International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
model as a tool to imagine, test and choose between possible
future urban growth scenarios in relation to. different
environmental and development conditions.
2. RESEARCH METHODOLOGY
2.1 Data Assemblages
Five types of input data are needed to run the SLEUTH urban :
growth model. They are: urban extent, road, excluded area for
development, slope, and shaded relief. Most of these input data
were assembled from the databases constructed by the author
through different research efforts over the past years. Each layer
was resampled into three levels of spatial resolutions, namely 60
m, 120 m, and 240 m.
2.1.1 Urban Extent: Urban extent is actually urban built-up land,
thus including all types of urban uses. Five layers of urban extent
data were extracted from a time series of land use/cover maps
that were produced with a hybrid approach combining
unsupervised classification and knowledge-based spatial
reclassification. Detailed description of these procedures can be
found from Yang (2002). These layers represent five different
dates, namely, 1973, 1979, 1987, 1993, and 1999.
2.1.2 Road: It contains not only major road networks but also
node points and large shopping malls. For convenience, this layer
is still named as ‘road’. The major highways were extracted
from the AND global highway database (http://www.and.com),
and then updated with satellite images to form 1973, 1987, and
1999 highway layers for three years. Major node points are
either (major) highway exits, junctions, or towns where major
highway(s) runs across. They may be of strategic significance
for commercial or industrial development. Three layers of large
mall polygons were extracted from the 1973, 1987, and 1999
Landsat images. A weighting system was established for
highways, nodes, and malls, respectively.
The layers of highways, nodes, and shopping malls in the same
year were combined to form a single ‘road’ layer. In this way,
three 'road' layers were produced for 1973, 1987, and 1999,
respectively. The ‘road' layer for the year of 2025 was produced
by overlaying the 1999 roads with the improved roadways and
new roadways according to the 2025 Regional Transportation
Plan (Atlanta Regional Commission, 2000).
2.1.3 Excluded Area for Development: Two layers of excluded
areas were assembled. The first layer is a binary image,
consisting of the water extracted from 1973 Landsat MSS image
and the public lands. The latter includes national parks/refuge
and wilderness areas, archaeological sites/areas, historic sites,
off-road vehicle sites/areas, wild and scenic areas, state parks,
USDA land, wildlife management areas, and county Parks. These
areas were not allowed for urban development. This layer was
mainly used for the model calibration.
For the future growth prediction, another layer was built, with
probabilities of exclusion included. All excluded areas in the
first layer were still preserved and assigned a value of 100.
Additionally, this layer contains three levels of buffer zones
around major streams in the study area.
2.1.4 Slope and Shaded Relief Image: In order to produce
terrain slope and shaded relief images, a seamless DEM image
was constructed by mosaicking 159 USGS 7.5' DEMs covering
the entire modeled area. Then, a terrain slope image was
computed and represented in percentage. Furthermore, a layer of
1228
the hillshaded image was computed from the DEM. This image
shows the topographic relief in the study area. It was used as a
background image for visualization purpose only.
2.2 Model Calibration
The purpose of model calibration was to determine the best values
for five control coefficients, namely, diffusion (overall
dispersiveness of growth), breed (likelihood of new settlements
being generated), spread (growth outward from existing spreading
centers), slope resistance (likelihood of settlements extending up
steeper terrain), and road gravity (attraction of urbanization near
road networks).
The calibration was built upon on a statistical approach. Thirteen
statistical measures were computed to quantify the historical fit
between the modeled results and historical urban extent data
extracted from remotely sensed imagery. The list of these
statistical measures and their detailed description are given
elsewhere (Yang and Lo, 2003). They were used to narrow down
the range and determine the best value for each control coefficient.
The possible range is between 0 and 100 and the possible
combinations for the five control coefficients are approximately
5!? or 7.89 x 109! Ideally, each combination should be assessed.
Given the computational resources available (a Sun Ultra Model
1, with 143 Mhz CPU and 64 Mb RAM), however, this would take
years to complete according to an earlier test. For the time and
computational resources constraints, the calibration was broken
down into three phases (Table 1). The coarse calibration was to
block out the widest range for each control coefficient. The fine
calibration was to narrow down the ranges to approximately 10 or
less The final calibration was to determine the best combination,
which had the following starting values: diffusion(55), breed (8),
spread(25), slope resistance (53), and road gravity (100).
Table 1 Calibration runs: input data, calibration files, number of
Monte Carlo iteration, computation time, and outputs.
Future Simulations
Items 2 >
Scenario 1 | Scenario 3
2000-2050
Time Span
Resolution (m) 240
Input Data ‘bantexte
urban extent 1999
(vear)
‘roads’ 1999 1999, 2025
excluded areas stream * stream
buffered buffered
zones not zones
considered | considered
slope same (only one layer can
be chosen)
hillshaded relief same (only one layer can
be chosen)
critical high
1.500
Self critical low 0.050
Modification
Constraints* boom 1.010
bust 0.090
critical slope 21 10
diffusion 71/88 100/100
Control breed 10/12 100/100
Cocfficients* * E mS
spread 32/40 15/15
slope resistance 73/100 10/40
road gravity 100/100 |200/2003***
Number of Monte Carlo
: 100
Computations
Random Samples 2840
* These are about 1 percent of the total pixel numbers.
** This number is for the Monte Carlo iterations.
