Sub-pixel analysis in combination with knowledge based decision rules to optimise a land cover classification

摘要

Land cover mapping is one of the earliest tasks of satellite data interpretation. Best classification results are usually achieved with supervised classification methods. Commonly used supervised classification methods as the Maximum Likelihood classifier are not appropriate for a detailed differentiation of certain land cover classes such as vegetation associations, The optimisation of land cover class identification by the use of a sub-pixel classifier is presented in this paper. Sub-pixel classifiers based on statistical rules use more of the embodied information in satellite data than pure pixel values. The concept of Linear Spectral Unmixing (LSU) involves the multispectral image information to derive abundance channels which can be more physically interpreted. Reference spectra of certain surface components serve to describe the spectral surface variability and to calculate abundance channels. This abundance information of the reference spectra is in a second step used to set up decision rules for land cover class discrimination; which depends on terrain knowledge. Auxiliary data, such as a digital elevation model or hydrological maps can be incorporated to further discriminate land cover classes applying additional rules. An example of the Basin of Tazenakht within the Dra catchment in South Morocco will show that a detailed vegetation classification can be achieved by using LSU in combination with knowledge based decision rules. The overall classification accuracy at 96% is high.