A structural pattern recognition paradigm and system to infer information on urban land use from fine scale earth observation data

• Main Author: Barr, S. L.
• Published: Swansea University 2002
• Subjects: 526.9
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.636044

Accurate and objective fine-scale information on urban land use is important for the development of cohesive urban planning policies. Earth observation images represent one source of such information although historically the results of automated urban land use mapping have been poor. It is argued that this due to the widespread use of spectrally-based pattern recognition approaches that assume a direct relationship between detected spectral response and urban land use. Evidence is presented that shows that detected spectral response of urban scenes is only indirectly related to land use. In response a <i>function-from-form</i> approach is developed to infer urban land use in two-stages: (i) the derivation of land cover from a multispectral image data set, and (ii) the inference of land use on the basis of the structure of the land cover. A novel structural pattern recognition system called SAMS (Structural Analytical Modelling System), consisting of XRAG (eXtended Relational Attribute Graph) - a data model that represents the structure of image regions - is developed to achieve this two-stage inference. SAMS is used in several experiments that reveal building proximity and morphology, as opposed to the topological spatial organisation of urban land cover, facilitates the statistical discrimination of different urban land use types. Further analysis reveals, however, that the spectral inference of urban land cover from multispectral images results in very complex apparent scene structure. In response to this, an automated approach is developed to reduce the complexity of scene region-structure. This procedure is, however, unable to achieve a building region-structure that allows unambiguous urban land use discrimination. Nevertheless, the results presented in this thesis demonstrate that it is possible to distinguish urban land use types using structural pattern recognition approaches to analyse urban land cover regions, although the accurate inference of urban land cover remains a significant barrier to its application.