Classification using hyper-spectral imagery

Abstract

The subject of this dissertation is the study of concepts and principles of Remote Sensing focusing on the analysis and processing of Hyperspectral Images. Hyperspectral imaging as a tool for monitoring, detection and classification of objects and areas on Earth can be used in a wide range of applications. Especially in the case of monitoring environmental and climatic changes, hyperspectral imaging can provide plenty of information in order to create an environmental database develop models and prevent possible natural disasters. Based on this framework and using Matlab software tool a processing system was proposed and developed. The procedure of processing raw hyperspectral data is described and a classification system is developed. This system uses classification algorithms in order to distinguish areas with vegetation, water resources, dry soil or man-made constructions using hyperspectral images from a remote sensor. Four supervised classification algorithms were developed and evaluated: Euclidean Minimum Distance, Mahalanobis Minimum Distance, Bayesian Classification and Maximum Likelihood approximation. The algorithms were applied on a satellite hyperspectral image of an area. Results have shown that the fastest algorithm is the Euclidean distance giving relatively accurate results. Weaknesses in the performance of the Mahalanobis distance algorithm were reveled in case of insufficient training sets. Bayesian Classification algorithm performed exceptionally well. However the best performance as far as accuracy is concerned were provided by the proposed approximation of the Maximum Likelihood Distance Algorithm, which also reaches satisfactory performance without the need of a long iterative process. During the study the spectral signatures for the defined classes were estimated. Comparison with existing studies has shown that the estimates are representative of the area under investigation confirming the integrity of the proposed procedure. The dissertation is partitioned in two sections. Chapters 2, 3, 4 and 5 provide the necessary background and state of the art in the subjects of remote sensing and Hyperspectral Imaging, while chapters 6, 7 and 8 present the classification algorithms, the implemented processing system, the results and conclusions. The Matlab code with necessary clarifications and explanations is also provided in the Appendix.