Intrinsic Dimensionality Estimation and Dimension Reduction of Hyperspectral Data for Classification Using Decision Tree C4.5, SVM and ANN

Classification is one of the prominent issue in processing of Hyperspectral images because of their variety applications such as urban planning, mining and military. Hyperspectral imaging due to high spectral resolution can provide much more information about the surveyed area. Ability of classifying objects in the scene can raise by using their spectral properties. However, the classification of Hyperspectral image is a very challenging issue because of their huge number of band. The large number of spectral bands in Hyperspectral data not only increases the computational burden but also can reduce classification accuracy. Moreover, there are usually high correlation between bands. Dimensionality reduction has become a crucial step for successfully classify Hyperspectral image and getting better results. Feature extraction by using the intrinsic dimensionality of the data is one useful approach to decrease the dimensionality. In this paper, Principle component analysis method using intrinsic dimension used as a pre-processing step to improve classification result. By using intrinsic dimension Estimation, can recognize proper bands number for processing and by PCA can extract optimal bands to perform classifications SVM, ANN and C4.5. The results reveal the implemented method can improve the overall accuracy of classified imagery nearly 5-10 %. ANN method with accuracy about 0.95 have highest accuracy among all methods.