Using an anisotropic diffusion scale-space for the detection and delineation of shacks in informal settlement imagery

• Main Author: Levitt, Stephen Phillip
• Format: Others
• Language: en
• Published: 2011
• Subjects: delineation; anisotropy; shacks; informal settlements
• Online Access: http://hdl.handle.net/10539/9633

PhD, Faculty of Engineering and the Built Environment, University of the Witwatersrand, 2010 === Informal settlements are a growing world-wide phenomenon. Up-to-date spatial information mapping settlements is essential for a variety of end-user applications from planning settlement upgrading to monitoring expansion and infill. One method of gathering this information is through the analysis of nadir-view aerial imagery and the automated or semi-automated extraction of individual shacks. The problem of shack detection and delineation in, particularly South African, informal settlements is a unique and difficult one. This is primarily due to the inhomogeneous appearance of shack roofs, which are constructed from a variety of disparate materials, and the density of shacks. Previous research has focused mostly on the use of height data in conjunction with optical images to perform automated or semi-automated shack extraction. In this thesis, a novel approach to automating shack extraction is presented and prototyped, in which the appearance of shack roofs is homogenised, facilitating their detection. The main features of this strategy are: construction of an anisotropic scale-space from a single source image and detection of hypotheses at multiple scales; simplification of hypotheses' boundaries through discrete curve evolution and regularisation of boundaries in accordance with an assumed shack model - a 4-6 sided, compact, rectilinear shape; selection of hypotheses competing across scales using fuzzy rules; grouping of hypotheses based on their support for one another, and localisation and re-regularisation of boundaries through the incorporation of image edges. The prototype's performance is evaluated in terms of standard metrics and is analysed for four different images, having three different sets of imaging conditions, and containing well over a hundred shacks. Detection rates in terms of building counts vary from 83% to 100% and, in terms of roof area coverage, from 55% to 84%. These results, each derived from a single source image, compare favourably with those of existing shack detection systems, especially automated ones which make use of richer source data. Integrating this scale-space approach with height data offers the promise of even better results.