Projecting Urban Land Cover on the basis of Population Dan Miller Runfola Postdoctoral Researcher National Center for Atmospheric Research CGD & RAL CU:Boulder Institute of Behavioral Science 1
The Challenge Need to know where urban areas are, spatially Urban Impacts Already developed approaches to projecting future population spatially (Bryan Jones) How (or can) we use this population data to project future changes in urban land area? 2
Current Products Urban Land Name Organization Temporal Resolution GRUMP CIESIN 1995 1km Global Land Cover UMD Circa 1990 1km MODIS MCD12Q1 NASA Annual 2001+ 500m GlobCover ESA Circa 2005,2009 300m NOAA Global Impervious NOAA Circa 2000 1km CLM (Jackson) NCAR 2004 1km Spatial Resolution (approx) 3
Product Comparison Optimally, all urban datasets would agree on where urban land is. To aid in projecting future urban land cover, we also hope that urban land and population are closely correlated. 4
Comparison of Urban Products - Visual In this set of slides: Red = Urban / Built Up / Artificial Blue = Water Mango = Everything Else
GlobCover 2004
GRUMP (1995)
NOAA (Circa 2000)
UMD (AVHRR Circa 1990)
MODIS (2001)
Quantity of Urban Land Cover: Comparison GlobCover JACKSON UMD NOAA MODIS GRUMP 0 5 10 15 20 Average Percent of Municipality that is Urban 11
Can we replicate the current approach? No clear correct urban product, so Can we replicate the currently utilized approach? Jackson population, regionally-based thresholds. 12
Challenges to This Approach Population projections at a coarser scale than base-year data used to create Jackson Population base data used in projections has differing definitions Continuous vs. Hard Classified Approach 13
LandScan - 2001
GPW (CIESIN) Unadjusted 2000
2.5 Medium Density
Low Density Resolution Average R^2 Minimum R^2 Maximum R^2 2.5.21.05.51 7.5.26.11.59 15.35.13.71 Medium Density Resolution Average R^2 Minimum R^2 Maximum R^2 2.5.33.07.66 7.5.38.19.76 15.45.16.76
High Density Resolution Average R^2 Minimum R^2 Maximum R^2 2.5.37.01.63 7.5.36.03.66 15.39.06.70 Tall Business District Resolution Average R^2 Minimum R^2 Maximum R^2 2.5.08.0001.31 7.5.10.001.47 15.15.007.52
Conclusions GPW can be used to approximate the Landscan/Jackson approach, but Only works in some regions / for some resolutions Highly variable accuracy (Range)
Where to from here? GPW-based thresholds for three countries of immediate interest: Brazil, China, India Test using calibration and validation cities independently to confirm thresholding accuracy Do this for two points in time (1990, 2000) to determine how well threshold temporal stationarity
Potentially Relevant Detours
What definitions of urban perform reasonably over time? (Hint: Not MERIS)
Mexico City Circa
Mexico City Circa 2009
How might resolution influence impacts analysis?
Why Helpful to know the coarsest scale at which finer scales no longer make an important difference (what degree of spatial allocation is helpful?). Robustness is the result wildly different?
First Cut Approach (1) Aggregated three datasets to five different resolutions: Flooding Frequency (CIESIN), Urbanization (Jackson LD), and Population (GPW); 2.5 2.0 1.5 1.0 0.5 degrees. (2) Multiplied population by flooding frequency to get a measure of exposure within each cell (repeated at each resolution). Regressed urbanization against population (repeated at each resolution).
New Data Flooding CIESIN Flooding Dataset Resolution is variable, but roughly 1 degree in most places. On the website they say 2.5 degrees, but that is not what is contained in the data. Data is based on observed flood frequency, 1985-2003. Deciles (each cell has a 1-10) For each cell, calculated the area (in square degrees) that is attributed to each decile. Sum-Product to get the final results (sum of all deciles*area). Can potentially result in values > 10.
Raw Data (1 Degree)
How might resolution influence Findings: impacts analysis? Resolution matters for flooding impacts, but the coarse resolution of our flooding data (~0.5 degree cells) prevents us from looking at the resolutions we might be most interested in conducting such an analysis at. Uncertain if the same result might hold for less spatially-manifest events (i.e., tornados instead of floods).
