference images. Then the maximum a posteriori (MAP) is designed as an optimal criterion and the result of classification is obtained by the simulated annealing algorithm. Experimental results show that it is an accurate, efficient and robust algorithm for surface features labeling.3. Studying the theories of independent component analysis (ICA) which is applicable to the spectral mixture phenomenon and the non-Gaussian statistics in hyperspectral images, a method of unsupervised classification is presented based on the independent component analysis and Markov random fields, then an unsupervised classification result is obtained for hyperspectral images. The validity of this classification algorithm is demonstrated by experimental results.