Multi- und hyperspektrale Klassifikation von Feuchtgebieten mittels Ableitungsanalysen und spektralem Entmischen

Multi- and hyperspectral classification of wetlands using derivative analysis and Spectral Unmixing

Diplomarbeit von Thomas Galka
Studiengang: Kartographie und Geomedientechnik
SS 2006


With the goal of developing automated and standardized techniques to classify multispectral and hyperspectral remote sensing data, this thesis has tested the capabilities of derivative analysis and spectral unmixing algorithms. For that purpose two hyperspectral datasets and one multispectral dataset were used to classify wetlands by these methods. To ensure the transferability of classification parameters between the different datasets the data were transformed in objective, physical values. These were bottom type specific reflectance values and their derivatives. Using an inversion method the influence of different atmospheric parameters has been diminished and the gray values have been changed into bottom reflectance values. The analysis of derivative spectra was used to extract spectral details which were appropriate to separate the classes of the wetlands. These parameters were used to devise independent configuration files for a standardized classification of remote sensing data. Furthermore by an inversion method of spectral unmixing the cover amounts of three endmembers have been acquired. Thereby it has been investigated which minimal spectral resolution is necessary for derivative analysis and which modifications have to be done to translate hyperspectral information into multispectral data. The purpose of this classification method was to make the processing of remote sensing data independent from additional ground truth measurements. It should facilitate the creation of standardized remote sensing products using a standardized process chain and make possible to simplify a calculation of the costs. These researches have shown that there is only a finite transferability of classification parameters between different investigation areas and datasets. Thus it has appeared that the classification of main classes by devolved classification parameters is indeed possible, but not an accurate analysis of the occurred species. For an accurate classification the classification parameters had to be extracted in each hyperspectral dataset. A manual detection of those parameters in the multispectral dataset hasn't succeeded. Thus it requires a statistic tool which will find the best separabilities of the classes automatically. In contrast to this the transferability of the reference spectra for the spectral unmixing between all datasets could be demonstrated. The results which have been achieved using the developed classification method are in accordance with the actual facts and show many details of the covering of the research areas. The classification method which has been developed in this thesis is particularly suitable for an accurate classification of limited areas with a finite number of classes.

Verfahrenskette zur automatisierbaren Klassifikation