GeoJournal DOI 10.1007/s10708-007-9102-z
A critical look at representations of urban areas in global maps David Potere Æ Annemarie Schneider
Springer Science + Business Media B.V. 2007
Abstract According to the UN, the number of urban dwellers is expected to increase from roughly 3.2 billion today to more than 4.9 billion by 2030. An accurate and regularly updated estimate of the extent and spatial distribution of urban land is an important first step in our search for realistic responses to the ecological and social consequences of what promises to be the most rapid urbanization in world history. By employing circa-2000 satellite remote sensing imagery, geographic information systems, and census data, six groups from government and academia in both the EU and the US have created global maps that can be used to describe urban land. We compare these maps from global to sub-national scales, for the first time applying Global Discrete Grids to the problem of global-scale map comparison. Although most of these maps share common data inputs, they differ by as much as an order of magnitude in their estimates of the total areal extent of the Earth’s urban land (from
D. Potere (&) Office of Population Research, Princeton University, 207 Wallace Hall, Princeton, NJ 08544, USA e-mail:
[email protected] A. Schneider Geography Department and the Institute for Computational Earth System Science, University of California at Santa Barbara, 5703 Ellison Hall, Santa Barbara, CA 93106-4060, USA e-mail:
[email protected]
0.27 to 3.52 million km2). A sub-national analysis of the spatial distribution of urban land reveals that inter-map correlations are highest in North America ( r = 0.90), intermediate in Europe, South and Central America, and Sub-Saharan Africa ( r = 0.78), and lowest in Asia ( r = 0.63). Across most regions, our analysis uncovers a degree of variance that is high enough to call into question the consistency of each group’s approach to urban land, pointing to the need for both a common urban taxonomy and a global urban assessment effort. Keywords GIS Global Land cover Mapping Remote sensing Urban City Population
Introduction How much of the planet is covered by urban land? How are the world’s cities distributed across continents, regions, and countries? With the United Nations (UN) Population Division predicting the largest absolute increase in urban dwellers for any quarter-century in human history (UN 2005), answers to these questions will play a critical role in our planning for the ecological and social developments of the next 25 years. Many countries track their urban population with censuses, but these census counts cannot be readily transformed into estimates of urban areal extent. Conversion from
123
GeoJournal
urban population to urban area is complicated by high variance in the density of urban settlement, and the large administrative districts that often underlie census results. To estimate the amount of land devoted to urban areas, it is necessary to build spatially explicit human settlement maps that combine ground-based census data with satellite observations of the built environment. This research compares six contemporary answers to the question of how much of Earth’s land surface is covered by urban areas. Although several of these global urban maps draw on common data sources, they show a remarkably high degree of variability at the scale of continents, regions, nations, and cities. Even at the global scale, the difference between these urban maps is extreme (Fig. 1). Six independent groups have drawn on a combination of satellite imagery, ground-based census data, and GIS maps to create estimates of circa-2000 urban areas with global extents that range from 276,000 to 3,524,000 km2. Our findings point to a need for a common set of standards for characterizing human settlement and a
global validation campaign based on those new standards.
Fig. 1 Global totals for six spatially explicit estimates of urban area (1000s of sq. km): the National GeospatialIntelligence Agency’s Vector Map Level Zero (VMAP0), the European Commission’s Global Land Cover 2000 v1.1 (GLC00), the Netherlands Environmental Assessment Agency’s History Database of the Global Environment v3 (HYDE3), the US National Oceanic and Atmospheric Administration’s Nighttime Lights-based Global Impervious Surface Area beta product (IMPSA), NASA’s Moderate Resolution
Imaging Spectroradiometer Urban Land Cover v4 2001 (MODIS), and Columbia University’s Global Rural-Urban Mapping Project version alpha (GRUMP). The dotted line is an estimate of urban area based on UN Population Division national-level urban statistics (UN 2003) and regional-level urban population densities for the year 2000 (Angel et al. 2005). Note the order of magnitude range for these totals. For scale, the difference in the GRUMP and VMAP0 totals is approximately equal to the land surface area of India
123
Background Why global urban maps? The density, spatial distribution, and physical characteristics of human settlement are important drivers of social and environmental change at multiple scales (Massey 2005). Based on national-level urban population projections, the UN now estimates that urban communities account for more than half of the human population (UN 2005). Aside from some city-level statistics on the largest urban areas, these UN estimates do not reveal where urbanites live and work at sub-national scales. There exists a largely unmet need for a map of human settlement patterns that is global, assessed for accuracy, at moderate resolution, and regularly updated. Such a map is crucial on several fronts:
GeoJournal
Global urban maps can help us track contemporary urban expansion We are living in an urbanizing1 world. Two components underlie the conclusion that this urbanization will lead to an expansion of urban land: (a) despite cautions regarding the confidence that should be placed on more distant projections (Cohen 2004), the number of urban dwellers is on course to increase by 1 billion people over the next 15 years and by nearly 2 billion by 2030 (UN 2005); and (b) mounting evidence suggests that density (population per unit of urban land area) is decreasing for a number of cities in both the developing and the developed world (Angel et al. 2005; Schneider and Woodcock in press). When combined, these two factors point to the potential for a rapid expansion of urban areas worldwide. The magnitude of the demographic trend and its developing world character are evident in Fig. 2, which traces UN urban population estimates and projections by world region from 1950 to 2030. Figure 2b reveals that whereas the urban residents of Europe, Japan, North America, and Australia/New Zealand (NZ) (bottom three green bars) accounted for 58% of the global urban population in 1950, today they represent only 29% of that total, and by 2030 they are projected to account for just 21%. Although we are confident that urbanization will continue during the next quarter-century throughout the developing world (Montgomery et al. 2003; Brockerhoff 2000), there is considerable uncertainty behind aspects of both the UN’s pre-2005 estimates (left of dashed line in Fig. 2b) and post-2005 projections (right of dashed line). Aside from the remaining unknowns in effectively modeling urbanization (Montgomery et al. 2003), there are significant challenges in providing a consistent, operational definition of the term ‘urban.’ In a review of the UN’s statistics on urban populations, Utzinger and Keiser (2006) point out that 228 member nations employ at least 10 categories of urban classification, drawing on various combinations of population size and density, administrative boundaries, and economic activities. Global urban maps produced using consistent methodologies that incorporate synoptic satellite observations could 1
The level of urbanization describes the fraction of a population living in urban areas. Urban expansion refers to an increase in the areal extent of urban land.
improve current census-based urbanization estimates and provide useful constraints for models of urban expansion.
Global urban maps can improve our understanding of the influence of urban areas on the biosphere As cities grow, the degree and complexity of their effects on ecosystems expands, prompting some to define a uniquely urban ecology (Kaye et al. 2006). Despite their small area relative to the total land surface, urban locations are known to impact environmental systems across scales, including species diversity (Davies et al. 2006), microclimate (Oke 1982; Unger et al. 2001), phenology and net primary production of urban-proximate vegetation (Zhang et al. 2004; Milesi et al. 2003), and global climate and biogeochemical systems (Calbo et al. 1998; Peters-Lidard et al. 2004). Relative to rural areas, cities appropriate a disproportionate share of the Earth’s carrying capacity in terms of both resource inputs and waste sinks (Folke et al. 1997). The ecological footprint of urban areas extends beyond city boundaries, including the conversion of natural ecosystems, the loss of valuable agricultural land, fragmentation of natural habitats, contamination of air, soil and water, increased water use and runoff, reduced biodiversity, and introduction of non-native species (Rees 1992; Pickett et al. 1997; El Araby 2002; Alberti 2005). In order to confront these environmental consequences of urbanization and urban expansion, we need to build maps of human settlement that are global, reliable, of adequate spatial resolution, and regularly updated.
Global urban maps can help us adapt when the biosphere/geosphere ‘‘pushes back’’ We live on a dynamic planet which is undergoing global-scale change. Traditional hazards such as volcanoes, earthquakes, tsunamis, hurricanes, cyclones, floods, and droughts may be more destructive in areas that have undergone urbanization or other geo-engineering efforts (Ehrlich 1991; Stren et al. 1992). The devastation of New Orleans during Hurricane Katrina is a good example; intense management of the Mississippi river destroyed
123
GeoJournal Fig. 2 (a) The map of regions used in this study (modified from the UN regional scheme). The color codes in this map match the legends in Figs. 4, 6 and 11. (b) Time trends in the global distribution of urban population. Source: World Urbanization Prospects: The 2003 Revision, UN Population Division
wetlands and barrier islands south of the city that might have helped to weaken the storm before it moved inland (Bourne 2000; Travis 2005). Although both urban and rural areas are vulnerable to natural disasters, the scale of the immediate humanitarian needs and the magnitude of the recovery efforts are often higher within densely settled urban areas. Added to this list are the hazards imposed by climate-change, including eustatic sea-level rise and significant shifts in precipitation and temperature regimes (IPCC Working Group II report 2007). The degree of damage inflicted by these hazards is strongly proximitydependent, which is why it is essential that we have spatially explicit estimates of urban populations at scales well-suited to the full suite of potential hazards. Regularly updated global urban maps could serve as the basis for vulnerability assessments, improving both our preparations for and responses to natural disasters.
123
Global urban maps can help illuminate the connection between settlement characteristics and human health and well-being For the first time in recorded history, urban environments influence the health and well-being of the majority of the world’s population. In addition to many of the environmental issues mentioned above, rapid population growth of cities in developing countries with a paucity of resources can lead to increased urban poverty (Brockerhoff and Brennan 1998; Wratten 1995), greater economic and social inequality (Massey 1996), and increased exposure and susceptibility to disease (Montgomery et al. 2003; Keiser et al. 2004). At the same time, sustainable urban expansion within a favorable development climate can lead to economic growth, higher incomes, technical innovation, more efficient land and energy use, better living conditions, cleaner water, and increased access to health care and
GeoJournal
education (Sassen 1994; Castells 1996; Quigley 1998; Van Vliet 2002). Overall, the only certainty with respect to urban growth is that it is going to occur rapidly for at least the next 25 years. The ultimate social, economic, and environmental impacts are complex (Montgomery et al. 2003), and will depend on the soundness of our urban planning decisions (Burgess 2000; McGranahan and Satterthwaite 2000). To make quality decisions regarding the future of urban expansion, regional and national planners cannot rely on a piecemeal patchwork of city-scale maps. What is needed is a consistently-produced, regularly-updated, and accuracy-assessed global urban map.
