Settlement classification from multi-scale spatial patterns of building footprints in a machine learning approach

Remote sensing techniques are commonly applied to map and monitor land uses, to measure the growth of cities, and to assist with urban planning. Increasing availability of very high spatial resolution (VHR) imagery (< 1 m) and computing power is now enabling settlement data in the form of building footprints to be extracted from imagery for whole countries. These settlement data provide information on city growth and urbanisation, but, while spatially detailed, extracted building footprints typically lack other attribute information that could identify building types or be used to differentiate intra-urban areas, thus limiting their potential uses. This presentation will discuss an approach to classifying settlement types from spatial patterns of urban morphology visible in building footprints extracted from VHR imagery. Settlement classifications can be used as part of geodemographic analyses, for urban planning studies, or integrated with other statistical measures. To create the classifications, data features describing the size, shape, angle, and density of structures are calculated at multiple spatial scales. Using supervised and unsupervised machine learning methods (Random Forests and Gaussian mixture models) in a high performance computing environment, these data are used to classify settlement types across a 100 m spatial resolution grid for large study regions. We present classification results from three areas in Africa (Kaduna, Nigeria; Kinshasa, DRC; and Maputo, Mozambique) and discuss the potential for applying this approach in other contexts. Overall this work provides an example of using computational methods to extract information from big geospatial datasets and gain new insights.