Thesis Type: Postgraduate
Institution Of The Thesis: Middle East Technical University, Turkey
Approval Date: 2013
Thesis Language: English
Student: Okan Bilge Özdemir
Supervisor: YASEMİN ÇETİNAbstract:
Hyperspectral image processing is improved by the capabilities of multispectral image processing with high spectral resolution. In this thesis, we explored hyperspectral classification with Support Vector Machines (SVM), Maximum Likelihood (ML) and KNearest Neighborhood algorithms. We analyzed the effect of training data on classification accuracy. For this purpose, we implemented three different training data selection methods; first N sample selection, randomly N sample selection and uniformly N sample selection methods. We employed Principal Component Analysis (PCA) as preprocessing method and conducted experiments with different number of principal components for all three classification algorithms. As a post-processing method following pixelwise classification, filtering with 3x3 window and majority voting with meanshift segmentation methods are used to incorporate spatial information over spectral information. The experiments showed that without using pre-processing and post-processing SVM procures better classification accuracies than the other algorithms for all training data sizes. ML is inferior for lower number of training data samples but improves its performance with lower number of principal components. K-NN algorithm provides almost the same accuracies for more than 10 principal components. PCA usage does not improve SVM performance but decreases classification time for larger scenes. Filtering with 3x3 window method improves the classification accuracy by 4-5%. However, spatial information usage by employing majority voting with meanshift segmentation method performs better than filtering 3x3 window. Classification with both pre-processing and post-processing improves classification accuracy and decreases classification time. The largest improvement is for the ML method with lower number of training data.