Description of global urban maps Table 1 describes the six global urban maps and two urban-related maps under review in this study, including the five-letter acronyms that will be used
from this point forward. Each of these maps has a unique approach to urban land, employing a diverse set of methodologies that draw on a sometimes overlapping pool of remote sensing imagery, groundbased census results, and geographic information systems (GIS) data layers (Table 2). Two of these maps—Vector Map Level Zero (VMAP0) and Global Landcover 2000 (GLC00)—are multi-class land cover maps that include an urban component. Two are binary (urban/non-urban) maps devoted entirely to urban land: Moderate Resolution Imaging Spectroradiometer Urban Land Cover (MODIS), and Global Rural-Urban Mapping Project (GRUMP). Two new maps, Global Impervious Surface Area Map (IMPSA) and History Database of the Global Environment (HYDE3) characterize urban land as a continuous variable—as the fraction of impervious surface and the fraction of urban land, respectively. The remaining two, LandScan 2005 (LSCAN) and Nighttime Lights (LITES), measure continuous variables closely related to urban land: the ambient
Table 1 The six global urban maps examined in this research, and two urban-related continuous products (bottom two rows) Code
Map
Producer
Specifications/source
VMAP0
Vector map Level zero
US National GeospatialIntelligence Agency
Land cover and map features, vector, 1:1,000,000 scale, geographic projection, http://www.geoengine.nga.mil/
GLC00
Global land cover 2000 (v1.1)
European Commission Joint Research Center
HYDE3
History database of the global environment v3
Netherlands Environmental Assessment Agency
Land cover, 24 classes, raster, 3200 arc-seconds (*1 km), geographic projection, http://www-gvm.jrc.it/glc2000/ Global fraction of urban land, raster, 50 arc-minutes (*10 km), geographic projection, http://www.mnp.nl/hyde/
IMPSA
Global impervious surface area (2000–2001)
Earth Observation Group, US National Geophysical Data Center
Global fraction of urban land, raster, 3000 arc-seconds (*1 km), geographic projection, http://www.ngdc.noaa.gov/dmsp/
MODIS
MODIS urban land cover (2001v4)
Boston University Department of Geography (US-NASA)
Global urban land, raster, *1 km resolution, sinusoidal projection, http://www-modis.bu.edu/landcover/
GRUMP
Global rural–urban mapping project (alpha version)
Earth Institute at Columbia University
Urban/rural map, raster, 3000 arc-seconds (*1 km), geographic projection, http://www.sedac.ciesin.columbia.edu/gpw/
LITES
DMSP-OLS nighttime lights v2 (2001, F15 satellite)
National Geophysical Data Center (US-NOAA)
LSCAN
LandScan 2005
US Oak Ridge National Laboratory (US-DOE)
Nighttime illumination intensity, raster, 3000 arc-seconds (*1 km), geographic projection, http://www.ngdc.noaa.gov/dmsp/ Ambient human population, raster, 3000 arc-seconds (*1 km), geographic projection, http://www.ornl.gov/sci/landscan/
Abbreviations: MODIS, Moderate Resolution Imaging Spectroradiometer; DMSP-OLS, Defense Meteorological Satellite ProgramOperational Line Scanner; NOAA, National Oceanographic and Atmospheric Administration; NASA, National Aeronautics and Space Administration; DOE, Department of Energy
123
GeoJournal Table 2 Major inputs used to produce the six global urban maps and LandScan
Imagery
Primarily image-based
Combination
MODIS
GLC00
IMPSA
LSCAN
·
·
·
·
HYDE3
GRUMP
VMAP0
·
Variousa
High resolution Medium res.
Landsat
Landsat
Coarse res.
MODIS
SPOT-VGT
Nighttime lights
1996–1997
1994–1995
·
·
Census & maps
Map-based
2000–2001c
Annual census
·
·
·
USd
UNg
UNg
·
·
·
·
·
·
·
Road vectors
·
Global maps VMAP
·
1994–1995
City gazetters
Maps/charts
·
GeoCoverb
Landsat
·
·
Lev. 0e
MODIS
·
·
·
Lev. 0 & 1f
Lev. 0
Lev. 0
2001
GLC-2000 LandScan
2001 2004
2005
a
Notes: Various commercial and government high resolution satellites GeoCover’s 30 m land cover product is based on circa-1990 and 2000 Landsat imagery
b c
Radiance-calibrated nighttime lights product
d
US Census Bureau
e
VMAP level zero is global in extent and at 1:1,000,000 scale
f
VMAP level one products are near-global in extent and at 1:250,000 scale
g
United Nations Population Division, 2003 World Urbanization Prospects
human population (defined as the population of a given area averaged over a 24-h period) and the intensity of stable nighttime illumination, respectively. These maps are beset by the problem of an operational definition of urban areas—the same challenge faced by the UN in their national-level census estimates of urban populations. The map legends reveal little of these groups’ underlying conception of the urban class: VMAP0 uses the term ‘‘built-up,’’ GLC00 makes use of ‘‘artificial surfaces and associated areas,’’ MODIS employs ‘‘urban and built-up,’’ and GRUMP utilizes ‘‘urban extent.’’ As part of preparations behind a forthcoming circa-2005 medium resolution (300 m) global land cover map, GLOBCOVER, Herold et al. (2006) have recently pointed out the importance of harmonizing map
123
legends to improve our understanding of the Earth’s land surface. Although not reflected in their legends, these six global urban maps model urban land based on a complex matrix of attributes, including: (1) remotely sensed impervious surface, (2) the concentration of mapped infrastructure, (3) population density, and (4) electrification. In the absence of a clear definition of urban areas, the best approach to understanding these maps is to explore their production processes in closer detail. VMAP0 (Table 2, rightmost column), one of the earliest available global maps, is a 1:1,000,000-scale GIS product that was created by digitizing a large collection of maps and navigational charts starting in the 1950s. The popular Digital Chart of the World (DCW) was based on the 1993-version of VMAP0 (Danko 1992). The urban polygons of VMAP0 often
GeoJournal
trace the outer edge of urban areas, without delineating interior patches of non-urban land. These urban polygons are not labeled with the date of the underlying map from which they were extracted, and are sometimes poorly geo-located. Nevertheless, because it is a global dataset and a conservative estimate of urban land area, VMAP0 is used as part of the input stream for GLC00, GRUMP, LSCAN, and HYDE3. MODIS (left column of Table 2) is primarily derived from the supervised classification of coarse resolution multi-spectral, multi-temporal daytime satellite imagery from the Moderate Resolution Imaging Spectroradiometer instrument aboard NASA’s Terra platform (Schneider et al. 2003, 2005). GLC00 is based on similar remote sensing data from the European SPOT-VEGETATION instrument, but was completed on a regional basis by 11 research teams each employing distinct methodologies (Bartholome and Belward 2005). In both cases, MODIS and GLC00 incorporated LITES data to constrain their classifications (1996–1997 and 1994–1995 LITES, respectively). The remaining maps rely on a combination of remote sensing and ground-based inputs. GRUMP integrates VMAP0, 1994–1995 LITES, census data, and a variety of ancillary GIS data sets (CIESIN 2004). IMPSA employs LSCAN 2004 and radiance-calibrated LITES 2000–2001 data to model impervious surface area globally (Tuttle 2007). The IMPSA model was trained using a medium-resolution map of impervious surface area for the US produced by the MultiResolution Land Cover Characteristics Consortium (Yang et al. 2003). The global IMPSA process represents a continuation of previous efforts focused exclusively on the US (Elvidge et al. 2004). HYDE3 combines LSCAN 2005 population density with UN urban population estimates (UN 2003), city gazetteers, and assumptions about mean urban population densities to estimate the fraction of urban land cover (Goldewijk 2001, 2005). One additional global land cover map that also contains a representation of urban areas is EarthSat’s GeoCover Land Cover (LC). GeoCover LC is a supervised land cover classification of the GeoCover image archive, a collection of circa-1990 and circa2000 30 m-resolution Landsat imagery (Tucker et al. 2004). Although the underlying GeoCover Landsat imagery is freely available to the public (Global Land
Cover Facility 2007), the land cover classification is a commercial product produced under NASA contract by EarthSat Corporation (now MDA Federal). GeoCover LC includes a class for ‘‘urban/built-up, developed areas greater than 60 m wide,’’ which was produced by labeling urbanized areas visible in the Landsat imagery. GeoCover LC serves as an important input into the LandScan product stream, but is not included in this analysis because it is not a truly global dataset and its prohibitive cost makes widespread use impractical. GeoCover LC has no coverage above 60 N, and has no land cover data (except for a land-water mask) for most of Europe, all of Australia/NZ, and much of Central Asia and contains only circa-1990 data for the US. In addition, many large regions of the tropics are not represented in GeoCover LC due to cloud cover in the Landsat imagery. The two other products included in this study, LITES and LSCAN, map urban-related variables. LSCAN models the ambient human population using GeoCover LC, MODIS LC, VMAP level one and above (vector maps produced by the National Geospatial-Intelligence Agency at scales of 1:250,000 or finer), VMAP0, GIS census products (similar to those used by GRUMP), Landsat data, and high resolution imagery (1–5 m). Although LSCAN was originally conceived during work related to LITES (Sutton et al. 1997; Dobson et al. 2000), the current LSCAN versions do not rely on LITES data (Bhaduri et al. 2002). The LITES map is created by the Earth Observation Group at the National Geophysical Data Center (NGDC) using data from a nighttime imaging satellite that has a primary mission of monitoring cloud cover by moonlight (the Defense Meteorological Satellite Program’s Operational Line Scanner satellite). NGDC models the average illumination intensity of human settlements and activities for all cloud-free observations within a given year through compositing many individual images (Elvidge et al. 2001; NGDC 2007). The Operational Line Scanner applies a variable gain during flight, and the gain settings are only continually monitored on a select number of orbits. Because of these variable-gains, it is not possible to convert standard LITES composites into radiance values. However, the NGDC group also produces a smaller number of radiance-calibrated composites using fixed-gain imagery. Whether or not the composites are radiance-calibrated, thus far, the
123
GeoJournal
Fig. 3 The six global urban maps and two urban-related maps for Beijing-Tianjin, China (top row), Mumbai, India (second row), Paris, France (third row), Moscow, Russia (fourth row), and Cairo, Egypt (bottom row). The first column is Nighttime Lights for 2000–2001, with white representing the areas of brightest emitted light. The second column is LandScan 2005, with the LandScan color-scheme of grey for zero population, and the yellow-brown color ramp for increasing population counts. Beginning with the Impervious Surface Area (IMPSA)
map in the third column are the six global urban maps. IMPSA and the History Database of the Global Environment v3 (HYDE3) (columns 3–4) represent the fraction of a cell covered by impervious surface or urban land, respectively. Columns five through eight are the binary global urban maps (urban areas shaded black): Vector Map Level 0, Global Land Cover 2000, MODIS Urban Land Cover, and the Global Rural– Urban Mapping Project (GRUMP). Each subset is approximately 150 · 150 km and all are oriented north-up
relationship between LITES and urban areas appears to be complex and context-dependent (Imhoff et al. 1997; Henderson et al. 2003; Small et al. 2005). Figure 3 depicts the six urban maps and two urban-related maps at work over Beijing-Tianjin (top row), Mumbai (second row), Paris (third row), Moscow (fourth row), and Cairo (bottom row). The first two columns are Nighttime Lights for 2000– 2001, and LandScan 2005. The global urban maps begin with the third column: IMPSA, HYDE3, VMAP0, GLC00, MODIS, and GRUMP (left to
right). Comparing GRUMP to LITES reveals the important role of thresholded nighttime lights in the GRUMP mapping process. Note that several lighted urban clusters do not appear in GRUMP because the LITES imagery is for 2000–2001, while GRUMP relied on 1994–1995 nighttime lights imagery to produce their urban maps. GRUMP’s use of buffers surrounding gazetteer city points is also evident in the cluster of circular urban patches in the lower left corner of the Beijing map. Even a quick glance at the VMAP0 column confirms the often dated nature of
123
GeoJournal
the maps; significant urban expansion has occurred in all of these cities since the charts and maps that underlie VMAP0 were digitized. Overall, these examples from Fig. 3 demonstrate the diversity of the six global urban maps, and the challenges of comparing them at global scales.
