Study Of Feature Extraction And Classification Of Hyperspectral Remote Sensing Image Based On Projection Pursuit And Nonlinear Principal Curves

For the feature extraction and classification of hyperspectral remote sensing image data, especially for the easy mixed classification problem, a projection index is established; by optimizing the projection index, a projection direction is calculated to classify the easy mixed classified pixels. An algorithm of multi direction projection pursuit is constructed to calculate multi directions of projection pursuit, the high dimensional data of hyperspectral remote sensing can be represented on a lower dimensional space more accurately. By combining the directions of PCA(Principal Component Analysis) and the directions of projection pursuit, an algorithm of feature extraction and classification is developed and the accuracy of classification is improved. For the bands selection of hyper spectral image, an algorithm called selected projection pursuit is established and a quick calculating method is developed, the calculating speed is increased 80 (the bands number of hyperspectral image) limes. The nonlinear principal curves is researched, it is the nonlinear expansion of PCA. The high dimensional data can be projected to a curve and the accuracy of classification will be improved. A simplified algorithm of nonlinear principal curves called nonlinear principal poly line is developed and its effect for feature extraction of hyperspectral data is researched. Based on the nonlinear principal poly line, a classification algorithm is established. The algorithm firstly calculates the nonlinear principal poly line go through the middle of trained samples of hyperspectral image, and then classifies every pixel by projecting the spectral vector of the pixel on the principal poly line. All the above algorithms are programmed by IDL language.