Hyperspectral Remote Sensing Image Feature Extraction And Classification

During the last 20 years, hyperspectral remote sensing has been playing an importantrole in many fields of both military and civil applications。compared Multi-spectrum remote sensing,the characteristics of hyperspectral remote sensing data are more of channels, higher spectral resolution, narrower bandwidth and larger amount of data。This for utilize remote sensing image carry on goal classify,discern and with importance value of research of following etc. But its enormous data amount and higher data bring greater difficulty to classifying of hyperspectral remote sensing image。The special properties of hyperspectral image data are first analyzed,Prove hyperspectral remote sensing image relatively stronger spectrum dependence , and analyzed its challenging influence on traditional classification methods。Before classification with usual classification method , feature selection and abstract should be carried through. The most common used method is Karhunen-Loeve (K-L) transform. We use the K-L transform as a tool to reduce the dimension of a set of vectors and eliminate the relevance of vectors. This is done by transforniing (rotate/scale) the base axis in such a way that the variance in the direction of on or more axis (dimension) is small enough to neglect. Towards a remote sensing image, first we get the correlation matrix from all bands data, then calculate the feature value and feature vectors, and last get the main ingredients which contain the availability message running out at 95%. Classification with the main ingredients can reduce calculation amount and improve classification result.The other classification method of hyperspectral remote sensing image is classification with original spectral characteristics. With limited training, samples, the accuracy of traditional classification methods is always unsatisfactory. SVM is a good method ,it can solve limited sample, high dimension, no-liner classification problem. the problems of kernel function selection , parameter determination and multi–classification are analyzed. Experiment results show that the generalization ability of SVM is strong, and its classification accuracy is better than traditional algorithms, no matter whether the training set is huge or small. Advantages of SVM are much more notable when the training set is small and the data dimension is high.