Previous studies Since 1990, there have been several attempts to estimate the total urban extent of the Earth. The basic approach is to combine global urban population estimates from census data with regional estimates of urban population density. Applying this technique to UN urban population statistics and urban density estimates from unnamed sources, Douglas (1994) arrives at a global figure of 2.47 million km2 of urban land circa-1985. In the same volume, Gru¨bler (1994) used similar global values on urban populations in combination with a small collection of city-scale studies on the built environment to estimate a global total of 1.3 million km2 for built-up area circa-1990. Angel et al. (2005) used satellite-derived regional urban density figures from a global sample of cities with populations of more than 100,000 to arrive at an estimate of approximately 400,000 km2 for urban land circa-2000. None of these efforts can be considered a global urban map, in the sense that none contain information regarding sub-national distributions of urban land. To date, there has been no systematic comparison of global urban maps. Although there have been a number of global and regional scale comparisons and validation efforts aimed at other, more common land cover classes, this study is the first effort to bring all the existing global urban products into a common analysis environment. Urban areas make up only 1–3% of the Earth’s total land area, making urban areas a rare class compared to much more common land cover types such as forest, grassland, and savanna. This minority status has led to scant urban-specific analysis in comparisons of global land cover products. See and Fritz (2006), for example, do not report any urban findings in their comparison of GLC00 and MODIS land cover. The comparison conducted by Giri et al. (2005) is the only global land cover comparison that reports results concerning urban land. While Giri et al.
purport to compare GLC00 and MODIS on a pixelby-pixel basis, because they used an early version of MODIS land cover, the urban component of their analysis is actually a comparison of GLC00 to VMAP0. Prior to the adoption of the Schneider et al. (2003) global MODIS urban map, MODIS Land Cover relied on the Global Land Cover Characteristics (GLCC) map, which in turn relied on VMAP0 for the urban class (Loveland et al. 2000; Friedl et al. 2002). Giri et al. report that the GLC00 total urban area is 93% of the VMAP0 total, which is consistent with our findings using the original GLC-2000 v1.1 product. Giri et al. (2005) also repeated their intermap contingency tables for each of 16 biomes, and found no statistically significant differences for any biome-class combination. Globally, the class most often confused for urban in MODIS and GLC00 was croplands, a finding that is not surprising when one considers that croplands are by far the most abundant urban-proximate land cover class. In the process of creating a synthetic map of land cover for carbon cycle modeling, Jung et al. (2006) compare MODIS, GLC 2000 v1.1, and GLCC. They report class-specific consistencies among the three maps, and indicate that urban areas (together with shrublands and grasslands) are problematic classes with relatively low inter-map consistencies. Because carbon modeling was the focus of their work, Jung et al. concentrate primarily on vegetative classes. The only other global comparative work was a study by Hansen et al. (2000), comparing two early Advanced Very High Resolution Radiometer (AVHRR) land cover products which both likely relied on VMAP0 for their urban classifications. Although not a comparative project, the recent global validation of GLC 2000 by Mayaux et al. (2006) included an urban class, but because the validation set was not weighted to account for the rarity of urban land, they report no urban-specific accuracy assessments. At regional and city scales, there have been a number of comparisons which include some of the urban maps considered here. Schneider et al. (2003) compared VMAP0, LITES and the MODIS-derived urban map for cities in North America, uncovering variable amounts of underestimation of urban land by VMAP0 and systematic overestimation of urban land by the 1994–1995 nighttime lights data (thresholded) when compared against more recent medium
123
GeoJournal
resolution maps of urban land derived from Landsat imagery. Tatem et al. (2005) compared a map of urban areas in Kenya to five global scale urban maps, and unsurprisingly reported that their medium resolution Kenya map based on Landsat and Radarsat imagery is of higher accuracy than the coarse resolution global maps. Finally, Small et al. (2005) assessed thresholded 1994–1995 and 2000 LITES data against Landsat-based maps for a global sample of 17 cities, confirming that no single LITES threshold is suitable for mapping urban land.
Research methods and results Map preparation Not all of the global urban maps share a common geospatial model (see Table 1). For our analytical environment, we selected a geographic projection, 3000 arc-second raster (*0.86 km2 cells, at the equator), and the WGS-84 datum. LITES, LSCAN, GRUMP, and IMPSA required no modification. HYDE3 was down-scaled to 3000 arc-second cells from 5 arc-minute cells (*9.27 km wide at the equator). The MODIS LC group provided a geographic reprojection of their native sinusoidal product. Although GLC00 is in a geographic projection, the pixel size was resampled from 32.1400 arcseconds to 3000 arc-seconds. Finally, the VMAP0 product is in a vector format, and was converted to raster. All of the modified products were checked against their native counterparts at various stages and biases were determined to be negligible throughout the map preparation process. To make comparisons across maps at continental, sub-continental (regional), and national scales, we needed a coarse-resolution map of international boundaries. Unfortunately, few options exist at 1 km spatial resolution with a consistent land and water boundary. We opted to use the boundary file produced by the LSCAN program, as it is at 3000 arcsecond resolution and is updated annually. To account for significant differences in the how each map delineated land versus water, we created custom international border and land-water boundary files for each product. In this step, land pixels from each map that fell outside of the LSCAN land-water mask were retained as land and assigned to the country of their
123
nearest land neighbor in the LSCAN country boundary file. This procedure prevents the land water mask from eliminating some of the coastal areas within several of the maps, which can result in serious underestimates of urban land. We then cross-walked the LSCAN country names (derived from the US Census Bureau) to their counterparts in the system employed by the UN Statistics Division and the International Organization for Standardization (UN Statistics Division 2007). These UN country codes are accompanied by a scheme for delineating continents and regions. The regional scheme employed throughout the rest of this analysis is a slightly modified version of this UN regional scheme (see Fig. 2a for a map of the continental divisions). These modifications include: (1) reassignment of Sudan from Northern Africa to West Africa in order to maintain Sudan’s traditional association with the super-region of Sub-Saharan Africa; (2) combining Australia/NZ with North America on the basis of demographic and economic similarities (e.g. moderate levels of projected population growth and in-migration); and (3) merging Japan with Western Europe due to Japan’s similarly high levels of contemporary urbanization and projected population decline. The global GLC-2000 v1.1 indicates that virtually no urban areas are present in southern China. To correct for some of these and other urban omissions, we relied on recently updated GLC regional maps. These updates are still based on circa-2000 imagery, but represent refinements in the regional protocols by the GLC 2000 teams for Europe and Asia. By including these improved regional maps, the GLC00 map used in this analysis contains 9% (28,000 km2) more urban area than the original global map (v 1.1). The following regions were allocated additional urban area through the use of the updated maps: Southern Europe (9,000 km2), Northern Europe (6,800 km2), East Asia (4,700 km2), South East Asia (4,500 km2), and South Central Asia (2,300 km2). MODIS LC also includes a small region of urbanomission (just visible as a thin stripe over southern Europe in Fig. 9e) which was not corrected. Both LITES and LSCAN are properly classified as urban-related products, because their target variables (nighttime emitted light and ambient human population) are not simple proxies for urban areas. The relationship between urban area and nighttime
GeoJournal
Fig. 4 Continental distributions of urban land (colored areas, values on left axes) and total urban area (solid black lines, right axes, logarithmic scale) for a range of thresholds in light intensity and population density applied to Nighttime Lights 2000–2001 (left figure) and LandScan 2005 (right figure), respectively. The dotted red lines indicate the thresholds of illumination intensity and population density selected for
constructing the Nighttime Lights (LITES-15) and LandScan (LS-3000) urban maps used for comparison purposes throughout this study (minimum DMSP intensity of 15 and minimum population density 3,000 persons per sq. km, respectively). These thresholds were selected such that each map includes approximately 3 billion people, which corresponds to the circa2000 UN urban population estimate
illumination or population density is contextdependent, nonlinear, and variable both within and across geographic regions and national borders. Because the precise nature of those relationships remains poorly understood, we chose to include LITES and LSCAN in this analysis through a simple thresholding procedure. In both datasets, pixels decrease in value—either illumination intensity or population density—as the distance from the urban center increases. Applying a threshold limits how many of the pixels are ‘turned on’ for classification as part of the urban areas. For both datasets, a high threshold generates a small urban area, while a more conservative value produces an expanded urban area. In order to select a minimum threshold of population density or nighttime emitted light intensity for each map, we explored the full range of possible thresholds (Fig. 4) in a manner similar to Small et al. (2005) and Schneider et al. (2003, 2005). Figure 4 traces the continental distributions of urban area (shaded areas, left y-axes) and total urban area (black lines, right y-axes) for a range of LITES and LSCAN thresholds (x axes). The black lines reveal the global urban areal extents that result from applying each minimum light intensity or population density threshold to the continuous datasets. The shaded colored areas show the proportion of urban
land within each world region across all thresholds. Thresholds greater than 10 digital numbers (DNs)2 for LITES and greater than 100 people-per-sq. km for LSCAN reveal a relatively stable distribution of urban areas across all of the world regions. Notably, the principal source of variability in both figures is the North America and Australia/NZ region (dark green). In the LITES figure (left), North America is the dominant region for all LITES thresholds above 30 DNs, a fact that is not surprising given the large areal extents of US cities and their high levels of energy consumption (Welch 1980; Doll et al. 2006). In contrast, LSCAN’s world regional distribution (right figure) more closely resembles the distribution of the world’s urban population from Fig. 2, where East Asia (salmon) and South Central Asia (purple) dominate across all thresholds. The black lines of Fig. 4 (left and right) indicate the very wide range of global urban area estimates that are possible when employing a simple global threshold (100,000–50 million km2). To create global urban maps from LITES and LSCAN, we selected
2
The DMPS-OLS satellite is sensitive to radiation in the visible-near infrared region of the electromagnetic spectrum (0.44–0.94 lm). Light intensity is encoded in 6-bit digital numbers.
