Elsevier

Applied Geography

Volume 41, July 2013, Pages 36-45
Applied Geography

Human population distribution modelling at regional level using very high resolution satellite imagery

https://doi.org/10.1016/j.apgeog.2013.03.002Get rights and content

Highlights

  • OBIA of QuickBird image useful for regional human population distribution modelling.

  • Distinct frequency distribution of houses for 5 factors driving settlement patterns.

  • In addition, removal of artefacts at admin boundaries with pycnophylactic approach.

  • Results show strong impact of river proximity and slope on population distribution.

Abstract

Modelling the distribution of human population based on satellite-derived information has become an important field of research, providing valuable input e.g. for human impact assessments related to the management of threatened ecosystems. However, few regional-scale studies have been conducted in developing countries, where detailed land cover data is usually absent, and the potential of very high resolution (VHR) satellite imagery in this context has not been explored yet. This study uses results obtained through object-based image analysis (OBIA) of QuickBird imagery for a subset of a highly populated rural area in western Kenya. Functions are established that approximate frequency distributions of QuickBird-derived locations of houses in relation to five factors. These factors are known to impact settlement patterns and data is available for the entire study area. Based on an overall probability coefficient (weight) calculated from the single functions, human population is redistributed at the smallest administrative level available (version A). In addition, the problem of artefacts remaining at administrative boundaries is addressed by combining the approach with the pycnophylactic smoothing algorithm (Tobler, 1979) (version B). The results show distinct patterns of population distribution, with particular influence of rivers/streams and slope, while version B in addition is free of boundary artefacts. Despite some limitations compared to models based on detailed land cover data (e.g. the ability of capturing abrupt changes in population density), a visual and numerical evaluation of the results shows that using houses as classified from VHR imagery for a study area subset works well for redistributing human population at the regional level. This approach might be suitable to be applied also in other regions of e.g. sub-Saharan Africa.

Introduction

Reliable information on the spatial distribution of population has been identified as key information for a better understanding of human impacts on land and water resources (Tian, Yue, Zhu, & Clinton, 2005). Traditionally, information on population numbers and demographic characteristics stem from census data made available per administrative unit (Liu, Clarke, & Herold, 2006). However, population is unlikely to be distributed uniformly within these conceptual census boundaries (Hay, Noor, Nelson, & Tatem, 2005; Langford, Higgs, Radcliffe, & White, 2008). Thus, the spatial pattern as provided by choropleth mapping does not reflect the actual population distribution (Mennis, 2003). To address this problem, areal interpolation methods have been developed that transform population data into a regular, continuous grid (raster) in which each grid cell (pixel) is assigned a representative estimate of the total population for that particular cell (Yue et al., 2005). As human population distribution is closely linked to various exogenous factors (Briggs, Gulliver, Fecht, & Vienneau, 2007), several sophisticated techniques, called intelligent or smart interpolation, have been introduced (Liu, Kyriakidis, & Goodchild, 2008). These GIS-based population distribution models disaggregate population of census units based on spatially explicit ancillary information of which land cover data is often used as main explanatory variable (e.g. Gallego, 2010; Tian et al., 2005).

To derive the ancillary information necessary for estimating population quantities, the use of satellite imagery has become commonplace. While some studies simply use land cover information derived from satellite imagery (e.g. Linard, Alegana, Noor, Snow, & Tatem, 2010; Viel & Tran, 2009), Li and Weng (2005) employ a stepwise regression analysis to sequentially test potential candidate variables from Landsat imagery such as vegetation indices or specific bands. Other techniques have been developed that make use of night-time light imagery (Briggs et al., 2007; Dobson, Bright, Coleman, Durfee, & Worley, 2000; Sutton, Roberts, Elvidge, & Baugh, 2001; Zhuo et al., 2009) or image texture indices (Chen, 2002). However, although ascribed high potential in the context of population distribution modelling (Hay et al., 2005), the opportunities of very high resolution (VHR) satellite imagery have not yet been fully explored (Liu et al., 2008).

A common technique of incorporating ancillary information into population distribution modelling is dasymetric mapping (e.g. Eicher & Brewer, 2001). By introducing empirical sampling, Mennis (2003) improved the determination of percentage assignment values for different land cover classes. Other techniques, sometimes referred to as intelligent or smart interpolation, use multiple other factors instead of or additional to land cover data, such as terrain or road network, and assign coefficients according to their influence on population. From the single coefficients an overall weight is calculated for each grid cell and then used for population redistribution (e.g. Dobson et al., 2000; Tian et al., 2005). However, the assignment of coefficients is mostly based on heuristic rules and assumptions without a solid evidence-base for such rules (Hay et al., 2005). Recently, some studies have attempted to improve this limitation, e.g. Briggs et al. (2007) by using satellite-derived light emissions data, or Liao, Wang, Meng, and Li (2010) by integrating generic programming and generic algorithm techniques. Still, VHR satellite imagery has been rarely employed for deriving coefficients that describe the relationship between population and various ancillary information sources.