Other Materials
Comparison of Urban Products - Visual In this set of slides: Red = Urban / Built Up / Artificial Blue = Water Mango = Everything Else
GlobCover 2004
GRUMP (1995)
NOAA (Circa 2000)
UMD (AVHRR Circa 1990)
MODIS (2001)
Comparison of Population Products - Visual In this set of slides: Brighter (more white) cells have larger populations.
LandScan - 2001
GPW (CIESIN) Unadjusted 2000
Misc Text for Reference
Jackson 2010 Population density proved to be the most reliable proxy for urban intensity after several satellite-derived landcoverproducts were evaluated for defining urban extent (e.g. MODIS, GLC2000, and DISCover). Not only do these products show significant disagreement on the location and size of urban areas, but they also do not provide levels of urban intensity, which is a required parameter of CLMU. Population density as provided by LandScan already considers slope, landcover, nighttime lights, proximity to roadways, and census data to determine where people are most likely living (Dobson et al. 2000). Population density is not a perfect determinant of urban concentration, but it provides an easy transition to temporal usage so it offers the ability to answer more questions than a satellite based snap-shot of urban areas. With emergence of new population datasets, the time series of population densities lengthens, assisting in the study of urban evolution over time. Using LandScan 2004 population densities (ORNL 2005), a lower limit of population density for each region was selected to define the urban-rural boundary. Once urban areas were defined, they were further delineated into four levels of urban intensity, including tall building district (TBD), high density (HD), medium density (MD), and low density (LD). Population density boundaries for each of these urban levels differ widely to account for different types of urban communities. For example, East Africa contains many small farm plots housing large families. This causes the population density to be relatively high for an agriculture-based community, and it has many characteristics of for example U.S. suburbs,suchas low buildings and a high percentage of green space.
Jackson 2010 continued Intra-urban boundaries (e.g. between low and medium density) were based on population density and observations of satellite imagery in at least ten sample cities per region (i.e. validation cities). Validation cities were chosen based on size (i.e. large urban population), location (i.e. to geographically represent the entire region), and best available resolution of satellite imagery in Google Earth. The images in Google Earth come from a variety of sources (e.g. TerraMetrics, NASA, DigitalGlobe), and the available resolutions vary as well, from 15m in most cases(earthsat 2007) to less than one meter (DigitalGlobe 2007). During comparisons of LandScan population density (Dobsen et al. 2000) to satellite images natural boundaries typically presented themselves in the form of a relatively large disparity of neighboring LandScan pixels corresponding to a change in urban land surface properties(e.g. vegetated fraction, building density, etc.) as seen in Google Earth. For each region, boundaries were assigned based on an initial validation city, then adjusted for each subsequent validation city. The initial cities were then re-checked to assure the adjustments still properly represented their intra-urban boundaries.
Jackson 2010 continued To maintain consistency from region to region during this process, definitions for the urban categories were established from the start. Tall building districts (TBD) here are defined as an area of at least 1 km2 with buildings greater than or equal to 10 stories tall, with a small fraction of vegetation (i.e. 5-15 percent of plan area). Many cities that may appear to have a TBD were not included in this dataset because the aerial extent was too small (i.e. less than 1 km2) or because the population density of the TBD was too sparse. High density (HD) areas may encompass commercial, residential, or industrial areas and are characterized by buildings 3-10 stories tall with a vegetated/pervious fraction typically in the range of 5-25 percent. Medium density (MD) areas are usually characterized by row houses or apartment complexes 1-3 stories tall with a vegetated/pervious fraction of 20-60 percent. Finally, the low density (LD) category covers areas with 1-2 story buildings and a vegetated/pervious fraction of 50-85 percent. LD includes a variety of urban types, from the suburbs of the United States to urban agricultural parts of East Africa. As a final step in defining urban extent, an inter-region comparison looked at disparities on the global scale such as population percentile rank of each boundary. This ensured that regions with similar qualities were consistent in their assignment of urban categories. In this process, 13 minor adjustments were made to the urban-rural and intra-urban boundaries to maintain consistency in urban levels from region to region
Global Maps
GRUMP
NOAA
GlobCover
AVHRR (UMD)
Jackson (no low density)