Drivers of Land-Use/Land-Cover Changes

and Dynamic Modeling for the

Atlanta, Georgia Metropolitan Area
C.P. Lo and Xiaojun Yang
Abstract
Landsat images and census data were integrated in a zone-based
cellular approach to analyze the drivers of land-uselland-cover
changes in Atlanta, Georgia, a postmodern metropolis. Land-
uselland-cover statistics, which were extracted from Landsat
MSS, TM, and ETM+ images for 1973,1979,1987,1993, and 1999
for the 2 3 metro counties of the Atlanta metropolitan area,
revealed rapid increases in high-density and low-density urban
use at the expense of cropland and forests during this period
of rapid population growth. To understand the underlying causes
of all these changes, demographic and socio-economic data
from the censuses were integrated with the land-uselland-cover
change data and location data. A total of 17 themes comprising
78 variables were used in the analysis at three different spatial
levels: the whole metropolis, county, and census tract, all unified
at the 60-meter grid-cell level. It was found that proximity to
highways, nodes, and shopping malls tended to promote urban
development in Atlanta, and the increasing affluence of the
population has induced rapid suburbanization, with con-
sequent adverse impact on the greenness and fragmentation of
the environment in recent years. The results of the driving force
analysis were incorporated into a dynamic model, namely,
cellular automaton, at the census tract level, which simulated
the land-uselland-cover change of Atlanta from 1999 to 2050.
It predicted the continued growth of edge cities and the loss of
forest, if unchecked, within a time span of 10 to 20 years. The
limitations of the cellular automaton model as applied to Atlanta
were also discussed.
Introduction
For the past three decades, the city of Atlanta, Georgia has expe-
rienced very rapid growth both in terms of population and spa-
tial extent as it emerged to become the premier commercial,
industrial, and transportation center of the southeast
(Research Atlanta, 1993). Population has increased 27 percent
from 1970 to 1980,33 percent from 1980 to 1990, and 40 per-
cent from 1990 to 2000. The city expanded outward at the
expense of crop and forest land. This has given rise to urban
sprawl along highways radiating from the city center. Research
on Atlanta's internal structure led to the formulation of the
urban realms model to depict the multi-nuclei nature of the city
in contrast to the usual single-core urban form of many Ameri-
can cities (Hartshorn and Muller, 1989; Fujii and Hartshorn,
1995). Edge cities are formed at the intersection of an urban
beltway and a hub-and-spoke lateral road (Garreau, 1991). All
these are characteristics of the postmodern city (Dear, 2000). A
study of the driving force of land-uselland-cover changes of
Atlanta will contribute to a better understanding of the proc-
esses of urban sprawl and their spatial consequences under
postmodern urbanism.
The study of drivers of land-uselland-cover change in the
past has focused primarily on biophysical variables, such as
elevation, slope, or soil type. However, it has been increasingly
realized that data on socio-economic drivers of change have to
be incorporated (Veldkamp and Lambin, 2000). The integration
of biophysical and socio-economic data has been hampered by
the spatial unit problem because the relevant spatial units for
biophysical processes are different from the spatial units of
decision making, on which most socio-economic data are based
(Martin, 1996).
In this paper, the drivers of land-uselland-cover changes in
Atlanta are studied using Landsat Multispectral Scanner
(MSS), Thematic Mapper (TM), and Enhanced Thematic Mapper
Plus (ETM+) data for 1973,1979,1987,1993, and 1999 supple-
mented by socio-economic data obtained from the United
States Census Bureau for 1970,1980, and 1990 at three spatial
levels: the whole Atlanta metropolitan area as delimited by the
13 urban counties, the urban counties, and census tracts, in
view of the fact that drivers of land-use change are scale depen-
dent (Veldkamp and Lambin, 2001). The findings are then
employed in a process-based cellular automata (CA) modeling
to simulate the urban growth of Atlanta up to the year 2050. CA
modeling allows a spatially explicit display of land-use deci-
sions made by planners. In applying to an urban environment,
this research aims to demonstrate a new approach to drivers of
land-uselland-cover change analysis, and the use of dynamic
modeling in solving the spatial unit problem commonly
encountered in data integration.
Land-Use/Land-Cover Change Detection from Landsat Data
Three generations of Landsat data were used to extract land-
uselland-cover information of Atlanta for 1973 (MSS), 1979
(MSS), 1987 (TM), 1993 (TM), and 1999 (ETM+) at roughly six-
year intervals. The Landsat data have been preprocessed by
geometric rectification and radiometric normalization (Yang
and Lo, 2000). The MSS and TM data were resampled to a spatial
resolution of 57 meters and 25 meters, respectively. A six-class
land-uselland-cover classification scheme was adopted: (1)
high-density urban use, (2) low-density urban use, (3)
cultivatedlexposed land, (4) croplandlgrassland, (5) forest,
- -
C.P. Lo is with the Department of Geography, University of
Georgia, Athens, GA 30602 (chpanglo@uga.edu).
X. Yang is with Environmental Studies, The University of West
Florida, Pensacola, FL 32514 (xyang8uwf.edu).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
-
Photogrammetric Engineering & Remote Sensing
Vol. 68, No. 10, October 2002, pp. 1073-1082.
0099-1112/02/681&1073$3.00/0
O 2002 American Society for Photogrammetry
and Remote Sensing
Oc t o b e r 2002 1073
TABLE 1. LANDUSE/LANDCOVER CLASSES AND DEFINITIONS
-
No. Classes Definitions
High-Density Urban Use Approximately 80 to 100 percent construction materials, e.g., asphalt, concrete, etc.; typically commercial
and industrial buildings with large open roofs as well as large open transportation facilities, e.g., large
airports, parking lots, and multilane interstatelstate highways; with low percentage of residential development
residing i n the city cores.
Low-Density Urban Use Approximately 50 to 80 percent construction materials; often residential development including mostly
singlelmultiple family houses and public rental housing estate as well as local roads and small open (transi-
tional) space as can be always found in a residential area; with certain amount of vegetation cover (up to
20 percent).
Cultivated/Exposed Land Areas of sparse vegetation cover (less than 20 percent) that are likely to change or be converted to other uses
in the near future; including clearcuts, all quarry areas, cultivated land without crops, and barren rock or
sand along riverlstream beaches.
CroplandIGrassland Characterized by high percentages of grasses, other herbaceous vegetation, and crops; including lands that
are regularly mowed for hay andlor grazed by livestock, golf courses and city parks, and regularly tilled and
planted cropland.
Forest Including coniferous, deciduous, and mixed forests (90 to 100 percent).