The vast majority of studies on population distribution modelling published in the past 20 years were conducted in Europe (e.g. Gallego, Batista, Rocha, & Mubareka, 2011) and China (e.g. Tian et al., 2005; Zhuo et al., 2009) or have a global coverage such as LandScan (Dobson et al., 2000) or Gridded Population of the World (GPW) (CIESIN, 2000). The outcomes of these global studies are increasingly applied also in the developing world, e.g. for assessing conflict risk under climate change in East Africa (O'Loughlin et al., 2012), or modelling malnutrition in Mali (Jankowska, Lopez-Carr, Funk, Husak, & Chafe, 2012). However, regional scale models from the developing world, particularly from rural areas that take into account specific drivers of settlement patterns, such as rivers for water supply, are scarce. The recent initiative AfriPop addresses this need by improving existing global products for Africa (Linard et al., 2010; Tatem & Linard, 2011), but their country-scale modelling might still not be sufficient for regional-scale planning and research purposes.

This paper explores the value of classification results obtained through object-based image analysis (OBIA) of VHR satellite imagery for implementing GIS-based regional population distribution models. The first objective is to use the locations of houses, as classified from QuickBird imagery for a subset of a larger study area in rural western Kenya, to establish functions that approximate the frequency distribution of the houses in relation to selected population impacting factors. The second objective is to redistribute population for the entire study area, preserving population figures at the level of the smallest available administrative unit, based on an overall weight calculated from the single frequency distribution functions. Finally, in order to address the problem of artefacts at administrative boundaries that has not been solved by existing approaches using proximity factors (e.g. Yue et al., 2005), the third objective is to combine our approach with a mass-preserving pycnophylactic smoothing algorithm (Tobler, 1979) and to evaluate the resulting patterns.

Section snippets

Study area

The study area is located in western Kenya between 34°37΄05″ and 35°09΄25″ E and 0°02΄53″ S and 0°32΄23″ N and encompasses an area of 60 km by 65 km around the Kakamega-Nandi forest complex (see Fig. 1). While the forests are considered the easternmost relict of the Guineo-Congolian rainforest belt (Wagner, Köhler, Schmitz, & Böhme, 2008) known for their unique biodiversity and species composition (e.g. Althof, 2005; Mitchell, Schaab, & Wägele, 2009), their surrounding areas are densely

Redistributed population density

In order to allow discerning the modelling results with sufficient spatial detail, for this paper two subsets (see Fig. 1 for their location) are visualised instead of the entire study area. The two subsets are considered to be representative in terms of characteristics of the modelling outcomes. Showing pronounced differences in population density between adjacent sublocations, the first subset (Fig. 4) is employed for a comparative presentation of the two modelling approaches A and B. The

QuickBird image classification results

Many studies on extracting buildings from VHR satellite imagery are known (e.g. Bhaskaran, Paramananda, & Ramnarayan, 2013; Haverkamp, 2004; Lhomme, He, Weber, & Morin, 2009). However, most research has been conducted for urban areas in Europe, the U.S. and Australia, whereas for developing countries detailed information on the location of individual housing is scarce (Devaux, Fotsing, & Chéry, 2007). Therefore, and since ancillary data such as from Lidar surveys (cp. Weidner, 1997) was not

Conclusions

This study has demonstrated that results obtained through object-based image analysis (OBIA) of VHR satellite imagery can provide a suitable means for implementing regional scale population distribution models, in particular for areas where sufficiently detailed land use/cover data is absent, a situation typically found in developing countries. For calculating an overall probability coefficient, the use of houses as classified from VHR imagery for a subset of the study area for establishing

Acknowledgements

The study was carried out within the research framework BIOTA East Africa (subproject E02, 01LC0625D1), funded by the German Federal Ministry of Education and Research (BMBF). We would also like to thank Eric Nyadimo from the Kenyan School Mapping Project for providing the GIS data on schools, as well as Uwe Deichmann (The World Bank) for providing a C code to implement the pycnophylactic interpolation approach.

References (52)

  • CIESIN (Center for International Earth Science Information Network)

    Gridded population of the world. Version 3 (GPWv3)

    (2000)
  • N. Devaux et al.

    Extraction automatique d’habitations en milieu rural de PED à partir de données THRS

    Cybergeo – European Journal of Geography 388

    (2007)
  • J.E. Dobson et al.

