UNIGIS Abschlussarbeiten

Der krönende Abschluss eines UNIGIS MSc Studiums ist sicherlich die Master Thesis. Mit ihr belegen unsere MSc-AbsolventInnen, dass sie den akademischen Grad "Master of Science (Geographical Information Science & Systems)" zu Recht führen.  Im UNIGIS professional Studiengang muss keine Abschlussarbeit verfasst werden. Dennoch nehmen einige Studierende die Möglichkeit war, ein Geoinformatikprojekt durchzuführen und entsprechend zu dokumentieren.

Sie sind auf der Suche nach aktueller Literatur zu Geoinformatik-Themen?
Hier finden sie die mitunter preisgekrönten Abschlussarbeiten unserer AbsolventInnen!

Roland Graf [01-2012]:

Objektklassifizierung mit Support Vector Machines

Diese Arbeit ist online verfügbar: Download

This Master Thesis introduces the theoretical background of Support Vector Machines (SVMs) and discusses their usage in object-based land use and land cover classification systems. The chapters at the beginning of the Master Thesis describe the foundation of image processing, the underlying concepts of supervised learning methods and define the notation of feature vectors for data representation in classification problems. After outlining the numerical classification process and basics of linear discriminant function analysis, the thesis reviews the mathematical concept of Linear Support Vector Machines and Soft Margin methods to find separating hyperplanes for feature vectors. In a subsequent step, the model will be extended for nonlinear classification problems. Kernel methods will be introduced to map the data from the original input space into a kernel feature space of higher dimensionality. For a series of experiments with SVMs, a Definiens eCognition SVM Plug-In and multiple MATLAB script files have been implemented. The second part documents some implementation details und evaluates the applicability of Support Vector Machines in object-based image classification systems. These chapters also describe the determination of preconditions and parameter settings for a successful SVM usage. Another goal is the formulation of recommendations for optimizing both, the data and the SVM model, to fine tune the classification results and to simplify and speed-up the classification process. The quality of classification and the effect of different parameter settings are evaluated with pre-classified data by graphical and numerical comparisons. Contour plots of cross validation accuracies for selected parameters settings and kernels as well as confusion matrices and processing time comparisons are analyzed to evaluate the discriminatory power of Kernel-SVMs and their practical usability. Finally, optional optimization efforts (feature transformation, feature selection and training data reduction) are discussed and some examples show whether the changes affect the results.

Zurck weitere Arbeiten ...