123
GeoJournal Table 3 The confusion matrix for the four Boolean maps of urban land cover
VMAP0
Table 4 For each combination of the four Boolean maps, this table shows Cohen’s kappa (j) statistic
VMAP0
GLC00
MODIS
GRUMP
VMAP0
1
0.41
0.54
0.76
VMAP0
1
GLC00
MODIS
GLC00
0.37
1
0.60
0.89
GLC00
0.39
1
MODIS
0.21
0.26
1
0.80
MODIS
0.30
0.36
1
GRUMP
0.06
0.08
0.16
1
GRUMP
0.11
0.14
0.27
GRUMP
1
Each entry represents the percent of Map A (in the rows) that is also present in Map B (in the columns)
The strength of agreement ranges from slight (0.01–0.20) to fair (0.21–0.40)
thresholds that generate total urban areal extents encompassing approximately 3 billion people, a total urban population which corresponds roughly with the non-spatial UN estimate for circa-2000. To assess the population of these thresholded maps, we use LSCAN. The vertical red dashed lines in Fig. 4 mark these thresholds: nighttime lights DN of 15 and population density of 3,000 persons-per-sq. km. Henceforth, LITES-15 and LSCAN-3000 refer to these thresholded urban maps. There is an order of magnitude difference between the total urban extents of these two thresholded maps (LSCAN-3000 contains roughly 380,000 km2 globally and LITES-15 is roughly 3 million km2), yet both estimates are well within the range of extents from the six global urban maps included in this study. Although there are no satisfactory criteria for selecting a global-scale population density or illumination intensity threshold to define urban areas, we include these two maps in our analysis to better understand the relationship among the six urban area maps and these two urban-related products (LITES, LSCAN). Because these maps are produced using simple global thresholds which can produce inconsistent local results, they are not likely to be realistic estimates of global urban area.
(shown in the rows) also present in map B (shown in the columns). Although VMAP0 is a part of the input data streams for both GLC00 and GRUMP (Table 2), only 41% and 76% of this map agrees with GLC00 and GRUMP, respectively. This result could be due to differences in geo-registration, since there is a roughly 10 3000 eastward and 3000 arc-second southward bias for GLC00 over Canada, USA, Central America, and the Caribbean relative to all of the other urban maps. It is not possible to correct for this shift because the geo-registration problem appears to be non-constant within the northwest hemisphere. Similar registration problems are not apparent for GLC00 in any other region of the world. A more methodical evaluation of geo-registration would require map assessment using medium-to-high resolution maps, and is not possible within the context of a comparative study. This example highlights one of the fundamental problems of conducting a global map comparison at the pixel-level: a shift of a single pixel can lead to significant inter-map differences. Another concern for interpreting this simple contingency table is the order of magnitude variance in total map area (Fig. 1). For instance, 60% of GLC00 agrees with MODIS (above the diagonal), while only 26% of MODIS agrees with GLC00. This is not surprising when one considers that MODIS is 2.4 times the size of GLC00. Even if every MODIS pixel were in agreement with GLC00, that would only represent 42% of MODIS urban land. Cohen’s Kappa statistic (j) is designed to help deal with these challenges of contingency table analysis (Cohen 1960; Monserud and Leemans 1992; Goldewijk and Ramankutty 2004). This more objective statistic measures strength of agreement, taking into account the potential for chance agreement (Table 4). Some have suggested that Cohen’s j values of 0.01–0.2 should be considered poor
Assessing per-pixel map agreement Contingency tables, sometimes referred to as confusion matrices, are a common way to begin inter-map comparisons. These tables are produced by overlaying a set of maps and estimating the areas of agreement and disagreement among them. Our first global comparison (Table 3) was conducted for the four dichotomous (urban/rural) maps at 3000 arcsecond resolution: VMAP0, GLC00, MODIS, and GRUMP. Each entry describes the fraction of map A
123
GeoJournal
Fig. 5 The shaded areas of a and b (right axes) are the distribution of urban area for IMPSA and HYDE3, respectively, for each fractional coverage (0–100% urban). Both distributions have a similar shape, with discontinuous spikes at 100% urban. Above this distribution are colored lines indicating the percent of IMPSA or HYDE3 in agreement
with the various mapped estimates from the four Boolean urban maps (left axes 0.0–1.0 scale). As is expected, the overall levels of these lines correspond to their relative total extent. In a, note the rapid decrease in agreement for all five maps when the IMPSA percent urban drops below 30%. For HYDE3 (b), a similar effect is only evident in GRUMP (blue)
agreement and values of 0.21–0.4 as fair agreement (Landis and Koch 1977). The overall impression from the Cohen’s j analysis is that on the per-pixel level these maps are quite distinct, as shown by the relatively low levels of agreement across all map pairs. The strongest agreements are between the GLC00/VMAP0 and MODIS/GLC00 map pairs. This result is expected, considering that VMAP is an input into the GLC00 product and that MODIS and GLC00 are both derived primarily from coarse resolution daytime imagery and LITES. For the two continuous urban maps, IMPSA and HYDE3, we can complete a more detailed per-pixel analysis by examining the relationship between their sub-pixel urban fractions and each of the four discrete maps (Fig. 5a and b). In both figures, the shaded grey bimodal distributions represent the probability density functions (pdfs) of the urban areas in IMPSA (Fig. 5a, left) and HYDE3 (Fig. 5b, right). These distributions show that the vast majority of pixels in IMPSA and HYDE3 are either entirely urban or less than 20% urban, with only a small amount of the total urban land falling between these two peaks. Atop each of these pdfs is the fraction of HYDE3 or IMPSA pixels at each sub-pixel urban fraction that are also present in one of the four discrete maps, which reveals their percent agreement. The overall level of each agreement curve corresponds well to the relative amount of total urban area for a given map,
with GRUMP (highest curve, blue), followed by MODIS (orange), GLC00 (black) and VMAP0 (lowest curve, red). It is interesting to note the rapid decline in agreement below 25% impervious surface or urban area. Although the four discrete urban maps are constructed to represent areas where the majority of the land is urban, here is evidence that they may actually be sensitive to urban fractions as low as 25% impervious surface or urban area.
Comparing the areal extent of urban land Another way to overcome the challenges of per-pixel map comparison is to aggregate individual pixels and conduct comparisons across a wide range of superpixel scales, including: global scales (Fig. 1), regional scales (Fig. 6, Tables 3, 4), national scales (Table 5), and the scale of individual urban patches (Fig. 7). Each of these scales is relevant to a different cast of potential map users, and may be more informative than pixel-level comparisons.
