Theses
Entwicklung eines Verfahrens zur Klassifikation heterogener Räume auf der Basis von hyperspektralen Flugzeugdaten
Art der Abschlussarbeit
Diplomarbeit
Autoren
- Mende, André
Betreuer
- Prof. Dr.phil.habil. Manfred Buchroithner
Weitere Betreuer
Deutsches Zentrum für Luft- und Raumfahrt Oberpfaffenhofen
Abstract
The main aim of this study is the development of a spectral library comparison classifier, which is based on a comprehensive spectral library. This classifier will be able to classify airborne hyperspectral images without prior knowledge about the test site. The concept-shows three different classification criteria, which are based on the spectral similarity analysis of the SID-SAM mixed measure. This similarity analysis compares all pixel spectra-to all reference spectra of the spectral library. An analysis of the best ten scores of each pix-el is the basis for the assignment of the class relationship for each classification criterion. Additionally, for the evaluation of the classification results five different reliability meas-ures are developed. The chosen test sites Berlin, Dresden, Zwickau, Mende and Fos-sur-Mer represent different spatial characteristics. This study shows that the classifier is able to deal with the incompleteness of the spectral library by using the delivered mask of poorly classified pixels in combination with the developed concept of the spatial and spectral cluster analysis. The results of this analysis return potential new training spectra for the completion and improvement of the spectral library. rnThe spectral library is based on a consistent and well-structured class hierarchy. There-fore, the compatibility to other classification approaches is provided. The classification re-sults of different test sites are comparable due to the common spectral library. Additionally, the classification results can be included to in an ensemble classifier to reduce the disadvan-tages and to improve the potentials of the combined classifier. The classification results are independent of calibration parameters of different test sites and dates. The calculation of the validation approaches for the test site Dresden returns the best reliability measure for the classification criterion of the best albedo match by an overall accuracy of 89.8 % and a kappa-coefficient of 0.888. Therefore the classification accuracy is very good for the con-sidered test site. rnThe best classification reliability exists for material classes, which include large spectral contrasts to other material classes. Therefore the spectral library comparison classifier is able to classify both natural and anthropogenic material surfaces in heterogeneous en-vironments without any prior knowledge about the test site. Additionally, the comparison of the validation results of the comprehensive spectral library and a reduced spectral library shows another important conclusion: Using a low number of representative reference spectra, all material classes with distinctive spectral characteristics will be classified very well. The only problem with using this reduced spectral library is the bad differentiation of material classes with similar spectral characteristics.rn
Url/Urn/Doi
kartographie.geo.tu-dresden.de
Zugeordnete Forschungsschwerpunkte
- Fernerkundungskartographie
Schlagwörter
spectral library comparison classifier
Berichtsjahr
2011