Seminar Ozean, Eis, Atmosphäre

Remote sensing data are used for many agricultural and ecological applications, which cover a wide range of investigation areas. Detection of land use classification, ecology monitoring and yield prediction in agriculture represent only a few purposes taking benefit from the analysis of satellite or airborne remote sensing images. In my presentation I will concentrate on a method for the use of satellite data for the detection of oilseed rape acreage, which is the basis for a study of the potential spread of genetically modified rape in northern Germany. This study is part of the GenEERA project is a cooperative project of several partners all over northern Germany (for more information see: http://www.uft.uni-bremen.de/risk/projekte/geneera/geneera.html).
Classification algorithms like maximum likelihood or minimum distance are commonly used for the analysis of spatial high resolved multi spectral satellite images ( e.g. LANDSAT, SPOT, IRS-1C/D ). But in most of the cases sufficient classification results can be achieved if they are applied for a single image only, because the reflectance of identical areas are changing from image to image due to crop growing and atmospheric influences. In the GenEERA project these techniques are not applicable, because the total coverage of the investigated area (northern Germany) requires a minimum of 6 scenes per year and because of crop rotation the GenEERA project requires to monitor several years. In total 45 LANDSAT images for five years have to be processed and this has to be done automatically. The best time for rape detection is the full blooming because of the acreage yellow color, but the period of full blooming is quite short (only two weeks). Due to the weather conditions in northern Germany, it is unlikely to obtain all satellite images which are needed within this period. This small period for rape detection can be extended up to six weeks if a combination of Red and Near Infrared is used instead of the yellow color. For a further increase of this period, it is possible to take the state of plant development into account. This can be done by comparing the local blooming dates and the NDVI, the so called Normalized Difference Vegetation Index as indicator for the crop growth state. The first classification is made by a pixel based threshold method. To improve the classification an image segmentation is used to identify single fields. With the average spectral information for all segment pixels, the algorithm becomes less vulnerable to inhomogeneities within a field and single false classified pixels can be secluded. In order to improve the spatial accuracy, this segments can be used for a soft classification algorithm for the bonder pixels of a segment. This can be done by using the spectral information for neighboring segments to get a mixing ratio for the border pixels.