Vol. 12, No. 1, 67-81, 2013

Abstract
Land use demands of tourism and residential development drive the spread of leisure infrastructure and new urban landscapes in the Mediterranean. In particular along the coasts, golf courses, irrigated landscaping and swimming pools are becoming characteristic features of already densely populated locations that are among the areas with the greatest water deficits. Against this background, the present paper assesses the potential of different high-resolution imagery in combination with innovative image analysis techniques for an automated, targeted mapping of water-related urban features. The mapping task is conducted with WorldView-2, IKONOS and airborne imagery in three different urban study areas in Spain and Greece. Object-based feature extraction and the Random Forests algorithm are applied to the classification problem of separating turf grass, swimming pools, other vegetation and non-vegetated areas. The classifier performance is evaluated against susceptibility to reduced training set size and high-dimensional features spaces, variations in the training set (stability of results), varying feature subspaces, and the inclusion of uncertain - hence potentially mislabelled - pixels in the model calibration stage. The results indicate that best discrimination can be obtained if complementing the standard spectral feature space by object features from multi-scale segmentation. Furthermore, it is confirmed that Random Forests can handle high-dimensional feature spaces with large amounts of potentially redundant or irrelevant features. Comparing the results across image sources and study areas reveals different quality levels, but indicates that water-related urban landscape elements can be mapped, if the remote sensing imagery meets several requirements such as very high spatial resolution.

ISSN 1729-3782