Modeling the Distribution of Human Population with Nighttime Satellite Imagery and Gridded
Population of the World

Human population is one of the key factors in any analysis of human-environment interactions and global change research. The environmental implications associated with population growth are not only related to the total number of people but also to their spatial distribution. Human populations are not uniformly distributed on the earth’s surface. For instance, in 1990, 50% of the human population inhabited less than 3% of the earth’s ice-free land area (Small, 2001). Almost half of the world’s current population lives in cities and the number is likely to increase to over 60 percent by 2030 (United Nations, 2001). Such concentration of people in spatially localized areas has significant environmental and social impacts at local, regional, and global levels.

Francesca Pozzi, Christopher Small, and Gregory Yetman

In order to understand and study such impacts, population and physical factors need to be made available as detailed, spatially dis­aggregated data and reduced to comparable scales. While many environmental data are available already as spatial datasets, census data usually require some form of spatial allocation, to convert irregularly shaped census units to globally or regionally consistent population grids. The Gridded Population of the World (GPW2) dataset combines 127,105 census estimates worldwide into a uniform grid with a spatial resolution of 2.5’ (4.5 km at the equator). However, the global median resolution for the census data used in the compilation is 31 km (Small and Cohen, 2003, see also www.ldeo.columbia.edu/~small/population.html). Additional information about the location and size of urban settlements can greatly improve spatial resolution when estimating the population distribution within a country. The Defense Meteorological Satellite Program Operational Line Scanner (DMSP/OLS) has been used to make global maps of stable lights corresponding to lighted urban areas (Elvidge et al., 1997a, 1999). With a ground resolution of 2.7 km, these data could provide a valuable proxy for human settlement locations and distributions.
In this study we look at the DMSP/OLS nighttime light data as a potential means to refine the spatial detail of the population distribution, as represented by the Gridded Population of the World. By comparing the Log10 of population density with the nighttime light frequency for a sample of regions of the world with spatially detailed administrative data, we developed an algorithm that relocates fractions of populations within large administrative units to locations of lighted settlements.

Description of the Datasets
Gridded Population of the World (GPW)
The Gridded Population of the World was initially developed in 1995 at the National Center for Geographic Information Analysis-NCGIA (Tobler et al., 1995) and was the first raster global dataset of population totals based solely on administrative boundary data and population estimates associated with those administrative units. GPW represented the first attempt to generate a spatially georeferenced global population dataset that could be integrated with environmental datasets.
Since that first release, higher resolution population datasets have been compiled for various regions of the world. In 2000, CIESIN released an up­dated version of the GPW: GPW Version 2 (CIESIN, 2000). GPW2 is based on an improved median resolution of 30 km for the input data layers of administrative units and increased detail of population data. No effort was made to “model” population distribution and no ancillary data were used to predict population distribution or to revise the population estimates. Other currently available population distribution models redistribute population within administrative units on the basis of ancillary information and specific assumptions about population clustering (e.g., Landscan, available at www.ornl.gov/gist/ land­­scan/). The only assumption made when creating GPW2 is that population is uniformly distributed within each administrative unit. For a detailed description of the gridding algorithm and discussion on sources of error in the GPW2 dataset, see Deichmann et al. (2001). GPW2 data for 1990 and 1995 are available at sedac.ciesin. columbia.edu/plue/gpw/. These include population data for the years 1990 and 1995, both unadjusted and adjusted to match United Nations population estimates for those years. Data on land area and population density are also included.
In this study we used unadjusted population data for the year 1990 and the land area data.

Nighttime Lights
The nighttime lights dataset used in this analysis is the World Stable Light Dataset, which was produced using time series data from the Defense Meteorological Satellite Program (DMSP) Operational Linescan System (OLS) for the period 1 October 1994 to 30 April 1995 (Elvidge et al., 1997a, 1997b). The visible band of the OLS is intensified at night, permitting detection of nocturnal visible-near infrared emissions from cities, towns and villages. The pixel values are measurements of the temporal frequency with which lights were observed normalized by the total number of cloud-free observations.
Some of the issues encountered in the development of the human settlements dataset are believed to depend on artifacts of the sensor spatial resolution and data processing (Elvidge 1997a, 2001). The “blooming” effect is an overestimation of the actual size of human settlements due to the large OLS pixel size, the OLS’ capability to detect subpixel light source and to geolocation errors. The dataset was originally produced at a nominal resolution of 30 arc seconds. For this analysis we aggregated the data up to 2.5 minute using bicubic re-sampling to match GPW2 cell resolution.