Regional comparison Figure 6 and Table 5 present the total areal extent of urban land for each of the six urban maps and two urban-related maps at the scale of world regions. The
123
GeoJournal
Fig. 6 The distribution of urban land per continental region for six global urban land cover maps: VMAP level 0, GLC2000 v1.1, HYDE v3, IMPSA, MODIS, and GRUMP alpha, plus the two urban derivatives constructed from LandScan and Nighttime Lights (LSCAN-3000 and LITES-15). The thickness of the horizontal bars are scaled to reflect the relative amount of urban land within each of the eight maps, globally. These
horizontal bars are divided into sections representing the relative distribution of urban land area within each of 10 regions (10-color scheme also used in Fig. 2a). Because both axes are scaled proportionately, it is possible to make direct areal comparisons between segments and across both rows and columns. The order of regional color blocks from left to right is identical to the top-down order presented in Table 5
thickness of each horizontal bar in Fig. 6 is scaled to reflect the amount of urban land within each of the eight maps relative to the other maps in the group. These horizontal bars are divided into sections representing the relative distribution of urban land within each of the 10 regions (the 10-color scheme is the same as in Fig. 2a). Because both axes are scaled proportionately, it is possible to make direct areal comparisons between segments and across both rows and columns. The most obvious feature of Fig. 6 and Table 5 is the previously-noted order of magnitude variance in total urban area, revealed by the pronounced differences in the thickness of the eight horizontal bars, from thickest to thinnest: GRUMP, MODIS, IMPSA, HYDE3, GLC00, and VMAP. The pattern of regional banding within each map (measured on the horizontal axis) reveals variance at the scale of world regions. For example, LSCAN-3000 and IMPSA have distinctive banding patterns relative the other six maps,
with far less urban land in more developed regions (left-most green bars) and more land in the bands to the far right. For LSCAN-3000 this reflects the fact that the most densely settled land in the world belongs to the two demographic billionaires of India and China. Because IMPSA uses LSCAN as an important input, its regional distribution is quite similar. The two maps differ in their estimates for North America, however, because IMPSA’s use of the Nighttime Lights data has increased the fraction of urban land relative to LSCAN-3000’s simple threshold of population density. For related reasons, the thickness and banding patterns of GRUMP and LITES-15 are very similar. There are several additional trends in Fig. 6 worth mentioning. The results for VMAP0 reflect the coldwar legacy of this data source, as shown in the wide bands for Eastern Europe (light green bar, third from left). The US agencies responsible for VMAP0 likely focused more mapping efforts on this region,
123
GeoJournal Table 5 The areal extent of each of the six global urban maps (in square kilometers) for each of the 10 regions (modified from the UN regional designations) Region
VMAPO
GLCOO
HYDE3
IMPSA
MODIS
GRUMP
North America, Australia & NZ
91,123
95,982
186,303
89,963
134,764
933,537
79,607 4,201 7,315
80,482 6,054 9,446
161,041 12,417 12,845
75,495 11,308 3,161
117,504 7,895 9,366
755,939 132,571 45,027
41,327
72,761
108,435
66,469
178,887
641,608
15,789 16,083 4,102 5,352
27,372 27,413 16,133 1,842
56,509 18,788 19,625 13,513
20,855 13,157 19,250 13,207
53,386 21,829 50,127 53,545
180,709 159,799 196,262 104,839
67,056
35,937
35,796
34,540
68,487
301,596
40,760 26,296
15,044 20,893
16,321 19,475
20,073 14,467
37,731 30,757
188,346 113,250
17,074
10,731
30,499
35,382
42,876
374,942
10,113 6,962
5,025 5,706
17,021 13,478
17,938 17,445
19,254 23,622
189,286 185,655
3,466
3,468
17,802
17,581
10,274
154,951
2,534 933
2,309 1,159
12,254 5,548
14,245 3,335
8,863 1,411
122,462 32,490
6,828
17,937
27,201
49,788
39,621
144,996
East Africa West Africa (plus Sudan) Southern Africa Middle Africa
1,585 1,483 2,459 1,301
3,286 5,378 7,883 1,390
6,181 10,448 7,717 2,856
17,887 2,430 4,965 6,506
10,136 15,468 10,482 3,535
32,310 45,967 49,977 16,741
Western Asia & North Africa
7,433
16,905
27,114
34,492
44,039
222,113
5,732 1,701
6,108 10,797
20,396 6,717
21,527 12,966
29,393 14,645
145,247 76,866
22,026
31,680
32,327
112,296
86,298
350,383
8,160 13,866
21,288 10,392
17,020 15,307
76,244 36,052
30,857 55,44
204,676 145,708
11,063
10,788
44,634
90,059
109,100
297,692
9,579 1,484
10,012 776
39,547 5,087
82,301 7,758
88,977 20,123
261,920 35,772
USA Canada Australia & NZ Western Europe & Japan Western Europe Northern Europe Southern Europe Japan Eastern Europe Russia Eastern Europe (excluding Russia) South America Brazil South America (excluding Brazil) Central America & Caribbean Central America Caribbean Sub-Saharan Africa
West Asia Northern Africa (excluding Sudan) South Central Asia India South Central Asia (excludinq India) East Asia China East Asia (excluding Japan & China)
123
GeoJournal Table 5 continued Region Southeast Asia & Pacific Islands Southeast Asia Melanesia Micronesia Polynesia Total (sq. km)
VMAPO
GLCOO
HYDE3
IMPSA
MODIS
GRUMP
8,981
11,819
21,874
40,933
12,597
102,290
8,866 97 4 13
11,714 102 3 0
21,560 216 40 58
40,252 553 68 60
12,522 71 4 0
97,440 3,454 688 707
276,377
308,007
531,985
571,504
726,943
3,524,109
elevating the relative amount of VMAP0’s Russian land far above that of any other map. At the same time, VMAP0 (and GLC00) include very small estimates for China, India and other developing regions, areas which have clearly witnessed rapid urbanization and urban expansion since the inception of this map in the 1950s, 1960s, and 1970s. The bottom-most band (LITES-15) reveals the clear relationship between development status and nighttime illumination, since the world regions with the largest amount of urban area in this map are those in North America and Western Europe. While more developed regions (the green bars) account for greater than half of the global total urban area in the LITES15 map, each of the remaining seven regions in less developed parts of the world have roughly equivalent amounts of urban land. If LITES-15 were actually intended as a realistic estimate of the global distribution of urban land, this bias towards more developed countries would be difficult to reconcile with what is actually on the ground.
Country-level comparison It is possible to continue this areal comparison at the scale of countries, which serve as the underlying units behind the world regional aggregation scheme. Although Table 5 reveals the amounts of urban area for several of the largest countries (e.g. China, India, USA, etc.), it is only possible to make a limited number of inter-country comparisons in this table. An alternative way to represent country-level information is to rank the top 10 countries with the largest urban extents for each of the six global urban maps and the two thresholded maps (Table 6). There is considerable variation in these rank lists, with only
123
USA, India, Russia, and China appearing in all of the top 10 lists. The list of countries is ordered left to right according to the UN’s estimate of total urban population for 2000 (UN 2005), and adjacent to each country in parenthesis is the 2000 World Bank ranking of gross domestic product (GDP) based on purchasing power parity. The top 10 rankings for IMPSA (and to a lesser-degree LSCAN-3000, HYDE3, and GRUMP) closely follow the order by UN urban population. This is not surprising considering the important role that demographic attributes play in the methodologies underlying these maps. Estimating inter-map national-level correlations using Kendall’s tau (s) provides a quantitative assessment of the relationships among the urban land area rankings for all 223 countries within the global dataset (Kendall 1938) (Table 7). The overall impression from this measure is one of far greater agreement than the per-pixel comparison (Tables 3, 4), as shown by the mean correlation of 0.73 and the range of 0.62–0.89. As in the regional level comparisons, Table 7 confirms that LSCAN-3000 has a particularly high correlation with the two products based in part on LSCAN, the HYDE3 (0.77) and IMPSA (0.88) maps. Likewise, LITES-15 is strongly correlated with two of its associated products, GRUMP (0.81) and MODIS (0.73). This country-level analysis reveals that international comparisons of the relative areal extent of urban land are far less sensitive to the selection of a particular global urban map than the per-pixel results imply. Of course, the large absolute differences between these maps remains important across all aggregation scales. In a final national-scale comparison, we modeled the total urban area of each country separately for each of the eight urban maps using both GDP (CIA 2000) and urban population (UN 2005) as predictors
GeoJournal Table 6 The ranks for the top 10 countries with the most urban land area for each of the six global urban maps and the two thresholded urban-related maps
in a simple additive linear model. For the eight maps, the mean adjusted R2 for the additive model was 0.87, with IMPSA and MODIS tied for the highest correlations (0.94) and VMAP0 had the lowest (0.69). Clearly, much of the national-scale variation in urban extent is strongly correlated with a country’s GDP and urban population size.
Urban patch size comparison We conclude our multi-scale areal comparison with a brief examination of the finest level of aggregate analysis: contiguous areas of urban land, defined as urban ‘patches.’ This type of analysis provides information on the effective minimum mapping unit used in each map, as well as some indication of whether the urban area is distributed in large clusters (e.g. extensive cities such as Chicago), or in a large number of small-sized patches (e.g. small towns and villages, 1 or 2 km2 in size). The number of distinct urban patches varies widely across the datasets (in descending order): LSCAN-3000 (167,000), MODIS (54,000), VMAP0 (32,000), GLC00 (22,000), GRUMP (21,000), HYDE3 (16,000), and LITES-15 (14,000). IMPSA is not included here because of the difficulty of delineating individual patches from a map with near-global coverage of urban pixels (many with extremely low fractions of impervious surface). LITES-15 finds itself at the bottom of this list, likely because of the blooming effect of bright city lights that causes the patches to be more contiguous than their counterparts in the other global urban maps (Imhoff et al. 1997; Schneider et al. 2003; Small et al. 2005). LSCAN-3000’s position at the top of the
list makes it by far the most fragmented dataset. This is influenced by LSCAN’s extensive use of the medium resolution GeoCover LC dataset, which depicts human settlements at much higher spatial resolution than any of the inputs used by the other maps (30 m vs. 1 km). When LSCAN aggregates this medium resolution data to 3000 arc-seconds, the resulting map is apt to be more fragmented than those maps that rely only on coarse resolution inputs. When examining the size distribution of urban patches within each of these datasets, it is helpful to first establish how many of these patches consist of single pixels (i.e., individual pixels that are isolated by more than one pixel from any neighboring pixel). Here too, LSCAN-3000 and MODIS have by far the most single-pixels (103,000 and 17,000, respectively), while GRUMP has the fewest (42). Figure 7 presents the probability density functions of patch sizes for all of the discrete urban maps. Because the geographic projection used for map comparison is not equal-area, single pixels in each map vary in size from 0.86 km2 at the equator, to 0.42 km2 at the southern tip of Greenland. Note the two peaks for MODIS (orange) and GLC00 (black) at patch sizes below 2 km2. These single- and double-pixel patches are much rarer in both the GRUMP (blue) and VMAP0 (red) maps, which do not rely on remotely sensed daytime imagery. This figure also confirms the massive size of most GRUMP patches (blue) relative to all of the other maps. GRUMP has a modal peak at 29–33 km2 and the majority of the patches are between 10 and 200 km2. Urban patches in the vicinity of GRUMP’s modal peak are found mostly in East and South-Central Asia.
