Generation and Assessment of Urban Land Cover Maps Using High-Resolution Multispectral Aerial Cameras

Abstract—New aerial cameras and new advanced geo-processing tools improve the generation of urban land cover maps. Elevations can be derived from stereo pairs with high density, positional accuracy, and efficiency. The combination of multispectral high-resolution imagery and high-density elevations enable a unique method for the automatic generation of urban land cover maps. In the present paper, imagery of a new medium-format aerial camera and advanced geoprocessing software are applied to derive normalized digital surface models and vegetation maps. These two intermediate products then become input to a tree structured classifier, which automatically derives land cover maps in 2D or 3D. We investigate the thematic accuracy of the produced land cover map by a class-wise stratified design and provide a method for deriving necessary sample sizes. Corresponding survey adjusted accuracy measures and their associated confidence intervals are used to adequately reflect uncertainty in the assessment based on the chosen sample size. Proof of concept for the method is given for an urban area in Switzerland. Here, the produced land cover map with six classes (building, wall and carport, road and parking lot, hedge and bush, grass) has an overall accuracy of 86 % (95% confidence interval: 83-88%) and a kappa coefficient of 0.82 (95 % confidence interval: 0.78-0.85). The classification of buildings is correct with 99 % and of road and parking lot with 95%. To possibly improve the classification further, classification tree learning based on recursive partitioning is investigated. We conclude that the open source software “R” provides all the tools needed for performing statistical prudent classification and accuracy evaluations of urban land cover maps. Keywords-land cover map; classification; assessment; thematic accuracy; multispectral camera; map revision I