Water All areas of open water, generally with greater than 95 percent cover of water, including streams, rivers,
lakes, and reservoirs.
and (6) water. Table 1 provides definitions of these six classes of
land-uselland-cover based on their image characteristics. An
unsupervised image classification approach, known as ISO-
DATA (Iterative Self-organizing Data Analysis) was adopted,
which produced natural clusters of homogeneous pixels. The
overall accuracies of classification for the five land-uselland-
cover maps produced varied between 87 percent and 90 per-
cent , while producer's and user's accuracies were well above
80 percent (Yang and Lo, 2002). These are acceptable accura-
cies. It is clearly revealed from the land-uselland-cover statis-
tics extracted from these five maps that between 1973 and 1999
the areas for both the high-density urban use and low-density
urban use have increased while those for cropland and forest
land have declined (Table 2). The increase has been particu-
larly impressive for the low-density urban use (hom 76,910 ha
in 1973 to 282,959 ha in 1999, or an increase of about 268 per-
cent in 26 years!) The low-density urban use consists mainly of
suburban residential housing. As a result of the urban sprawl,
almost 37 percent of cropland and 27 percent of forest land
have been lost.
Method of Driving Force Analysis
Data Preparation
The land-uselland-cover changes in Atlanta as noted above
occurred as a result of the interactions of a number of environ-
mental as well as demographic, social, and economic forces.
Therefore, integrating biophysical variables as extracted from
the Landsat images for individual pixels in raster format with
area unit based socio-economic data from the decennial cen-
suses in vector format is needed. For this study, a zonal
approach was adopted. A zonal file with the appropriate areal
unit, in either vector or raster format, is produced. This file is
used to extract characteristics of each data layer zone by zone
by means of cross-tabulation. Thus, a zone-based table was cre-
ated in which each row is a zone identification number and
each column is a theme or a data layer associated with each
zone. Such a table can be easily analyzed using GIS built-in spa-
tial analysis functionality or imported to a stand-alone statisti-
cal software package for more advanced analysis. In this study,
two basic areal units were employed: counties and census
tracts, so that the drivers of land-uselland-cover changes at two
different spatial scales can be compared. One notable limita-
tion of this approach is the modifiable areal unit problem
(MAUP) (Openshaw, 1984; Green and Flowerdew, 1996).
A total of 17 data layers or themes were used in this analy-
sis. They can be grouped into six major categories: (1)
administrativelstatistical boundaries, (2) land-uselland-cover
data, (3) landscape ecological measures, (4) topographic meas-
ures, (5) population and income, and (6) location measures.
The extraction of each specific data layer under each category
is briefly explained below.
AdministrativelStatistical Boundaries
The Atlanta metropolitan area under study is covered by 13
counties whose boundaries can be extracted from the 1992
TIGER line files. The boundaries of 339 census tracts for 1980
were extracted from the 1980 Census database published by
GeoLytics, Inc. The boundaries of 444 census tracts for 1990
were extracted from the 1995 TIGER street centerline files. All
these boundary data have a nominal scale of 1:100,000. These
are vector data and have been converted to raster data with a
resolution of 60 meters, which is compatible with the spatial
resolution of Landsat MSS data used in land-uselland-cover
mapping.
TABLE 2. ~ANDUSE/LANDCOVER STATISTICS FOR THIRTEEN METRO COUNTIES IN ATLANTA: 1973-1999
1973 1979 1987 1993 1999
No Land-UseILand-Cover Area (ha) % Area (ha) % Area (ha) % Area (ha) % Area (ha) %
1 High-density urban 29722 2.85 38015 3.64 54280 5.2 67633 6. 48 87477 8. 3
2 Low-density urban 76910 7.36 129174 12.37 177825 17.03 214484 20.54 282959 27. 1
3 Cultivated/exposed land 14534 1. 39 20595 1.97 15511 1. 49 21132 2.02 5358 0.51
4 Croplandlgrassland 159345 15.26 117365 11.24 117686 11.27 96700 9.26 101122 9.68
5 Forest land 750366 71.85 724967 69.42 663673 63.55 625984 59.94 545148 52.2
6 Water 13404 1. 28 14166 1.36 15306 1.47 18348 1.76 22217 2.13
In Total 1044281 100 1044281 100 1044281 100 1044281 100 1044281 100
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Land-UselLand-Cover Data
Only five classes of land-uselland-cover, namely, high-density
urban use, low-density urban use, cropland/grassland, forest,
and water for 1973,1987, and 1999 were used for this analysis.
Each land-uselland-cover class was converted to a binary mask
of 0 (for the background) and 1 (for the specific land-uselland-
cover type). This has resulted in 15 layers of data. The land-
uselland-cover data layers have also been resampled to a spa-
tial resolution of 60 meters.
Landscape Ecological Measures
TWO indices were computed: (a) greenness and (b) fragmenta-
tion. Greenness was extracted from the visible and near-infra-
red bands of the Landsat images, using the Tasseled Cap
Transformation (TCT). The second component of TCT is green-
ness. The formulas used for this computation were (Crist and
Kauth, 1986; Jensen, 1995):
For Landsat Ms s data:
Greenness = -0.2830"bl - 0.6600kb2 + 0.5770*b3
For Landsat TM and Em+ data:
Greenness = -0.2848*bl - 0.2435*b2 - 0.5436*b3
where bl, b2, ..., b7 are the band numbers. It should be noted
that, for the Landsat TM and ETM+ data, the thermal infrared
band (b6) was not used in the computation of greenness.
Fragmentation measures the tendency of the land cover to
break up into many small patches, which can affect species
diversity and density of animal species. It is therefore a meas-
ure of ecological quality of a habitat. The following formula
was used to compute fragmentation using a kernel-based
approach (Monmonier, 1974):
Fragmentation = (c - l)l(k - 1) (3)
where ci s the number of different land-uselland-cover classes
present in the kernel, and ki s the number of cells considered.
A 7 by 7 kernel was used for this computation so that k equaled
49.
Greenness and fragmentation were calculated for all the
three years of study at the spatial resolution of 60 meters.
Topographic Measures
Two topographic measures, namely, terrain elevation and
slope, were obtained from the USGS 7.5-minute Digital Elevation
Model (DEM). The 159 DEMs that cover the study area were
mosaicked to form a single image with a 30-meter pixel size.
The resulting DEM mosaic exhibited some discontinuities. To
produce a seamless mosaic, a 5 by 5 median filter was applied
and a mask image was created in which the discontinuities
were filled with the median values of the surrounding pixels.
This mask image was overlaid back to the areas of the original
DEM mosaic where the discontinuities occurred. The improved
DEM mosaic was then resampled to a spatial resolution of 60
meters. Each pixel of the DEM gives the elevation of the terrain in
feet. A terrain slope (in percent) was also calculated using the
built-in function of ERDAS Imagine (ERDAS, 1997) .
