Vol. 4, No. 1, 94-105, 2005

Abstract
The strong impact of land use on the climatic and hydrological cycle necessitates a detailed knowledge of the land use/land cover for land surface modelling. Within the research project GLOWA-Danube a knowledge-based classification method is applied on optical remote sensing data. Two important facts have to be addressed in the classification procedure: The large size as well as the heterogeneity of the test site.

Seven LANDSAT TM and ETM+ images were acquired covering the entire catchment. Due to the strong influence of topography, all scenes were georeferenced to the DTM and corrected for atmospherical and illumination effects. A multisource and multistage classification procedure was developed, including a GIS knowledge base consisting of elevation and slope data, soil information and the mean annual precipitation distribution. Water, snow and clouds were masked separately using thresholds. Also the separate retrieval of settlements represents a further improvement to the methodology. This class often showed overlapping with open or sparsely vegetated soil, harvested grassland and mature cereals. The best results were obtained using the backscatter of a RADARSAT image, the NDVI and principal components derived from the LANDSAT images. The core algorithm is an enhanced knowledge-based fuzzy logic classifier. It includes, beside the spectral information, a physiogeographical knowledge base to assign a pixel to a class. The finally resulting land use map of the Upper Danube catchment consists of 27 classes. A critical point is the accuracy assessment. Although a comparison with census data on a community base gives a good estimation, pixel-bx-pixel or field-based accuracy assessments have to be conducted as well. But this can only be done for areas where reliable ground truth data are available.

ISSN 1729-3782