A Generalised-Nearest-Neighbor (GNN) technique for the improved parametrization of spatialy distributed models using remote sensing data

One fundamental prerequisite for parameterizing spatially distributed hydrological models is the correct characterization of land surface properties. When using remote sensing techniques, it is often necessary to first classify areas (pixel) into a number of predefined soil or vegetation categories, before any quantitative measure can be derived. Also, uncertainties related to this process are rarely considered especially when appropriate ground information is sparse. Classifications methods commonly used in remote sensing (Nearest Neighbor, Maximum likelihood) apply distance measures that are adopted a priori. In contrast to these standard approaches we here propose a general procedure to find an adaptive metric that combines a local variance reducing technique and a linear embedding of the observation space into an appropriate Euclidean space. Besides significant improvements with regard to the overall accuracy of the method and its robustness when the cardinality of the calibration data set is low, the method also allows to explore the uncertainties of the classification by an ambiguity measure. Its spatial distribution helps to identify critical areas/classes or data limitation within the classification process, therefore reducing the potential costs. Several land cover classifications exemples from Germany using Landsat scenes are presented to illustrate this method. We will also demonstrate how the principal idea of this method can be extended to a variety of other application in spatial hydrology, especially where systems are dominated by non-linear behavior.