Population and Income
Population and income at county and census tract levels were
extracted using different approaches. At the county level, cen-
sus data for 1970,1980, and 1990 and population estimates for
2000 were obtained from the U.S. Bureau of Census. The popu-
lation and income for the three study years for each of the 13
counties, 1973,1987, and 1999, were interpolated assuming
that population and per capita income increased at the same
rate during a decade in a specific county. Thus, for the estimated
1987 population in a county, the following sequence of equa-
tions was used:
where pop80 and pop90 are the known total population for
1980 and 1990, respectively, in the county, and xis the annual
rate of population increase, which can be computed using the
following formula:
where In is natural logarithm and e is the mathematical con-
stant, namely, 2.718281828459. Once xi s known, the total pop-
ulation in 1987 (pop87) for the county was estimated as
The per capita income for the 13 counties (pci) was computed
using the following formula:
13
2 (pcii * popj)
pci = '=I
POP
where i denotes each of the 13 counties.
The computation of population and per capita income for
the three different years at census tract level is much more
complicated than that at the county level because of data
unavailability and the changes in census tract number and
boundaries through time. The population and income at the
census tract level are available for the census years of 1970,
1980, and 1990. However, the 1970 data were not used because
of the large number of census tract boundary changes caused
by highway construction and the poor compatibility to the 1980
and 1990 data. Using the 1980 and 1990 data to estimate popu-
lation and income for 1973,1987, and 1999 is not easy because
of differences in the number of census tracts and their bound-
aries. It is important that the same number and shape of census
tracts should be used to ensure comparability of the results of
analysis. While those tracts in 1980 that have been split in 1990
can be re-combined, newly created census tracts cannot be
restored to be compatible with those in 1980.
A dasymetric mapping approach based on the work of
Langford et al. (1991) and Langford and Unwin (1994) was
adopted to harmonize the data with the spatial units used. The
1980 census tract vector layer was converted to two raster lay-
ers at a 60-meter spatial resolution, one showing population
and the other per capita income per pixel. The population and
per capita income for each tract were re-distributed according
to the spatial pattern of low-density urban use (which repre-
sents residential use) of the land-uselland-cover map for 1979,
the closest year to 1980. Pixels in other types of land uselland
cover were assigned a value of zero. A population density layer
was obtained by excluding the zero pixels for each tract in the
calculation. By overlaying the 1990 census tracts layer over the
1980 adjusted population and income layers, the 1980 data
were made to conform to the boundaries of the 1990 census
tracts.
To estimate the population and per capita income for 1973,
1987, and 1999 at the census tract level, the increase rates for
population (x,) and per capita income (xi) for each tract
between 1980 and 1990 were calculatedusing an equation sim-
ilar in form to that of Equation 5. Based on the statistics for the
PHOTOGRAMMETRICENGINEERING& REMOTE SENSING
Oct ober 2002 1075
entire study area, it was found that the annual increase rate of
population tended to rise during 1970-1990, but to slow down
after 1990, while the increase rate of per capita income tended
to be lower during the past three decades. Adjustments to the
yearly increase rates for the 1970s and 1990s were needed,
which involved multiplying scaling factors to the population
and income increase rates, as shown in the following
equations:
where tpop87, tpop73, and tpop99 are estimated population for
1987,1973, and 1999 by census tracts; tpop80 and tpop90 are
census population for 1980 and 1990 by census tracts; tpci73,
tpci87, and tpci99 are estimated per capita income for 1987,
1973,1999 by census tracts; and tpci8O and tpci90 are per cap-
ita income from the 1980 and 1990 census by tracts.
Location Measures
Four location measures, namely, urban centerproximity, high-
way proximity node point proximity, and shopping mall prox-
imity, were generated. They were used only in the analysis at
the census tract level. Inside the Atlanta metropolitan area,
there are urban centers with commercial, retailing, and service
activities. There were a total of 133 such urban centers of vary-
ing sizes within the 13 counties of Atlanta in 1990. Because
they are not equal in importance, buffers of varying radii were
created based on the product of their population and areal
extent. These buffer rings of urban centers were converted into
a binary image with 1 for the buffers and 0 as the background.
Highways and roads have played a very important role in the
formation of edge cities, the spatial expression of urban sprawl.
A hierarchy of highways and roads can be identified in
Atlanta: limited-access divided highways, duty highways, and
local neighborhood roads. The AND Global Highway Database
was acquired to extract these highways and roads for the analy-
sis, which were then updated with satellite images to form
highway layers for the three years of study (1973,1987, and
1999). The relative importance of highways was determined by
the traffic volume and highway classification. Variable width
buffers were then created around the highways to reflect their
relative importance. A binary image was created with 1 for the
area close to highways and 0 for the background. Node point
proximity is represented by highway exits, junctions, or towns
where major highways run across, which are favored sites for
commercial and industrial activities to locate. These node
point data were extracted from the 1998 AND Global Highway
Database. Similar to the other two proximity measures
described above, a weighting system was assigned to create
radius-variable buffer rings around these node points, from
which a binary mask image of 1 as the area close to node points
and 0 as the background was formed. Finally, shopping mall
proximity suggests accessibility to a large population with sub-
stantial buying power. The sizes of the shopping malls were
used to determine the radii of the buffer rings around them.
Three layers of shopping mall proximity were created by dig-
itizing the large shopping mall polygons on-screen from the
1973,1987, and 1999 Landsat images. They were also con-
verted into binary images with 1 as the area close to the malls
and 0 as the background.
Statistical Analysis of RelatIonshIps
The data created basically consist of landscape metrics as well
as demographic, economic, and topographic measures at three
different spatial levels of aggregation: (1) the overall Atlanta
metropolitan area, (2) 13 counties, and (3) 444 census tracts.
They were all subjected to statistical analysis with the objective
of discovering the interrelationships among these variables,
and hence explanations for the driving forces and characteris-
tics of land-uselland-cover changes.
For the entire Atlanta metropolitan area, simple visual
analysis is adequate to reveal trends with three years of data
with 18 variables for one single observation (Table 3). At the
county level, there are 13 observations for 66 variables. At the
census tract level, there are 444 observations for 60 variables.
