Woods Hole Research Center

Input Data

For calibration, SLEUTH requires inputs of historic urban extent for at least four time periods, a historic transportation network for at least two time periods, slope and an excluded layer.

New techniques to map impervious surfaces from Landsat Thematic Mapper (TM)and Enhanced Thematic Mapper-Plus (ETM+) imagery allowed us to map urban change for 1986, 1990, 1996 and 2000 for the Washington, DC-Baltimore, MD region. Two time steps for roads (1986, 1996) were derived from the primary road network defined in the 1:100,000 scale USGS digital line graphs. A USGS 7.5 minute digital elevation model (DEM) was used to create an input layer for slope. For calibration, the excluded layer consisted of water, which was 100% excluded from development, and federal, state, and local parks, which were 80% excluded due to limited development that had occurred within some parks. All input files were rasterized ar a 45 m cell resolution to the spatial extent of the study area and checked for overlay accuracy.

Subset of calibration data set, showing the increase in developed lands between 1986 and 2000 in northern Virginia.

Calibrating SLEUTH

The goal of calibration is to derive a set of values for the growth parameters that can effectively model development patterns during the historic time period, in this case 1986-2000. In the SLEUTH modeling environment, this is achieved through a brute force method, where the user indicates a range of values and the model iterates using every possible combination of parameters. For each parameter set combination, the model simulates growth multiple times in a Monte Carlo process and compares simulated growth to actual growth, computing several least squares regression measures, including the number of urban pixels, the number and size of urban clusters, urban edge pixels, as well as other fit statistics.

The modeling software calculates these statistics internally and writes the results to a log file that is used to evaluate the performance of the different parameter sets. The user must decide what aspects of urban form (i.e. urban shape, number of urban clusters, the overall amount of growth, etc.) should be most accurately represented. WHRC researchers found that parameter sets that maximized the compare metric, which compares the amount of simulated urban growth to the amount of actual urban growth, were able to accurately capture the rate of growth as well as matching urban shape. Each parameter can take on a value between 1 and 100, and the parameter values that we derived were: dispersion = 52, breed = 45, spread = 26, slope = 4, and road growth = 19.

The brute force Monte Carlo calibration method is computationally intensive. For this data set, over a week of processing time on a USGS parallel computer (16 node Beowulf PC cluster) was required for calibration.

Accuracy Assessment

To perform an accuracy assessment, WHRC scientists initialized the model with the 1986 urban extent and growth was predicted out to the year 2000 using the parameter set derived during calibration. The resulting map showed the probability of any given cell becoming urbanized in 2000. In order to compare simulated patterns of growth with mapped patterns, the probability image was reclassified into a binary representation of urban extent using a probability threshold of 50%. Simulated development patterns were compared to mapped development at three scales: pixel, small watershed, and county.

The overall accuracy at the pixel scale was quite high (93.1%), but when only areas where change was predicted or observed (roughly 22% of the study area) were considered, the overall accuracy dropped to 19%. This points to a limitation in the model for predicting the exact location of urbanized pixels, but accuracy at the pixel scale is not crucial for a regional assessment. The model performed quite well when estimates of development were aggregated to the 11-digit (HUC 11) watershed scale (r2 = 0.72, P104 < 0.01) and at the larger spatial units of counties (r2 = 0.86, P44 < 0.01).