Klassifikation einer Zeitserie von AWiFS-Multispektraldaten

Abstract

In regard of the EU project TIRAMISU different LULC classification methods are tested and compared. Additionally, their implementation in the DLR internal image processing program is checked for proper operation. This is done by use of a time series of 14 IRS-P6 AWiFS images from 2011 of an area in southern Germany. After analysing, correcting and bypassing errors in the data basis, the k-means, maximum-likelihood and fuzzy classification are applied. Additionally a hierarchical classifier is developed. In consideration of lack of quality in the data basis, all the methods deliver good overall accuracies in between 77% and 81%. The self-developed decision tree achieves the best result. With respect to TIRAMISU it has to be noted that for an area with the natural landscape present in southern Germany (small plots of surface types), a geometric resolution of 60 meters is too low. On the other hand, the use of a time series turns out to be useful due to a high reliability. Identified objectives for improvements as well as errors of the image processing program are documented. Which classification methods are supposed to be used for the TIRAMISU project cannot be determined until the data basis and its characteristics are defined. To achieve the required high reliability, multiple methods are recommended to be applied in parallel.