Methodology and Case Studies
The GPW gridding approach is based on the assumption of uniform spatial distribution of population within each of the 127,105 administrative units; thus, the spatial detail of the gridded data is directly related to the spatial resolution of the administrative data on which they are based. Nighttime light imagery resolves lighted settlements as small as 2.7 km in diameter, thereby providing more spatially explicit information on spatial distribution of population in areas lacking detailed census data.
To produce the new population distribution dataset, we i) examined the relationship between population density and light frequency in areas with spatially detailed census data; ii) developed a transfer function to map light frequency into population density; and iii) developed a mass-conserving algorithm that relocates fractions of populations within large administrative units to locations of lighted settlements. (See Figure 1 and Figure 2.)
The algorithm that relocates fractions of populations within large administrative units to locations of lighted settlements was based on the above transfer function, and on the decision rule that relocates population at densities lower than the threshold given by the transfer function but leaves population at higher densities unchanged. This assumes that the points at lower densities correspond to lighted settlements in regions with coarse census districting and that points with higher densities represent settlements in areas with higher regional densities than the reference countries (Northeast United States, France, Spain and Portugal).
The algorithm was developed as an Arc Macro Language (AML) script using ESRI’s ArcInfo. Initially a mask of areas where the predicted Log10 population densities were greater than the Log10 population density of GPW2 was created. Then the predicted population values were totaled by administrative units and subtracted from the administrative unit totals. In this way the total population within each administrative unit is maintained but the spatial location has been improved by the additional polygons within the administrative units. The result is spatial refinement in areas poorly sampled by the census data. As initial case studies, countries with relatively large populations but poor spatial detail in administrative boundary data were chosen. Figure 3 illustrates the results for Japan.

Conclusions
The results of this analysis shows that in countries with relatively large populations but poor spatial detail in administrative boundary data, the partial reallocation of population into urban centers provides a more spatially explicit representation of population distribution than the original GPW2. The primary advantage of this product over the input GPW2 dataset is the addition of considerable spatial detail in sparsely populated areas not well represented in census data. The analysis was done in­dependently of biophysical or social par­ameters, thus making minimal as­sump­tions about factors influencing population distribution. Nonetheless the type of information about the location and size of urban settlements provided by the nighttime lights imagery can greatly improve the results of modeling efforts when estimating the population distribution within a country.

Acknowledgments
This article first appeared in the Proceedings of the Pecora 15 Conference, in Denver, Colorado, in November 2002. Reproduced with permission, the American Society for Photogrammetry and Remote Sensing, Pozzi, F., et al. "Modeling the Distribution of Human Population with Nighttime Satellite Imagery and Gridded Population of the World." Integrating Remote Sensing at the Global, Regional, and Local Scale - Proceeding of the Pecora 15/Land Satellite Information IV conference and ISPRS Commission I Mid-term Symposium/FIEOS. Bethesda: ASPRS, 2002.
The authors acknowledge the support and resources of CIESIN’s Socioeconomic Data and Applications Center (SEDAC) and the Columbia Earth Institute. The views expressed in this article are those of the authors and are not necessarily those of CIESIN or Columbia University. The authors would also like to acknowledge Dr. Chris Elvidge for providing the World Stable Lights Datasets to CIESIN and Dr. Uwe Deichmann for his inputs on population modeling.

About the Authors
Francesca Pozzi and Gregory Yetman are Research Associates at the Center for International Earth Science Information Network (CIESIN), Columbia University.
Christopher Small is a geophysicist at Lamont-Doherty Earth Observatory, Columbia University.

For more information on this article, and full list of references, please e-mail Francesca Pozzi at fpozzi@ciesin. columbia.edu

Back