These variables include population densities, population den-
sity changes, per capita income, and per capita income
changes; mean elevation, mean slope, percentages of county in
urban center, road, node, and shopping mall buffers; propor-
tions in high-density urban use, low-density urban use, crop-
land, forest, and water as well as their changes over time; mean
greenness values, mean landscape fragmentation and their
changes over time, all for the three years of 1973,1987, and
1999. Simple correlation analysis was used, in which Pearson
correlation coefficients were calculated for pairs of the vari-
ables in order to determine if an association exists as well as
the magnitude and direction of the significant association. The
correlation coefficients are deemed statistically significant at
the 0.01 level (2-tailed) or 0.05 level (2-tailed), based on the p
value, the probability that a statistical result as extreme as the
one observed would occur if the null hypotheses were true
(Norusis, 1998). If the observed significance level is less than
0.05 or 0.01, the null hypothesis of no association is rejected.
After the analysis, it was found that there was a larger number
of correlation coefficients that were more statistically signifi-
cant at the census tract level than at the county level, clearly
supporting the impact of MAUP mentioned previously.
For the data at the census tract level, a multivariate regres-
sion was also conducted to examine the relationship between
landscape metric proportions (and their changes) as dependent
variables with a group of independent variables. The stepwise
variable selection was employed during the regression in order
to determine which variable to include in the multivariate
regression. All variables to be entered must pass the tolerance
criterion specified as the probability of F (the square oft value)
less than 0.050. A variable was also not entered if it would
cause the tolerance of another variable already in the regres-
sion model to drop below the tolerance criterion, specified as
the probability of F larger than 0.100. All entered variables
with tolerance larger than the specified level were removed
from the model. A total of 35 multiple regression models was
computed, which relate land-uselland-cover types, landscape
indices, as well as demographic and socio-economic variables
together. The R2 (coefficient of determination) calculated for
each model indicates the proportion of variability of the
dependent variable accounted for by the regression model. It
was adjusted to account for the complexity of the regression
model relative to the complexity of the data. Another parame-
ter, tolerance, was also obtained to measure the strength of the
linear relationships among the independent variables. The
multiple regression models revealed large tolerance values,
indicating that there is little evidence of multicollinearity
among independent variables.
1076 Oct ober 2002
TABLE 3. LANDSCAPE METR~CS AND DEMOGRAPHIC, ECONOMIC, AND TOPOGRAPHIC MEASURES
1973-1987 1987-1999
Items 1973 1987 1999 change in percent change in percent
land-uselland-cover high-density urban 2.846 5.198 8.377 2.352 82.626 3.179 61.159
proportion (%) low-density urban 7.365 17.028 27.096 9.664 131.212 10.068 59.122
cultivated/exposed land 1.392 1.485 0.513 0.094 6.722 -0.972 -65.457
croplandlgrassland 15.259 11.270 9.683 -3.989 -26.144 -1.586 -14.075
forest 71.855 63.553 52.203 -8.302 -11.553 -11.350 -17.859
water 1.284 1.466 2.127 -0.182 14.190 0.662 45.152
mean greenness 27.733 24.867 16.983 -2.866 -10.334 -7.884 -31.705
mean fragmentation index (X20000) 8.731 10.250 10.339 1.519 17.401 0.089 0.869
per capita income (X1000) 5.509 18.368 29.244 12.859 233.413 10.877 59.216
population density (per hectare) 1.614 2.331 3.159 0.717 44.404 0.828 35.543
mean elevation high-density urban 9.578 9.626 9.663 0.048 0.504 0.037 0.381
(100 feet) low-density urban 9.516 9.484 9.432 -0.032 -0.340 -0.052 -0.549
croplandlgrassland 9.653 9.563 9.463 -0.090 -0.931 -0.010 -1.041
forest 9.469 9.494 9.525 0.025 0.263 0.031 0.331
mean slope high-density urban 4.862 4.986 5.557 0.124 2.550 0.571 11.452
(percent) low-density urban 5.509 5.812 6.220 0.303 5.500 0.408 7.020
croplandlgrassland 5.081 5.327 5.277 0.246 4.842 -0.050 -0.939
forest 8.095 8.303 8.542 0.208 2.569 0.239 2.878
Results of Analysis
Drivlng Forces of Land-Use/LanbCover Changes
For the Atlanta metropolitan area as a whole, Table 3 reveals
that the mean landscape fragmentation index and the propor-
tions of high-density urban, low-density urban, and water uses
have increased since 1973, while the mean greenness and the
proportions of croplandlgrassland and forest have decreased.
At the same time, population density and per capita income
have rapidly increased, thus suggesting their impact on the
land-uselland-cover changes and the landscape characteristics.
It is also revealed from Table 3 that the mean elevation for high-
density urban use and forest has increased from 1973 to 1999,
while that for low-density urban use and cropland/grassland
has decreased during the same period. The implication is that
high-density urban use tended to develop on land with higher
elevations, while low-density urban use occupied land with
lower elevations. Much of the urban development occurred in
forested areas at lower elevations, and as a result, only forest
land at higher elevations has not been touched. For all these
urban developments, land with increasingly steeper slopes
was used.
At finer spatial levels, the results of analysis are much
more complicated as the land-uselland-cover changes and
demographic, economic, and terrain characteristics have
become more differentiated. The results of analysis for four
types of land-uselland-cover, namely, high-density urban use,
low-density urban use, cropland/grassland, and forest, are dis-
cussed below at the county and census tract levels. All indi-
cated coefficients of correlation in the discussion are
statistically significant at either the 0.01 or 0.05 level, as
explained above.
High-Density Urban Use
At the county level, the proportions of high-density urban use
for 1973,1987, and 1999 were positively correlated with popu-
lation density (0.92,0.93, and 0.92), urban center proximity
(0.74,0.73, and 0.75), node proximity (0.77,0,72, and 0.74),
mall proximity (0.68,0.73, and 0.73), and highway proximity
(0.66,0.58, and 0.60). The change in the high-density urban
use proportion from 1973 and 1987 was positively correlated
with population density change (0.66). Income and terrain fac-
tors did not exhibit significant statistical relationship with
either the high-density urban use proportions or their changes
through time.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
At the census tract level, more subtle relationships were
revealed. The proportions of high-density urban use for 1973,
1987, and 1999 correlatedpositively with road proximity (0.53,
0.59, and 0.56), node proximity (0.5 1,0.5 7, and 0.54), urban
center proximity (0.26, 0.26, and 0.411, population density
(0.24,0.26, and 0.41), and mean elevation (0.13,0.15, and 0.16),
but negatively with mean slope gradient (-0.39, -0.42, and
(-0.42). The change in high-density urban use proportion
fsom 1973 to 1999 correlated positively with road proximity
(0.38), node proximity (0.36), urban center proximity (0.26),
mean elevation (0.14), and mall proximity (0.10), but nega-
tively with mean slope (-0.30). It appeared that, at the county
level, population was more significant than were location fac-
tors in explaining high-density urban use change. However, at
the census tract level, location and tenrain conditions were
clearly more significant than was population in explaining the
high-density urban use change over time. This is because com-
mercial and industrial developments (the two major compo-
nents of high-density urban use) were attracted to highways,
node points, and urban centers as revealed at the much finer
census tract level than that of county.
