[Zurück]

M. Gaina:
"Classification of satellite images by including spectral and textural information";
Betreuer/in(nen): J. Jansa, C. Ressl; Department für Geodäsie und Geoinformation, 2019; Abschlussprüfung: 19.06.2019.

With the advance in sensor technology in the field of remote sensing from space, new challenges emerge. The high-resolution images offer a wide range of new applications, but at the same pace, the interpretation requires new approaches, from pure spectral interpretation to a more holistic one. This thesis focuses only on a small aspect of that sort of interpretation and on one specific application, which has been gaining increasing importance nowadays, where the carbon dioxide balance has become an issue. Forests are important CO2 sinks, and therefore, it makes sense to concentrate on the interpretation of forest stands, in this case in the area of central Europe. Thus, the principal objective of this investigation is to focus on forest classification and to interpret different types of forest stands in high-resolution satellite images. The used images have been captured by Pléiades 1B satellites, whose spatial resolution provides quite good textural information, which may be utilized to distinguish between different types of forest patches. Together with the spectral information, one may expect even an improvement of the classification quality compared to the interpretation of sole multispectral object properties. Therefore, this research concentrates on assessing standard strategies for image classification if the spectral and textural information is to be taken into consideration. The key for characterizing texture in forest areas was found in using a set of Haralick textural features known already for many decades, therefore for the special purpose a thorough investigation of generating suitable textural features has been carried out and their properties have been studied. One of the standard classification algorithms in remote sensing is the Maximum Likelihood classification. The question arises of course, whether the Maximum Likelihood classification would be appropriate enough for the textural classification. Therefore, the distribution of the classes in the feature space for the textural parameters has been investigated and then the decision has been made to use the Maximum Likelihood for classifying the multispectral as well as for textural parameters, from which Mean, Contrast, and Entropy delivered promising results, which have proved of value in previous research with other satellite data. Further, the quality assessment of the data has been made, where the resulted accuracies are quite high, around 80%, and lie in the expected range, although a significant improvement by including textural features cannot be observed. In the frame of these investigations, commercial software (ENVI Image Analysis (the Environment for Visualizing Images)), open source products, and a few other minor tools have been used for visualization, analysis, and processing, besides own software developments. There are still a few open issues for future work, whose investigation would have exceeded the effort for a diploma thesis, in particular, the influence of combining various Haralick features and of varying the parameters for their generation.

https://repositum.tuwien.ac.at/urn:nbn:at:at-ubtuw:1-124865