Reducing structural clutter in land cover classifications of high spatial resolution remotely-sensed images for urban land use mapping

Abstract

A new generation of very high spatial resolution (1–5 m) satellite sensors is due to be launched within the next five years. Among other things, images acquired by these sensors offer considerable potential for the derivation of information on urban land use. It has been suggested that this can be achieved via a two-stage process involving (i) a standard (per-pixel) multispectral classification algorithm to identify the principal land cover parcels present in the observed scene and (ii) the application of structural pattern-recognition techniques to infer land use from the morphological properties of these parcels and the spatial relations that exist between them. It is implicit in this approach that the initial classification is of sufficient accuracy to allow land use to be inferred from these structural properties and relations. This assumption is investigated using airborne multispectral image data with a nominal spatial resolution of 2 m. It is shown that these data allow many features of interest in urban areas to be identified and delineated, but that they contain a significant amount of unwanted spatial detail (or ‘scene noise’). The latter results in structural ‘clutter’ in the initial land cover classification, which limits the potential to infer land use in the second stage of the data-processing chain. To address this issue, a simple, region-based, reflexive-mapping procedure is developed. This operates at the parcel (cf. pixel) level. The procedure is successful at removing much of the structural clutter, and performs well in comparison with traditional majority filtering; however, the inference of urban land use from the resulting data remains problematic.