Low-Density Urban Use
At the county level, the proportion of low-density urban use
(which basically is residential use) for 1973,1987, and 1999
correlated positively with population density (0.90,0.86, and
0.72). Other variables that correlated positively with low-den-
sity urban use proportion for at least one year included urban
center proximity (1973 and 1987), node proximity (1973 and
1987), mall proximity (1973 and 1987), highway proximity
(1973), and per capita income (1973). The mean slope gradient
was found to correlate negatively with the proportion of low-
density urban use in 1987 and 1999. The change in low-density
urban use proportion from 1973 to 1999 correlated negatively
with urban center proximity (-0.69), mall proximity (-0.68),
node proximity (-0.66), and highway proximity (-0.60).
At the finer census tract level, the proportion of low-den-
sity urban use for 1973,1987, and 1999 correlated positively
with population density (0.79,0.83, and 0.54), node proximity
(0.64,0.51, and 0.281, and highway proximity (0.60,0.41, and
0.15), but negatively with mean slope gradient (-0.43, -0.33,
and -0.24). Other variables that correlated positively with the
low-density urban use proportion for at least one year included
per capita income (1973 and 1987) and mall proximity (1987
and 1999). The change in the proportion of low-density urban
October 2002 1077
use from 1973 to 1999 correlated positively with population
density change (0.53), but negatively with road proximity
(-0.41) and node proximity (-0.37). Thus, population has
been the most important factor for explaining changes in the
proportion of low-density urban use at both the county and
census tract levels. At the census tract level, it is very clear that
highway proximity and node proximity were conducive to the
rapid growth of low-density urban use, manifested by suburban
housing. However, as more suburban housing has been built, its
correlation with the location measures became less important,
thus explaining the decreasing value of the correlation coeffi-
cient with time as well as the negative correlation between the
proportion of 1973 to 1999 low-density urban use change and
these location measures. It is also interesting to note that the
proportion of low-density urban use correlated positively with
per capita income for the years 1973 (0.62) and 1987 (0.16) only.
The weakening correlation in more recent years indicated that
in the beginning suburban housing was affordable only to peo-
ple with higher income, but in recent years, as the general
affluence of the population increased, more and more people
could afford suburban housing.
CroplandlGrassland
This category of land-uselland-cover includes grassland,
which may be used for agricultural or recreational purposes,
such as golf courses and public parks. At the county level, the
proportions of croplandlgrassland for 1987 and 1999 corre-
lated negatively with population density (-0.82 and -0.66),
urban center proximity (-0.70 and -0.62), node proximity
(-0.68 and -0.63), and mall proximity (-0.68 and -0.62).
Changes in the proportion of this class of land-uselland-cover
from 1973 to 1999 were negatively correlated with per capita
income change (-0.64).
At the census tract level, the proportion of croplandlgrass-
land use for 1973,1987, and 1999 correlated negatively with
population density (-0.32, -0.55, and -0.42), node proximity
(-0.39, -0.59, and -0.46), highway proximity (-0.34, -0.53,
and -0.41), and urban center proximity (-0.20, -0.38, and
-0.35). The correlation with per capita income was negative in
1973 and 1987 (-0.36 and -0.341, but positive in 1999 (0.15).
Other variables that correlated negatively with croplandlgrass-
land for at least one year's data included mean elevation (1973
and 1987) and mall proximity (1973 and 1987). The change in
the proportion of croplandlgrassland from 1973 to 1999 corre-
lated positively with road proximity (0.12) and node proximity
(0.15), and was negatively correlated with mean elevation
(-0.20). Thus, population and location appeared to be two sig-
nificant variables associated negatively with the proportion of
croplandlgrassland. Much of the loss of cropland/grassland in
recent years occurred in places far away from major highways
and node points as more land was acquired for suburban devel-
opment. Another interesting observation is that per capita
income correlated negatively with the proportion of cropland1
grassland for 1973 and 1987 (-0.36 and -0.34), but became
positively correlated in 1999 (0.15). The implication is that,
with increased affluence, people have more leisure time, thus
explaining the increase in the areas of golf courses and public
parks in Atlanta in recent years, while the agricultural use of
the cropland/grassland has declined.
Forest
Much of the suburban development took up forest land. Analy-
ses conducted at the county level revealed that the proportion
of forest land in 1973,1987, and 1999 correlated positively with
mean slope gradient (0.60,0.69, and 0.71, respectively), thus
indicating that forests are found at increasingly steeper gradi-
ents as a result of the loss of forest land at the more accessible
gentle slopes for housing development, Population density
exhibited negative correlation with the proportion of forest
land for 1987 and 1999 (-0.76 and -0.69), another indication
of urban encroachment in recent years. The change in forest-
land area from 1973 to 1999 correlated positively with mean
slope gradient (0.70).
At the census tract level, the proportion of forest land in
1973,1987, and 1999 correlated positively with mean slope
gradient (0.54,0.57, and 0.59), but negatively with road prox-
imity (-0.61, -0.64, and -0.57), node proximity (-0.59,
-0.67, and -0.61), population density (-0.55, -0.59, and
-0.60), and urban center proximity (-0.28, -0.38, and -0.36).
In other words, forests are found at unfavorable sites for devel-
opment. The change in forest-land area from 1973 to 1999 cor-
related positively with road proximity (0.35), node proximity
(0.29), and mean elevation (0.19), but negatively with popula-
tion density change (- 0.48), mall proximity (-0.251, and mean
slope gradient (-0.21). The independent variables that
explained about 44 percent of the variability in the forest class
proportion change during the same period were population
density change, mean elevation, and urban center proximity.
At the census tract level, population density and its change
have been the most consistent variables associated negatively
with the proportion of forest and its change. The correlation
between forest class proportion and per capita income was
negative in 1973 and 1987 (-0.66 and -0.23), but was positive
in 1999 (0.29). This reflects the process of suburbanization. In
earlier years, most of the affluent population concentrated not
too far away from large urban centers with a sparse forest cover.
Later, this group of people moved into the urban fringes for bet-
ter environment with dense forest cover. Therefore, affluence
tended to be associated with the density of forest cover.
Ecological Impacts of Land-Use/Land-Cover Changes
Two landscape measures, namely, greenness and fragmenta-
tion, have been included in the analysis at the three spatial lev-
els. As has been observed for the Atlanta metropolitan area as a
whole, mean greenness has decreased while mean fragmenta-
tion index has increased since 1973. They are further examined
at the county and census tract levels below.