123
GeoJournal Table 7 The rank correlation matrix (Kendall’s tau, s) for a sample of 223 countries for each map pair VMAP0
GLC00
HYDE3
IMPSA
MODIS
GRUMP
LSCAN-3000
LITES-15
VMAP0
1
–
–
–
–
–
–
–
GLC00
0.69
1
–
–
–
–
–
–
HYDE3 IMPSA
0.74 0.69
0.69 0.65
1 0.78
– 1
– –
– –
– –
– –
MODIS
0.75
0.72
0.78
0.76
1
–
–
–
GRUMP
0.76
0.71
0.81
0.77
0.79
1
–
–
LSCAN-3000
0.70
0.63
0.77
0.88
0.74
0.75
1
–
LITES-15
0.70
0.68
0.73
0.67
0.73
0.81
0.65
1
The mean tau is 0.73, with a standard deviation of 0.06 (all correlations are highly significant). The IMPSA-LSCAN, GRUMPHYDE3, and GRUMP-LITES-15 s may be considered high (greater than 0.80)
Fig. 7 The frequency of urban patch sizes (x-axis, log scale, spline curves) for each map (excluding IMPSA and HYDE3). Note the two peaks for MODIS (orange) and GLC00 (black) for patches below 2 km2. These peaks correspond to single and two-pixel patches, which may indicate speckle in these maps. These size distributions are unique to remote sensing-based
products and much rarer in the GRUMP (blue) and VMAP0 (red) maps. The modal peak in GRUMP is for patches between 29 and 33 km2 in total area. The majority of these mid-sized patches are found in East and South Central Asia. The maximum frequency of the MODIS patches was 5,213 at an area of 0.67 km2 (not plotted)
Comparing the spatial pattern of urban land
with these issues is through the use of Geodesic Discrete Global Grids (DGGs) (Sahr et al. 2003). DGGs are a class of equal-area, uniformly distributed partitions of the Earth’s surface. These partitions provide a scheme for assessing the urban area of a given map that is independent of arbitrarily defined political boundaries or world regions, and free of the limitations imposed by varying raster cell sizes in a geographic projection. Here, we employ five resolutions of hexagonal DGGs (Fig. 8) with cell sizes ranging from 800 to 70,000 km2. By computing the square root of the facet area, we can also derive a
Any comparison of the spatial pattern of urban land in these six global urban maps is made more challenging by the rarity of the urban class, the very large differences between the total urban area contained within each map, and the potential problems of geo-registration. Although using national borders and kappa statistics can help alleviate these problems, the distribution of country sizes has high variance and it is not possible to explore sub-national inter-map differences with this approach. One way of dealing
123
GeoJournal
Fig. 8 Five levels of a Discrete Global Grid system with hexagonal facets over the United Kingdom and globally. In order to conduct inter-map comparisons across a wide-range of spatial scales, a series of five Discrete Global Grids was employed with equal-area facets of between 800 and
70,000 km2 in total area. The DGG displayed globally in a is at the coarsest facet size of 70,000 km2 (the purple facets in b). The map legend also shows the effective spatial resolution for each facet size, estimated by the square root of the facet area (29–264 km)
rough estimate of spatial resolution (Small and Cohen 2004). In the following figures, the fraction of each facet covered by urban area from the six global urban maps was estimated at several DGG resolutions. Map visualization is the first area of analysis where DGGs are helpful. Displaying any global 3000 arc-second resolution map with sufficient detail is a challenge; at a reasonable resolution of 300 dots-perinch, the 40,000 · 20,000 pixel raster would be roughly 3.4 · 1.7 m. Even at this resolution, the urban class would be difficult to discern since urban areas occupy at most 3% of the Earth’s 140 million km2 of land. By aggregating the 30 arcsecond pixels of our global urban maps to the 51 kmresolution DGG facets, we have created a series of maps that effectively portray urban land at a global scale (Fig. 9a–f). In all of the maps, gray areas represent completely urban-free facets. Of the six maps, IMPSA (Fig. 9d) has by far the fewest facets that are completely free of urban land, and GLC00 (Fig. 9b) has the most. In Fig. 9g, the amount of urban area for each grid cell is averaged across all six maps (VMAP0, GLC00, HYDE3, IMPSA, MODIS, and GRUMP). The total extent of this mean urban map is
984,000 km2. Because GRUMP’s urban area is larger than any of the other five maps, GRUMP has influenced much of the distribution in this mean urban map, evident in even a quick comparison of Fig. 9f and g. The largest blocks of intensely urban areas include the Eastern USA, Western Europe, India, and Eastern China and Japan. The largest contiguous urban-free areas across all six maps are the Sahara desert, interior Australia, Siberia, Mongolia, Northern Canada and Greenland, and to a lesser degree the tropical rainforests of South America and the Kalahari Desert of Southern Africa. Turning to the question of quantifying the similarities and differences of these six maps, the DGGs can again be helpful. Hexagonal DGGs offer a global equal-area partitioning system with minimal shape distortion over the entire Earth’s surface. DGG facets are much more useful units of aggregation than either the 3000 arc-second pixels or the arbitrarily shaped and sized country borders used in sections B and C, respectively. In Fig. 10, we report Pearson correlation coefficient curves across five DGG resolutions (29, 51, 88, 153, 264 km, shown on the x-axis) for the 15 pair-wise map combinations of VMAP0, GLC00, HYDE3, IMPSA, MODIS and GRUMP facets. In
123
GeoJournal
Fig. 9 a–f Shows the percent urban land per facet for all six global urban maps, aggregated to a discrete global grid with hexagonal facets 2,500 km2 in area. The mean map, g, is the amount of urban area for each grid cell averaged acro ss all six
123
maps (VMAP0, GLC00, HYDE3, IMPSA, MODIS, and GRUMP). The sum of all these urban areas is approximately 984,000 km2. For all a–g, grey shades indicate the absence of urban land
GeoJournal
Fig. 10 The global correlations in percent urban area between all 15 pair-wise map combinations for five levels of discrete global grid (DGG) aggregation. The DGGs employed for aggregation purposes have hexagonal facets approximately 29, 51, 88, 153, and 264 km across. The y-axis is the correlation
coefficient (points are marked by triangles on each trend line), and the x-axis is the square root of this total area, which represents the level of spatial precision implied by the facet size. Only those aggregates with some urban land present in at least one of the maps were included in the samples
these regressions, we included only those facets that contained urban land in at least one map. Although this multi-resolution global map comparison is similar to the approach used by Goldewijk and Ramankutty (2004) and Greyner et al. (2006), their studies used a series of traditional raster grids in a geographic projection. While this raster approach may be more straightforward to implement than hexagonal DGGs, the drawback is that individual cells sizes vary widely by latitude. This effect creates significant shape distortions at high latitudes and non-uniform sampling, with sparser sampling in tropical versus temperate regions. The highest overall inter-map correlation curves belong to the VMAP0-GLC00-HYDE3 group (three top-most plots in Fig. 10), where the overall correlation is at or about r = 0.8. This makes sense considering that VMAP0 was a very important input for GLC00 (in some regions, the sole input), and that
HYDE3 was the only map to draw on both GLC00 and VMAP0. The lowest overall correlation curves were for IMPSA-VMAP0, IMPSA-GLC00, and MODIS-VMAP0, where the overall correlation was at or below r = 0.7. Considering that MODIS and IMPSA are the only two maps that did not draw on VMAP0, and the aforementioned similarity between VMAP0 and GLC00, this result is also to be expected. The slopes of the correlation curves are more difficult to interpret, and may be related to the size distribution and density of urban patches within each urban map. One trend is clear: for almost every intermap comparison that excludes GRUMP, the maps appear less correlated as the resolution becomes coarser. The opposite is true for all of the GRUMP comparisons, where correlation improves with increasing coarseness. This is likely tied to the extraordinarily skewed size distribution of the GRUMP urban patches (Fig. 7). At the finest DGG
123
GeoJournal
Fig. 11 These box-plots capture the 15 pair-wise Pearson correlations between maps at five resolutions within each of the 10 world regions. For instance, the five leftmost box-plots (dark green) represent the inter-map correlations for North America, Australia, and New Zealand at five resolutions (finest to coarsest, 29–264 km, left to right). For North America, the mean inter-map correlation (black horizontal bars) across all maps improves with coarser resolution. Each of the red
horizontal bars represents a mean for a group of regions: the top bar ( r = 0.90) is for North America, Australia, and New Zealand, the middle bar ( r = 0.78) is for Europe and Japan, South and Central America, and Sub-Saharan Africa, and the bottom bar ( r = 0.63) is for Western Asia and North Africa, South Central Asia, East Asia, and Southeast Asia and Pacific Islands
resolution of 29 km, each facet contains roughly 900 km2 of land. GRUMP has by far the most urban patches that are of sufficient size to saturate these facets, potentially reducing inter-map correlations. In separate scatter plots (not reported here), we noted that the relationship between GRUMP and the other maps appears to be exponential (versus linear) at finer resolutions, contributing to the poor inter-map correlations involving GRUMP at resolutions below 153 km. The DGG aggregates are also an effective tool for uncovering regional patterns in the inter-map correlations. The box-plots of Fig. 11 chart the distributions (median, inter-quartile range, and outliers) of the Pearson correlations for the 15 global urban map pairs (those pairs shown in Fig. 10) across all five resolutions. Each region has five box-plots showing correlations from the finest to the coarsest
DGG (29–624 km, left to right). The colors correspond to the world regional scheme used throughout this analysis. From Fig. 11, it is clear that the North America and Australia/NZ region (far left) has by far the strongest inter-map correlations ( r = 0.90), indicating the highest agreement in terms of the intraregional distribution of urban land. By comparison, the Asian regions (four regions, far right) have the lowest inter-map correlations ( r = 0.63), and the remaining regions are intermediate ( r = 0.78, red horizontal bars in Fig. 11). The maps show the widest range of correlations for China (the East Asia region, salmon color). Most of the variance in this region is caused by very low inter-map correlations for all the VMAP0 and GLC00 comparisons (mean correlations of 0.35 and 0.40, respectively) relative to the other four maps (mean correlations greater than 0.60). As previously
123
GeoJournal
discussed, this result may be an artifact of GLC00’s significant urban omissions over East Asia. VMAP0’s underestimates are likely tied to China’s exceptionally rapid urban expansion over the past 20 years, and the older vintage of many of VMAP0’s input maps.
