ANALYSIS OF MULTITEMPORAL AND MULTISENSOR REMOTE SENSING DATA FOR CROP ROTATION MAPPING

Abstract. For accurate regional modelling of (agro-)ecosystems, up-to-date land use information is essential to assess the impact of the permanent changing vegetation cover of agricultural land on matter fluxes in the soil-vegetation-atmosphere (SVA) system. In this regard, officially available land use datasets are mostly inadequate, since they only provide generalised information concerning agricultural land use. In this contribution, we present our work for the year 2008 on the generation of multi temporal, disaggregated land use data with the goal to derive a crop rotation map for the years 2008–2010 for the study area of the research project CRC/TR 32. For this purpose, the Multi-Data Approach (MDA) was used to integrate multitemporal remote sensing classifications with additional spatial information by the means of expert knowledge-based production rules. Our results show that the information content of a land use dataset is considerably enhanced by combining crop type information of multiple observations during each growing season. For a sufficient temporal coverage, the usage of multiple sensors is generally inevitable. Thus, datasets of ASTER, Landsat TM & ETM+ as well as IRS-P6 were incorporated. In terms of classification accuracy our analysis yielded similar results with support vector machines (SVM) and the classical maximum likelihood classifier (MLC) for all sensors, with SVM being mostly only slightly better. For the refinement of land parcel boundaries and the reduction of misclassification, the incorporation of the 'field block' (FB) vector information was very effective. 'Field blocks', provided by the chamber of agriculture, are coherent agricultural areas with (relatively) permanent boundaries. As a result, a much more accurate differentiation of agricultural land and non-agricultural land was achieved. With the enhanced annual MDA land use data of the three consecutive years containing crop type information sufficient information is available for the derivation of crop rotation. Again, adapted knowledge-based production rules are used for this purpose.