Landscape Greenness
At the county level, the mean greenness for 1987 only corre-
lated negatively with population density (-0.81), urban center
proximity (-0.68), node proximity (-0.67), mall proximity
(-0.64), and per capita income (-0.63). At the census tract
level, the mean greenness for 1973,1987, and 1999 correlated
positively with mean slope gradient (0.21,0.42, and 0.50), but
negatively with road proximity (-0.49, -0.62, and -0.57),
node proximity -0.44, -0.62, and -0.57), population density
(-0.26, -0.40, and -0.51), urban center proximity (-0.18,
-0.34, and -0.31), and mean elevation (-0.11, -0.16, and
-0.10). Those variables that explained 51 to 53 percent of the
variance in the mean greenness of 1987 and 1999 were road
proximity, mean slope gradient, mean elevation, population
density, and per capita income. The change in the mean green-
ness from 1973 to 1999 correlated positively with mean slope
gradient (0.13), but negatively with population density change
(-0.16), highway proximity (-0.15), and node proximity
(-0.14). Those variables that explained about 59 percent of the
variance in the mean greenness change were mean slope gradi-
ent, population density change, mall proximity, urban center
proximity, and road proximity. Essentially, population, high-
way proximity, node proximity, and urban center proximity
have been negatively associated with mean greenness and its
change. The decrease in greenness has worsened in recent years
as a result of the increase in population. Much of the decrease
in greenness occurred in gentle terrains close to major high-
ways and nodes. It is interesting to note that the relationship
between greenness and per capita income was negative in 1973
Oct ober 2002 PHOTOGRAMMETRICENGINEERING& REMOTE SENSING
and 1987, but positive in 1999. As a result, it is similar in inter-
pretation to the relationship between the proportion of forest
land cover and per capita income discussed above as the conse-
quence of suburbanization.
Landscape Fragmentation
At the county level, the mean fragmentation index for 1973,
1987, and 1999 correlated negatively with mean slope gradient
for all three years and mean elevation for 1973 and 1987 only.
The change in the mean fragmentation index from 1973 to 1999
correlated positively with mean elevation (0.72) and mean
slope gradient (0.70). In other words, landscape fragmentation
occurred mostly at sites with lower elevations and gentler
slopes, which urban development prefers. At the census tract
level, the mean fragmentation index correlated positively with
population density in 1973 and 1987 (0.52 and 0.16). The
change in mean fragmentation index from 1973 to 1999 corre-
lated positively with mean slope gradient (0.39) and popula-
tion density change (0.32), and negatively with node proximity
(-0.64), highway proximity (-0.58), urban center proximity
(-0.33), and mall proximity (-0.17). Variables that explained
34 percent of the variance of the mean fragementation index
change were node proximity, mean elevation, urban center
proximity, mall proximity, and population density change.
The implication is clearly that, as population increased, the
environment was changed by the construction of highways and
malls, and, as a result, the landscape became more fragmented.
The increased fragmentation of the landscape occurred in sites
with steeper slopes, again an indication of the intensifying
suburbanization.
The above analyses of land-uselland-cover changes and
their ecological impacts in the Atlanta metropolitan area indi-
cated the importance of spatial scale in affecting the results.
Because a zone-based approach has been used to integrate sat-
ellite image data with socio-economic data from the census at a
unified spatial resolution of 60 meters, analysis at the much
finer census tract level makes more sense, and is statistically
more significant. This has been proven by the fact that popula-
tion density, per capita income, and location measures are more
significant variables at the census tract level than at the county
level in explaining the land-uselland-cover changes engen-
dered by the suburbanization process, which expresses itself
spatially in the formation of edge cities. The interactions of the
socio-economic and location factors are clearly very
important.
Modeling the Future Urban Growth and Landscape Changes
Baslcs of Cellular Automaton Moddlng
The driving force analysis suggested that gentle slopes, high-
ways, nodes, and shopping malls generally promoted urban
development in the Atlanta metropolitan area as a result of the
great demand for suburban housing stimulated by economic
development and population growth. It is therefore possible to
simulate the future growth of Atlanta by using land-uselland-
cover, highways, nodes, and terrain data. The cellular automa-
ton (CA) model as developed by Clarke and Gaydos (1998) was
selected for this simulation. The model is dynamic, scale inde-
pendent, and future oriented. The behavior rules used to guide
urban growth in the model consider not only the spatial proper-
ties of neighboring cells but also existing urban spatial extent,
transportation, and terrain slopes. These behavior rules have,
therefore, realistically accounted for the driving forces in the
formation of edge cities in a postmodern metropolis. The model
can also modify itself if extensive growth or stagnation leads to
aberrations from the linear normal growth development. The
model can be verified through rigorous past-to-present calibra-
tion using historical land-uselland-cover data.
The five growth control ~arameters to be initiated in the
model, whicgdetermine the ;umber of Monte Carlo iterations,
are diffusion coefficient, breed coefficient, spread coefficient,
slope resistance, and road gravity. These pa&meters must be
determined with intensive model calibration in which each
coefficient combination needs to be tested individually and
the modeled result compared to historical urban and land-use1
land-cover data by suitable statistical methods. The urban
growth computation is based on a set of transition rules as
defined by the four types of growth: spontaneous, diffusive,
organic, and road-influenced (Project Gigalopolis, 1999). Dur-
ing the urban growth computation, a second level of growth
rules, termed self modification, is invoked when the model's
growth rate is larger or smaller than a critical number. In that
case, the model will modify certain parameters to ensure that
the growth rate is in the normal range.
Version 2.1 of the model permits uSGS Level I land-use/
land-cover transitions to be incorporated. Six themes of data
were employed to run the model. These include (1) the urban
extent, (2) the land-uselland-cover data, (3) highways with
nodes and large shopping malls, (4) slopes computed from the
U. S. Geological Survey 7.5-minute DEM data, (5) hillshaded
images also computed from the USGS 7.5-minute DEM, and (6)
excluded areas for urban development consisting of public
lands, water bodies, and streams. Both the urban extent and
land-uselland-cover were extracted from Landsat images for
five years as previously explained. All the input data layers
were then converted into a single raster format, namely, the
ERDAS WIAGINE format. Under IMAGINE, all images were
further resampled into a spatial resolution of 240 m, a choice
determined by the limitation of computing resources at the
time of conducting this research. The next step was to convert
these images into an 8-bit GIF format required for input to the
model. The model then went through three forms of calibra-
tion: coarse, fine, and final to determine the best values for the
five growth control parameters mentioned above. Only four
control years were used: 1979,1987,1993, and 1999.