Discussion and conclusions As both recipients of ecosystem services and modifiers of ecosystem processes, humanity is an important part of the biosphere. In order to better understand humanity’s role in ongoing global change processes, there is a need for a global, accuracyassessed, moderate resolution, and regularly updated map of contemporary human settlement. Global urban maps can make an important contribution here, by accounting for the residences of more than half of the human population. Such maps are of considerable interest to a wide range of users, including: regional and national planners, disaster management specialists, humanitarian and development aid coordinators, epidemiologists, demographers, conservation biologists, climatologists, urban ecologists, and ecological economists. Six groups from government and academia in both the EU and the US have created global maps that can be used to describe contemporary urban land (Vector Map Level Zero, Global Landcover 2000, History Database of the Global Environment, Global Impervious Surface Area, MODIS Urban Land Cover, and Global Rural-Urban Mapping Project). Two new maps, GLOBCOVER and a second-generation MODIS product, are forthcoming in 2007, but were not complete at the time of publication. Despite the considerable resources allocated to the task of creating these global urban maps, this comparative study has revealed that the six existing maps differ by as much as an order of magnitude in their estimates of the total areal extent of the Earth’s urban land (0.27–3.52 million km2). Differences in these six maps persist at the scale of regions, countries, and urban patches. To better visualize these maps and conduct a quantitative map comparison, we employed a hexagonal system of Discrete Global Grids (DGGs). An analysis of the spatial distribution of urban land based on these DGGs across a wide range of spatial resolutions (29– 264 km) has revealed that inter-map correlations are
highest in North America ( r = 0.90), lowest in Asia ( r = 0.63), and intermediate in Europe, South and Central America, and Sub-Saharan Africa ( r = 0.78). These large inter-map variances may be driven by a combination of several factors, including differences in the timing of map construction, differences in map resolution and class enumeration, and fundamental differences in each group’s approach to urban land. Timing matters because the input data sets for these urban maps were collected from the mid-1990s to late 2005. In regions such as East Asia, urban expansion is proceeding so rapidly that only a few years can make a large difference on the ground. Scale and resolution are of concern because all of these maps combine coarse resolution inputs with binary classifications (urban/non-urban classes). Because urban land does not occur in neat 3000 arcsecond square blocks, there are considerable problems inherent in any attempt to infer total areal extent from coarse-resolution binary classifications (Latifovic and Olthof 2004; Ozdogan and Woodcock 2006). Of the aforementioned sources of inter-map variance, perhaps the most important is a fundamental divergence in each group’s approach to the modeling of urban land. The problem of creating a meaningful and workable definition of ‘urban’ is not trivial. In the absence of a clear set of definitions, each group constructs an implicit model of urban land that can be inferred from their methodologies. The six global urban maps that emerge are sensitive to many attributes commonly associated with urban land, including: high population density, extensive builtenvironment, electrification, and proximity to transportation infrastructure. The degree to which any one of these attributes contributes to an urban classification is likely regionally-dependent and is not specified by any of the mapping groups. In order to characterize what each global map means when it classifies an area as urban, this comparative study is not sufficient. A global assessment of these maps is the subject of ongoing research. We are conducting a multi-resolution assessment using collections of regional and city-scale validation data, including medium resolution maps of impervious surface area, fine resolution satellite imagery, and several ground truth campaigns. This ongoing project will allow us to make more meaningful statements about map accuracy, geo-registration, and the underlying models of urban land implicit within each
123
GeoJournal
global urban map. The aim of the assessment will be to assemble a suite of composite urban maps that are tailored to particular user-groups. This map fusion exercise should allow us to address specific needs of users by first devising a group-specific definition of urban land, and next fusing components of existing global urban maps (and urban-related maps) that best meet the requirements of those definitions. Acknowledgments This research was made possible in part by the Department of Energy Computational Science Graduate Fellowship (GSGF) and Princeton University’s Office of Population Research (OPR). We’d like to thank Professor Shlomo Angel of New York University for his thoughtful comments and suggestions at all stages of this work. Thanks also to Professors Doug Massey and Burt Singer of Princeton’s Office of Population Research for their comments, as well as the recommendations of three anonymous reviewers.
References Alberti, M. (2005). The effects of urban patterns on ecosystem function. International Regional Science Review, 28(2), 168–192. Angel, S., Sheppard, S. C., & Civco D. L. (2005). The dynamics of global urban expansion. Washington, DC: The World Bank, available online at: http://www.williams.edu/ Economics/UrbanGrowth/WorkingPapers.htm, last accessed April 15, 2007. Bartholome, E., & Belward, A. S. (2005). GLC2000: A new approach to global land cover mapping from Earth observation data. International Journal of Remote Sensing26(9), 1959–1977. Bhaduri, B., Bright, E., Coleman, P., & Dobson, J. (2002). LandScan: Locating people is what matters. Geoinfomatics, 5, 34–37. Bourne, J. (2000). Louisiana’s vanishing wetlands: Going, going…. Science, 289, 1860–1863. Brockerhoff, M. (2000). Urbanizing world. Population Bulletin, 55, 3. Brockerhoff, M, & Brennan, E. (1998). The poverty of cities in developing regions. Population and Development Review, 24(1), 75–114. Burgess, R. (2000), The compact city debate: A global perspective. In M. Jenks & R. Burgess (Eds.), Compact cities: Sustainable urban forms for developing countries. London and New York: Spon Press. Calbo, J., Pan, W., Webster, M., Prinn, R. G., & McRae, G. J. (1998). Parameterization of urban sub-grid scale processes in global atmospheric chemistry models. Journal of Geophysical Research, 103, 3437–3467. Castells, M. (1996). The rise of the network society. Malden, Massachusetts: Blackwell Publishers. Center for International Earth Science Information Network (CIESIN), Columbia University; International Food Policy Research Institute (IFPRI); The World Bank; and Centro Internacional de Agricultura Tropical (CIAT)
123
(2004). Global Rural–Urban Mapping Project (GRUMP), Alpha Version: Urban Extents. Socioeconomic Data and Applications Center (SEDAC), Columbia University. Palisades, New York. Available at http://www.sedac. ciesin.columbia.edu/gpw, last accessed April 15, 2007. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37–46. Cohen, B. (2004). Urban growth in developing countries: A review of current trends and a caution regarding existing forecasts. World Development, 32(1), 23–51. Danko, D. M. (1992). The digital chart of the world project. Photogrammetric Engineering and Remote Sensing, 58(8), 1125–1128. Davies, R., Orme, C., Olson, V., Thomas, G., Ross, S., Ding, T., Rasmussen, P., Strattersfield, A., Bennett, P., Blackburn, T., Owens, I., & Gaston, K. (2006). Human impacts and the global distribution of extinction risk. Proceedings of the Royal Society B, 273, 2127–2133. Dobson, J. E., Bright, E. A., Coleman, P. R., Durfee, R. C., & Worley, B. A. (2000). Landscan: A global population database for estimating populations at risk. Photogrammetric Engineering and Remote Sensing, 66(7), 849–857. Doll, C. N. H., Muller, J. P., & Morley, J. G. (2006). Mapping regional economic activity from night-time light satellite imagery. Ecological Economics, 57(1), 75–92. Douglas, I. (1994). Human settlements. In W. B. Meyer & B. L. Turner II (Eds.), Changes in land use and land cover: A global perspective. Cambridge, UK: Cambridge University Press. Ehrlich, P. R. (1991). Population diversity and the future of ecosystems. Science, 254, 175. El Araby, M. (2002). Urban growth and environmental degradation. Cities, 19, 389–400. Elvidge, C. D., Imhoff, M. L., Baugh, K. E., Hobson, V. R., Nelson, I., Safran, J., Dietz, J. B., & Tuttle, B. T. (2001). Nighttime lights of the world: 1994–95. ISPRS Journal of Photogrammetry and Remote Sensing, 56(2), 81–99. Elvidge, C., Milesi, C., Dietz, J., Tuttle, B., Sutton, P., Nemani, R., & Vogelmann, J. (2004). US constructed area approaches the size of Ohio. EOS: Transactions of the American Geophysical Union, 85(24), 233. Folke, C., Jansson, A., Larsson, J., & Costanza, R. (1997). Ecosystem appropriation by cities. Ambio, 26, 167–172. Friedl, M., McIver, D., Hodges, J., Zhang, X., Muchoney, D., Strahler, A., Woodcock, A., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao, F., & Schaaf, C. (2002). Global land cover mapping from MODIS: Algorithms and early results. Remote Sensing of Environment, 83, 287–302. Giri, C., Zhu, Z. L., & Reed, B. (2005). A comparative analysis of the Global Land Cover 2000 and MODIS land cover data sets. Remote Sensing of the Environment, 94(1), 123–132. Goldewijk, K. (2001). Estimating global land use change over the past 300 years: The HYDE database. Global Biogeochemical Cycles, 15(2), 417–434. Goldewijk, K. (2005). Three centuries of global population growth: A spatially referenced population density database for 1700–2000. Population and Environment, 26(5), 343–367.
