Thesis (Ph.D.) - Cairo University - Faculty of Computers and Information - Department of Information Systems

Hyperspectral images (HSI) are used as a source of information to detect and discriminate the differences between land scales on Earth. In addition, HSI is used to study the changes that occur on these areas. Many unsupervised learning algorithms have been proposed to analyze and interpret the contents of Hyperspectral Images. However, these algorithms are tested against images with ground truth data or benchmark. For clustering algorithms, the number of clusters were identified according to prior knowledge known of such terrains. In practical, such knowledge are unknown. Identifying the correct number of clusters may differ from one algorithm to another. This study introduces a novel methodology to propose the correct number of clusters for clustering algorithms. This is done using by training a Support Vector Machine (SVM) classifier on the spectral information of a benchmark data. The spectral information is binary labeled either by correctly or incorrectly clustered. The purpose of the SVM classifier is to learn where the clustering algorithm can correctly cluster a pixel in the spectral space. Next, the clustering is applied on HSI with no ground truth. SVM classifier is used to validate the performance of the applied unsupervised learning algorithm applied to indicate the appropriate number of clusters. External cluster evaluation measures, Normalized Mutual Information of Purity Index, are used to accomplish this task

Issued also as CD