Identifying residential neighbourhood types from settlement points in a machine learning approach

Warren C Jochem,Tomas J Bird,Andrew J Tatem

doi:10.1016/j.compenvurbsys.2018.01.004

Warren C Jochem, Tomas J Bird + Show 1 more

Open Access

https://doi.org/10.1016/j.compenvurbsys.2018.01.004

Copy DOI

Abstract

Remote sensing techniques are now commonly applied to map and monitor urban land uses to measure growth and to assist with development and planning. Recent work in this area has highlighted the use of textures and other spatial features that can be measured in very high spatial resolution imagery. Far less attention has been given to using geospatial vector data (i.e. points, lines, polygons) to map land uses. This paper presents an approach to distinguish residential settlement types (regular vs. irregular) using an existing database of settlement points locating structures. Nine data features describing the density, distance, angles, and spacing of the settlement points are calculated at multiple spatial scales. These data are analysed alone and with five common remote sensing measures on elevation, slope, vegetation, and nighttime lights in a supervised machine learning approach to classify land use areas. The method was tested in seven provinces of Afghanistan (Balkh, Helmand, Herat, Kabul, Kandahar, Kunduz, Nangarhar). Overall accuracy ranged from 78% in Kandahar to 90% in Nangarhar. This research demonstrates the potential to accurately map land uses from even the simplest representation of structures.

Highlights

As populations around the world become more urbanised, in developing countries, the ability to quantify and study the growth and changing function of cities in detail has become more important for urban growth, informal settlements, poverty, environmental and health concerns (Duque, Patino, Ruiz, & Pardo-Pascual, 2015; Herold, Liu, & Clarke, 2003; Kuffer, Pfeffer, & Sliuzas, 2016; Kuffer, Pfeiffer, Sliuzas, & Baud, 2016; UN Habitat, 2016)
We describe several metrics calculated from the spatial point patterns of settlement points which are used to characterise the density and distribution of settlements
We report the cross-tabulation of pixel-level predicted vs. State of Afghan Cities (SoAC) residential types as well as positive and negative predictive values, sensitivity, specificity and overall accuracy measures for the predictions in each province

Summary

Introduction

As populations around the world become more urbanised, in developing countries, the ability to quantify and study the growth and changing function of cities in detail has become more important for urban growth, informal settlements, poverty, environmental and health concerns (Duque, Patino, Ruiz, & Pardo-Pascual, 2015; Herold, Liu, & Clarke, 2003; Kuffer, Pfeffer, & Sliuzas, 2016; Kuffer, Pfeiffer, Sliuzas, & Baud, 2016; UN Habitat, 2016). The Sustainable Development Goals (United Nations, 2014) and the New Urban Agenda (United Nations, 2017) have brought additional focus for policymakers on land use planning to create resilient, sustainable, and inclusive cities. To meet such goals, data on intra-urban differences in land uses is needed. Similar analyses using large collections of geospatial vector data (points, lines, polygons) have received far less attention in the literature than remote sensing approaches, though several studies have noted the potential to identify classes of buildings or urban land uses (Barr, Barnsley, & Steel, 2004; Hecht, Meinel, & Buchroithner, 2015; Longley & Mesev, 2000; Steiniger, Lange, Burghardt, & Weibel, 2008). Overall the results suggest that our method has potential to extract meaningful information from even the simplest geometric representation of structures

Objectives

Methods

Results

Conclusion