Based On The Mixed Pixel Model Hyperspectral Data Classification

Hyperspectral data classification is one of the most important problems in the field of remote sensing technology. The basic premise of hyperspectral data classification is that the spectral features of various materials are different from each other. With the development of sensor technology, hyperspectral sensor can collect in as many as several hundreds spectral bands at once. Because of the low spacial resolution, there are many mixed pixels in the hyperspectral image. The spectra of the mixed pixels and individual pure spectra are measurably different. Ignoring the mixed pixels, traditional algorithms in hyperspectral data classification are deteriorated. To solve the mixed-pixel problem, this paper puts emphases on both feature abstraction and classifier design.First, the independent component analysis (ICA) algorithm is used for feature abstraction. According to the linear mixing model, the spectra of mixed pixels are considered a linear combination of various individual pure spectra. In this paper, different material types are also considered to be statistical independent. Compared with principal component analysis (PCA) algorithm, the number of independent components is much smaller, meanwhile the independent components are informative for classification.Then the neural network is used for classification. Both the back-propagation algorithm based neural network and RBF neural network are applied to the classification. The classification precision of 220-band hyperspectral data is over 82%, which is remarkably superior to that of the conventional bayes classsifier.In the end, four different ways are researched for neural network classifier design. The BP neural network classifiers based on one-versus-rest algorithm and multilayer algorithm work well. According to the experiments upon the 220-band hyperspectral data, the classification precision is approaching to 90%.