GeoJournal Goldewijk, K., & Ramankutty, N. (2004). Land cover change over the last three centuries due to human activities: The availability of new global data sets. Geojournal, 61, 335–344. Global Land Cover Facility, website: http://www.glcf.umiacs. umd.edu. Greyner, R., Orme, C., Jackson, S., Thomas, G., Davies, R., Davies, T., Jones, K., Olson, V., Ridgely, R., Rasmussen, P., Ding, T., Bennett, P., Blackburn, T, Gaston, K., Gittleman, J., & Owens, I. (2006). Global distribution and conservation of rare and threatened vertebrates. Nature, 444(2), 93–96. Gru¨bler, A. (1994). Technology. In W. B. Meyer & B. L. Turner II (Eds.), Changes in land use and land cover: A global perspective. Cambridge, UK: Cambridge University Press. Hansen, M. C., & Reed, B. (2000). A comparison of the IGBP DISCover and University of Maryland 1 km global land cover products. International Journal of Remote Sensing, 21, 1365–1373. Henderson, M., Yeh, E. T., Gong, P., Elvidge, C., & Baugh, K. (2003). Validation of urban boundaries derived from global night-time satellite imagery. International Journal of Remote Sensing, 24(3), 595–609. Herold, M., Woodcock, C., Gregorio, A., Mayaux, P., Belward, A., Latham, J., & Schmullius, C. (2006). A joint initiative for harmonization and validation of land cover datasets. IEEE Transactions on Geosciences and Remote Sensing, 44(7), 1719–1727. Imhoff, M. L., Lawrence, W. T., Stutzer, D. C., & Elvidge, C. D. (1997). A technique for using composite DMSP/OLS ‘‘City Lights’’ satellite data to accurately map urban areas. Remote Sensing of Environment, 61, 361–370. Intergovernmental Panel on Climate Change. (2007). Working Group II Report: Climate Change 2007, Impacts, Adaptation and Vulnerability—Summary for Policymakers, IPCC Secretariat, Geneva, Switzerland. Jung, M., Henkel, K., Herold, M., & Churkina, G. (2006). Exploiting synergies of global land cover products for carbon cycle modeling. Remote Sensing of Environment, 101, 534–553. Kaye, J., Groffman, P. M., Grumm, N. B., Baker, L. A., & Pouyat, R. V. (2006). A distinct urban biogeochemistry? Trends in Ecology and Evolution, 21(4), 192–198. Keiser, J., Utzinger, J., De Castro, M. C., Smith, T. A., Tanner, M., & Singer, B. H. (2004). Urbanization in sub-Saharan Africa and implication for malaria control. American Journal of Tropical Medicine and Hygiene, 71(2), 118–127. Kendall, M. (1938). A new measure of rank correlation. Biometrika, 30, 81–89. Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics33, 169–174. Latifovic, R., & Olthof, I. (2004). Accuracy assessment using sub-pixel fractional error matrices of global land cover products derived from satellite data. Remote Sensing of Environment, 90, 153–165. Loveland, T. R., Reed, B. C., Brown, J. F., Ohlen, D. O., Zhu, J, Yang, L., & Merchant, J. W. (2000). Development of a global land cover characteristics database and IGBP
DISCover from 1-km AVHRR data. International Journal of Remote Sensing, 21, 1303–1330. Massey, D. (1996). The age of extremes: Concentrated affluence and poverty in the twenty-first century. Demography, 33(4), 395–412. Massey, D. (2005). Strangers in a strange land: Humans in an urbanizing world. New York, NY: W.W. Norton & Company. Mayaux, P., Eva, H., Gallego, J., Strahler, A. H., Herold, M., Agrawal, S., Naumov, S., Miranda, E., Bella, C., Ordoyne, C., Kopin, Y., & Roy, P. (2006). Validation of the Global Land Cover 2000 map. IEEE Transactions on Geoscience and Remote Sensing, 44(7), 1728–1739. McGranahan, G., & Satterthwaite, D. (2000). Environmental health of ecological sustainability: Reconciling the brown and green agendas in urban development. In C. Pugh (Ed.), Sustainable cities in developing countries: Theory and practice at the millennium (pp. 53–72). London, UK: Earthscan Publishers. Milesi, C., Elvidge, C. D., Nemani, R. R., & Running, S. W. (2003). Assessing the impact of urban land development on net primary productivity in the southeastern United States. Remote Sensing of the Environment, 86(3), 401–410. Monserud, R, & Leemans, R. (1992). Comparing global vegetation maps with the Kappa statistic. Ecological Modeling, 62, 275–293. Montgomery, M., Stren, R., Cohen, B., & Reed, H. (2003). Cities transformed: Demographic change and its implications in the developing world. Washington, DC: National Academies Press. National Geophysical Data Center. (2007). Nighttime Lights data available at: http://www.ngdc.noaa.gov/dmsp/ sensors/ols.html. Oke, T. R. (1982). The energetic basis of the urban heat island. Quarterly Journal of the Royal Meteorological Society, 108, 1–24. Ozdogan, M., & Woodcock, C. E. (2006). Resolution dependent errors in remote sensing of cultivated areas. Remote Sensing of the Environment, 103(2), 203–217. Peters-Lidard, C. D., Kumar, S., Tian, Y., Eastman, J. L., & Houser, P. (2004). Global urban-scale land-atmosphere modeling with the land information system. Symposium on Planning, Nowcasting, and Forecasting in the Urban Zone, 84th American Meteorological Society Annual Meeting, 11–15 January 2004, Seattle, Washington, USA. Pickett, S., Burch, W., Dalton, S., Foresman, T., Grove, M., & Rowntree, R. (1997). A conceptual framework for the study of human ecosystems in urban areas. Urban Ecosystems, 1, 186–199. Quigley, J. M. (1998). Urban diversity and economic growth. Journal of Economic Perspectives, 37, 426–434. Rees, W. E. (1992). Ecological footprint and appropriated carrying capacity: What urban economics leaves out. Environment and Urbanization, 4, 121–130. Sahr, K., White, D., & Kimerling, A. (2003). Geodesic discrete global grid systems. Cartography and Geographic Information Science, 30(2), 121–134. Sassen, S. (1994). Cities in a world economy. Thousand Oaks, CA: Pine Forge-Sage Press. Schneider, A., Friedl, M. A., Mciver, D. K., & Woodcock, C. E. (2003). Mapping urban areas by fusing multiple
123
GeoJournal sources of coarse resolution remotely sensed data. Photogrammetric Engineering and Remote Sensing69(12), 1377–1386. Schneider, A., Friedl, M. A., & Woodcock, C. E. (2005). Mapping urban areas by fusing multiple sources of coarse resolution remotely sensed data: Global results. In Proceedings of the 5th International Symposium of Remote Sensing of Urban Areas, 14–16 March, Tempe, Arizona. Schneider, A., & Woodcock C. E. (in press). Compact, dispersed, fragmented, extensive? A comparison of urban expansion in 25 global cities using remotely sensed data, pattern metrics and census information. Urban Studies. See, L. M., & Fritz, S. (2006). A method to compare and improve land cover datasets: Application to the GLC2000 and MODIS land cover products. IEEE Transactions in Geoscience and Remote Sensing, 44(7), 1740–1746. Small, C., & Cohen, J. (2004). Continental physiography, climate, and the global distribution of human population. Current Anthropology, 45(2), 269–277. Small, C., Pozzi, F., & Elvidge, C. D. (2005). Spatial analysis of global urban extent from DMSP-OLS night lights. Remote Sensing of the Environment, 96, 277–291. Stren, R., White, R., & Whitney, J. (1992). Sustainable cities: Urbanization and the environment in international perspective. Boulder, CO: Westview Press. Sutton, P., Roberts, D., Elvidge, C., & Meij, H. (1997). A comparison of night-time satellite imagery and population density for the continental United States. Photogrammetric Engineering and Remote Sensing, 63, 1303–1313. Tatem, A. J., Noor, A. M., & Hay, S. I. (2005). Assessing the accuracy of satellite derived global and national urban maps in Kenya. Remote Sensing of the Environment, 96(1), 87–97. Travis, J. (2005). Scientists’ fears come true as hurricane floods New Orleans. Science, 309, 1656–1659. Tucker, C. J., Grant, D. M., & Dykstra, J. D. (2004). NASA’s global orthorectified Landsat data set. Photogrammetric Engineering and Remote Sensing, 70, 313–322.
123
Tuttle, B. (2007). Global mapping of impervious surface area. San Francisco, CA: 2007 Annual Meeting of the Association of American Geographers. Unger, J., Sumeghy, Z., Bottyan, Z., & Musci, L. (2001). Land use and meteorological aspects of the urban heat island. Meteorological Applications, 8, 189–194. United Nations Human Settlements Programme (UN-HABITAT). (2003). The challenge of slums: Global report on human settlements, 2003 (310 pp.). London, United Kingdom: Earthscan Publications. UN Population Division. (2005). United Nations World Urbanization Prospects—The 2005 Revision, online at http://www.esa.un.org/unup, last accessed April 15, 2007. UN Statistics Division (2007). http://www.unstats.un.org/unsd/ methods/m49/m49.htm. Utzinger, J., & Keiser, J. (2006). Urbanization and tropical health – then and now. Annals of Tropical Medicine and Parasitology, 100(5–6), 517–533. Van Vliet, W. (2002). Cities in a globalizing world: From engines of growth to agents of change. Environment and Urbanization, 14, 31–40. Welch, R. (1980). Monitoring urban population and energy utilization patterns from satellite data. Remote Sensing of Environment, 9, 1–9. Wratten, E. (1995). Conceptualizing urban poverty. Environment and Urbanization, 7, 11–38. Yang, L., Huange, C., Homer, C., Wylie, B., & Coan, M. (2003). An approach for mapping large-area impervious surfaces: Synergistic use of Landsat 7 ETM+ and high spatial resolution imagery. Canadian Journal of Remote Sensing, 29(2), 230–240. Zhang, X. Y., Friedl, M. A., Schaaf, C. B., Strahler, A. H., & Schneider, A. (2004). The footprint of urban climates on vegetation phenology. Geophysical Research Letters, 31(12), L12209.
