Automated Detection of Canola/Rapeseed Cultivation from Space: Application of new Algorithms for the Identi cation of Agricultural Plants with Multispectral Satellite Data on the Example of Canola Cultivation

The dispersal of new genes resulting from the cultivation of genetically modified plants holds risks that are difficult to assess. In this context the situation of cultivation is of particular interest since fields are potential sources of the transfer of new genes to non-modified or related plants. The aim of this work is the identification of canola cultivation areas in northern Germany in the studied period from 1995 to 2002. The sizes of the fields and the investigation area pose requirements on the satellite data best met the LANDSAT Thematic Mapper and Enhanced Thematic Mapper and the Indian Remote Sensing Satellite Linear Imaging Scanning Spectrometer/3.The first processing step, the georectification is done by a passpoint correlation which is improved by an additional correction step, based on the correlation of image clips.The next processing step is the identification of clouds and their shadows. Opaque clouds can be identified by their brightness and low top temperature. Thin clouds are identified based on the Haze Optimized Transform method. The third processing step, the classification, is performed by the Mahalanobis Distance Clasifier (MDC) because it only requires training data for one single surface type. The accuracy of the MDC is enhanced by a segmentation of the MDC result used to identify single wrongly identified pixels and to perform region growing to include pixels missed by the MDC.The results are approximated by rectangles of equal orientation and area which allows a simple evaluation of the field distances and other parameters of interest. The results are used to produce statistics to investigate these parameters for the cultivation of canola in northern Germany. The results of the classification are compared to validation data, i.e., edges and positions of known canola fields and agricultural statistics for 1995 and 1999. This validation showed that the total acreage of canola is identified with 70 to 90% accuracy.