ISOMAP-BASED SUBSPACE ANALYSIS FOR THE CLASSIFICATION OF HYPERSPECTRAL DATA

A new object-oriented mapping approach is proposed based on nonlinear subspace feature analysis of hyperspectral data. A nonlinear manifold learning approach ISOMAP were utilized to obtain subspace feature representation of hyperspectral remote sensing imagery. Afterwards, the extracted subspace feature images were fed into the object-oriented system. Multiresolution segmentation algorithm was utilized to extract objects from subspace feature images and support vector machines (SVM) classifier was then used to classify the object-based feature images, texture features derived from gray level co-occurrence matrix (GLCM) and wavelet filter at the pixel level of the feature images with the use of SVM classifier were used as benchmarks to evaluate the proposed algorithm. Classification results show that the proposed object-oriented nonlinear subspace analysis approach can give significantly higher accuracies than the traditional pixel-based and texture-based subspace classification.