    LandScan: a global population database for estimating populations at risk

    Photogrammetric Engineering & Remote Sensing

    (2000)
  • EEA (European Environment Agency)

    The thematic accuracy of Corine land cover 2000: Assessment using LUCAS (land use/cover area frame statistical survey)

    (2006)
  • C.L. Eicher et al.

    Dasymetric mapping and areal interpolation: implementation and evaluation

    Cartography and Geographic Information Science

    (2001)
  • D. Freedman et al.

    On the histogram as a density estimator: L2 theory

    Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete

    (1981)
  • F.J. Gallego

    A population density grid of the European Union

    Population and Environment

    (2010)
  • F.J. Gallego et al.

    Disaggregating population density of the European Union with CORINE land cover

    International Journal of Geographic Information Science

    (2011)
  • D. Haverkamp

    Automatic building extraction from IKONOS imagery. Paper presented at the ASPRS 2004 annual conference, 23–28/05/2004

    (2004)
  • S.I. Hay et al.

    The accuracy of human population maps for public health application

    Tropical Medicine & International Health

    (2005)
  • T.S. Jayne et al.

    Land constraints in Kenya's densely populated rural areas: implications for food policy and institutional reform

    Food Security

    (2012)
  • Kenea, W. (2007). Rural livelihood, land management and biodiversity – Community participation for biodiversity...
  • S. Lhomme et al.

    A new approach to building identification from very-high-spatial resolution images

    International Journal of Remote Sensing

    (2009)
  • G.Y. Li et al.

    Using Landsat ETM + imagery to measure population density in Indianapolis, Indiana, USA

    Photogrammetric Engineering & Remote Sensing

    (2005)
  • Y. Liao et al.

    Integration of GP and GA for mapping population distribution

    International Journal of Geographic Information Science

    (2010)
  • C. Linard et al.

    A high resolution spatial population database of Somalia for disease risk mapping

    International Journal of Health Geography

    (2010)
  • Cited by (46)

    • Mapping population on Tibetan Plateau by fusing VIIRS data and nighttime Tencent location-based services data

      2022, Ecological Indicators
      Citation Excerpt :

      Unlike nighttime light data, human activities are inevitable within a specific range of human settlement, and inversion of population distribution based on human activity carriers becomes an essential idea for population estimation from optical satellite data. Scholars monitor the population distribution by surface human settlements (Lung et al., 2013), buildings (Freire et al., 2020), land use (Dmowska and Stepinski, 2017), and land cover. Among them, high-resolution land cover data can improve the accuracy of population modelling (Linard et al., 2011).

    • Population mapping in China with Tencent social user and remote sensing data

      2021, Applied Geography
      Citation Excerpt :

      Li et al. (2018) used NPP/VIIRS nightlight data to simulate the spatial distribution of the population of Beijing and found that the overall estimation accuracy achieved with nightlight data was about 62% as calculated with reference to the census data. In addition to nightlight satellite data, intermediate-to high-resolution optical satellite data are also often used to estimate the population distribution by monitoring human settlements on the ground (Lung et al., 2013). In contrast to luminous satellite data, human physical settlements, such as residences and buildings, are the most important features to exploit for estimation of the population distribution (Ye et al., 2019).

    • Dasymetric mapping of urban population in China based on radiance corrected DMSP-OLS nighttime light and land cover data

      2018, Science of the Total Environment
      Citation Excerpt :

      Most applications, especially for large areas, generate the continuous weight using auxiliary geospatial data and remote sensing data, by developing the statistical relationship between population and a series of explanatory variables. The commonly used geospatial variables include: distance based variable (e.g., distance to urban, river, road, etc.), topographic variable (e.g., elevation, slope, etc.), and LULC data, among others (Azar et al., 2013; Bajat et al., 2011; Bryan and Brian, 2013; Dong et al., 2015; Lung et al., 2013; Mao et al., 2012; McKee et al., 2015; Wang and Zhou, 1999; Yue et al., 2005). NTL, specifically that from Defense Meteorological Satellite Program (DMSP) measures light emission from human activities and can be used to disaggregate population, especially for large areas (Elvidge et al., 1999; Sutton, 1997; Sutton et al., 1997).

    View all citing articles on Scopus
    1

    Present address: European Commission – Joint Research Centre, Institute for Environment & Sustainability (IES), Via E. Fermi, 27439, TP 272, 21027 Ispra (VA), Italy.

    2

    Present address: Bundesamt für Naturschutz, AS Insel Vilm, 18581 Putbus, Germany.

    3

    Present address: FAO-SWALIM, UN Somalia, Ngecha Road, Off Lower Kabete Road, P.O. Box 30470-00100, Nairobi, Kenya.

    View full text