Using the best values for the growth parameters, computer
simulation of land-uselland-cover changes began, initially
from 1974 to 1999. To minimize the error level in the simula-
tion, 100 Monte Carlo computations were used. The simulated
results were compared with the actual land-uselland-cover
maps extracted from Landsat images, thus providing visual
verification of the accuracy of the model calibration. Animated
movies were also generated that allowed the general trend of
urban development in Atlanta to be verified. By comparing the
number of pixels of the modeled land-uselland-cover type that
spatially matched the corresponding land-uselland-cover
derived from the Landsat images, the simulation accuracy was
poor for cultivatedlexposed land but best for forest and water.
Urban use was about 40 percent accurate (Table 4). However,
the overall accuracies varied from 61 percent to 74 percent.
Simulation of future changes was continued for a time span
from 2000 to 2050.
Results of Computer Slmulatlon: Atlanta's Gmwth to the Year 2050
To save space, only the results of the land-uselland-cover
change simulation for 2010,2020,2030,2040, and 2050 are dis-
played in Plate 1. The simulation assumed that suburbaniza-
tion will continue unchanged as the factors for the growth of
postmodern metropolis remain unchanged. The simulated
urban area for 2050 will be 844,656 ha within the limit of the
13 metro counties of Atlanta (Table 5). The net increase in urban
land between 1999 and 2050 will be 474,220 ha, or 25 ha per
day, representing an increase of 128 percent for the entire
period. As a result of such a dramatic growth, urban land will
occupy about 81 percent of the total modeled land by 2050. The
number of urban clusters will decrease from 19,815 in 1999 to
2,234 in 2050, while the average size of each cluster will
PHOTOGRAMMETRIC ENGINEERING& REMOTE SENSING Oct ober 2002 1079
TABLE 4. SPATIAL FIT OFTHE MODELED RESULTS WITH THE KNOWN IANDUSE/LANDCOVER I N DIFFERENT CONTROL YEARS*.
Land-UselLand-Cover
Control Years urban use croplandlgrassland cultivatedlexposed land forest water total
1979 modeled 8486 44468 5356 198101 4916 261327
spatial fit 3484 13057 958 159403 2804 179706
% correct 41.01 29.36 17.89 80.47 57.04 68.77
1987 modeled 22255 44300 5229 184632 4911 261327
spatial fit 6264 22927 1518 159714 3911 194334
% correct 28.15 51.75 29.03 86.50 79.64 74.36
1993 modeled 33774 44050 5177 173415 4910 261326
spatial fit 14881 11248 815 129891 2709 159544
% correct 44.06 25.53 15.74 74.90 55.17 61.05
1999 modeled 45598 43578 5060 162184 4907 261327
spatial fit 17680 17769 1468 123315 4217 164449
% correct 38.77 40.78 29.01 76.03 85.94 62.93
*All the values given in this table are pixel numbers, with the exception of percentage values. The pixel size is 240 meters.
increase from about 25 ha in 1999 to 586 ha in 2050. This
means that smaller urban clusters will grow outward and join
together to form much larger clusters. Ecologically, such a mas-
sive growth of urban land will cause substantial change to the
landscape. The average slope steepness for urban land will
increase from 5.87 percent in 1999 to 8.29 percent in 2050, indi-
cating that many forested areas will be converted into urban
use. This is the nightmarish scenario of unchecked urban
sprawl.
There are certain limitations of the CA models that should
be noted. First, the urban use extent is underpredicted. A major
cause is that the model has not considered other factors control-
ling new development, such as urban or regional development
policies, human behaviors, tax, income, or zoning. Another rea-
son is that the model emphasizes linear growth although, in
the real world, non-linear urban growth is also quite common.
Second, for application to a rapidly suburbanizing city such as
Atlanta, the model's transition rules for new urban develop-
ment may need to be adjusted. The urban growth model over-
whelmingly favors the so-called "organic growth," or
expansion from established urban cells to their surroundings.
This growth pattern is generally true for the development of
high-density urban uses such as commercial, industrial, and
large transportation facilities. But for low-density urban uses
dominated by residential, new developments tend to move
away from existing urban facilities in search of a better living
LEGEND
Plrdlctrd uhl r s d land
C u w land
CroplandlOrauland
Fonrt land
m-
The boundaty of 13 counties
is shown
Plate 1. Simulation of the spatial consequences of future urban growth and landscape
change in the Atlanta metropolitan area for 2010, 2020, 2030, 2040, and 2050.
1 1080 Oct ober 2002 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Urban
Cropland/Grassland
Cultivated/Exposed
Forest
Water
Total
environment. In this sense, the other three types of urban
growth, namely, spontaneous, diffusive, and road-influenced
growth, should be more strongly represented, which will be
the objective of the next stage of modeling research.
Because of the limitations of the model and the fluctua-
tions in the world economy, it is unrealistic to believe that
Atlanta will remain unchanged for a period of 50 years. Near-
term simulation to 2010 and 2020, for a time span of 10 to 20
years, should be more realistic. Within the limit of the 13 metro
counties of Atlanta, urban land will increase by about 44 per-
cent from 1999 to 2010; and 26 percent from 2010 to 2020. For-
est areas will decline by 26 percent from 1999 to 2010 and 29
percent from 2010 to 2020. All this will happen if urban sprawl
continues to proceed in the Atlanta metropolitan area for the
next 10 to 20 years.
Conclusions
In this paper, a zone-based cellular approach to integrating sat-
ellite remote sensing data with demographic and socio-eco-
nomic data from the census was adopted to discover the driving
forces of urban development in Atlanta, Georgia, a postmodern
metropolis that has seen rapid population growth in the past 30
years. Simple and multiple correlation analyses conducted
revealed that population and site location characteristics were
important in determining the land-uselland-cover changes for
the three years of the study: 1973,1987, and 1999. Per capita
income has shown a strong relationship with cropland and for-
est. Suburbanization, originally driven by the desire of the
affluent group of population to live in a more spacious envi-
ronment, has caused the loss of cropland and forests in the
more accessible terrain, with the consequence that existing
cropland and forests are found in steeper terrain in more remote
areas. Both the high-density urban use and low-density urban
use exhibited the importance of location and site characteris-
tics in their development since 1973. The relationship between
greenness and affluence, which changed from negative to posi-
tive correlation in more recent times was also revealed. All
these relationships were much more significant at the census
tract level, the appropriate spatial scale of discernable people-
environment and people-people interactions for the city.
