Development of a fuzzy expert system for detailed land cover mapping in the Dra catchment (Morocco) using high resolution satellite images

Abstract

Land cover and the distribution of vegetation are important parameters in models describing the earth system. The work described in this thesis was performed in the framework of the IMPETUS project in order to deliver this information in a detailed manner for the river Dra catchment area in south Morocco. On the basis of a modified and regionally adapted CORINE land cover classification scheme a mapping was performed using LANDSAT ETM+ data. Intensive field campaigns were conducted to collect ground truth data and compile a validation dataset. <br /> In order to generate a land cover map with high thematic content in high confidence a non standard, knowledge based, expert classification system was implemented. This includes the utilisation and fuzzification of the outputs of two sub-pixel classifiers: linear spectral unmixing and the spectral angle mapper. <br /> The creation of the land cover map requires several data (pre-) processing steps. Emphasis was placed on the automation of high resolution satellite data processing. Cartographic base material from ASTER data was automatically generated by the use of digital photogrammetric methods. An automated ground control point generation program was incorporated for an image to image registration. An automatic relative radiometric image normalisation technique is introduced and utilised to generate a radiometrically correct LANDSAT ETM+ image mosaic of the catchment for May 2000. The expert system was transferred to LANDSAT TM data from April 1987 for full automatic change detection without loss of classification detail.<br /> Methods to treat obliquely recorded very high resolution satellite data, such as CORONA and IKONOS are discussed. NOAA/AVHRR data on a daily basis were processed automatically and compared to TERRA/MODIS data. Multispectral IKONOS data were used in a biophysical analysis and related to information from high resolution LANDSAT data. Methods for further improvement and the continuation of research are outlined and discussed.