The driving force analysis indicates that for the past 30
years an increasingly affluent population has generated the
demand for high quality housing based on such site characteris-
tics as proximity to highways, nodes, and shopping malls,
which become important factors promoting the growth of edge
cites in Atlanta. Dynamic modeling by cellular automaton (a)
has been found to be suitable in taking into account of these fac-
tors. The model has succeeded in simulating the land-use1
land-cover changes from 1999 to 2050, projecting the coalesc-
ing of small urban clusters to form bigger ones, and within the
limit of the 13 counties, there will no longer be any more subur-
ban areas to develop. Such a scenario may not be realistic in
the long range because the economy of the United States and
the world will change. On the other hand, the near-term projec-
tion of land-uselland-cover change to the years 2010 and 2020
may be realistic, which predicts the continuation of highway-
driven suburbanization as population and per capita income
continues to grow. We expect the continued decline of forest
land and cropland, and the landscape will become further frag-
mented at least up to 2010. The main limitation of the CA
model is that not all forms of urban growth have been repre-
sented. Overall, this study has demonstrated the interplay
among biophysical, location site, and socio-economic charac-
teristics of the population in shaping the growth of Atlanta as a
postmodern metropolis.
Acknowledgments
The research reported in this paper has been supported by a
NASA EOS Interdisciplinary Science (IDS) research grants NAS8-
97081 and NASA H-33023D awarded to C.P. Lo as well as an NSF
Doctoral Dissertation Improvement Grant for Xiaojun Yang.
Clarke, K.C., and J. Gaydos, 1998. Loose-coupling a cellular automaton
model and GIS: long-term urban growth prediction for San Fran-
cisco and WashingtonIBaltimore, International Journal of Geo-
graphic Information Science, 12(7):699-714.
Grist, E.P., and R.T. Kauth, 1986. The Tasseled Cap de-mystified, Photo-
grammetric Engineering 6. Remote Sensing, 52(1):81-86.
Dear, M.J., 2000. The Postmodern Urban Condition, Blackwell Publish-
ers, Oxford, United Kingdom, 352 p.
ERDAS, 1997. ERDASField Guide, ERDAS Inc., Atlanta, Georgia, 656 p.
Fujii, T., and T. Hartshorn, 1995. The changing metropolitan structure
of Atlanta, Georgia: locations of functions and regional structure
in a multinucleated urban area, Urban Geography, 16(8):680-707.
Garreau, J., 1991. Edge City: Life on the New Frontier, New York: Dou-
bleday, New York, N.Y., 546 p.
Green, M., and R. Flowerdew, 1996. New evidence on the modifiable
areal unit problem, Spatial Analysis: Modelling in a GIs Environ-
ment, (P. Longley and M. Batty, editors), GeoInformation Interna-
tional, Cambridge, United Kingdom, pp. 41-54.
Jensen, J.R., 1995. Introductory Digital Image Processing: A Remote
Sensing Perspective, Prentice-Hall, Upper Saddle River, New Jer-
sey, 316 p.
Langford, M., D.J. Maguire, and D.J. Unwin, 1991. The areal interpola-
tion problem: Estimating population usin remote sensing in a
GIS framework, Handling Geographical Information: Methodology
and Potential Applications (Ian Masser and Michael Blakemore,
editors), Longman Scientific and Technical, Burnt Mill, Harlow,
United Kingdom, pp. 55-77.
Langford, M., and D.J. Unwin, 1994. Generating and mapping popula-
tion density surfaces within a geographical information system,
The Cartographic Journal, 31:21-26.
Martin, D., 1996. Geographic Information Systems: Socioeconomic
Applications, Second Edition, Routledge, London, United King-
dom, 210 p.
Monmonier, M.S., 1974. Measures of pattern complexity for choropleth
maps, The American Cartogmpher, 1(2):159-169.
Norusis, M.J., 1998. SPSS 8.0 Guide to Data Analysis, Prentice Hall,
Upper Saddle River, New Jersey, 563 p.
Openshaw, S., 1984. The Modifiable Areal Unit Problem, CATMOG
38, Geo Abstracts, Norwich, United Kingdom.
Project Gigdopolis, 1999. The Clarke Urban Growth Model (Version 2.11,
National Center for Geographic Information and Analysis (NCGIA),
I
PHOTOGRAMMETRIC ENGINEERING& REMOTESENSING October 2002 1081
University of California, Santa Barbara, URL: http:llwww. Yang, X., and C.P. Lo, 2000. Relative radiometric normalization
ncgia.ucsb.edulprojects/gig (date last accessed 10 June 2002). performance for change detection from multi-date satellite
Research Atlanta, Inc., 1993. The Dynamics of Change: An Analysis of
images, Photogrammetric Engineering &+ Remote Sensing,
Growth in Metropolitan Atlanta Over the Past lhro Decades, Policy
66(8):967-980.
Research Center, Georgia State University, Atlanta, Georgia, 82 p.
- , 2002. Using a time series of satellite imagery to detect land
Veldkamp, A,, and E.E Lambin, 2001. Predicting land-use change, use and land cover changes i n the Atlanta, Georgia metropolitan
Agriculture, Ecosystems and Environment, 85:l-6. area, International Journal of Remote Sensing, 23(9):1775-1798.
L
MAXIMIZE YOUR COMPANTS EXPOSURE
BY HAVING YOUR IMAGES FEATURED IN
ASPRS PROMOTIONAL MATERIALS!
As technology Is constantly changing, it is important for the
society t o provide the most current representations of this in-
dustry t o our members, readers, and potential clients or em-
ployees.
We need your help with locating applicable images that con-
vey exceptional examples of the work done by our profession
so that we can expand the variety of images in our photo image
bank. We are asking all ASPRS Sustaining Members t o participate by donating five t o ten of their company's images
illustrating good examples of GIs, Photogrammetry or Remote Sensing.
The images would be featured throughout ASPRS promotional materials - primarily conference brochures and
advertisements, membership pieces, and the ASPRS Career Brochure and accompanying materials.
In exchange for your donation, your company will receive photo recognition wherever the image is used, thereby
increasing your company's visibility. We encourage you to take advantage of this opportunity-please participate.
To donate images, please follow the submission instructions below. For more information, please contact ASPRS
Production Manager, Rae Kelley by phone: 301 -493-0290 ext. 107 or email: rkelley@asprs.org.
Submission Instructions
Color
300 DPI, tiff or eps format
40-50 word description including image type, GIs, Photogrammetry or Remote Sensing
Submit on a disk or t o the ASPS FTP site (Contact Rae Kelley for logon instructions.)
Thank you in advance for your support!
1082 Oc t ober 2002 PHOTOGRAMMETRIC ENGINEERING & REMOTESENSING