Land Use Change Modeling with SLEUTH: Improving Calibration with a Genetic Algorithm

Abstract

SLEUTH is a cellular automaton computer simulation model that uses historical land use and other data to project growth and land use change into the future. The model has seen over 100 applications worldwide, and has been among the leading cellular automaton (CA) models applied in simulating land use change at many different spatial scales. The model is highly dependent on the use of historical data to derive the behavioral parameters that best capture the structure and dynamics of the location-specific growth history. While several improvements have been made to the model to increase calibration speed, the current brute force calibration technique has proven popular, in spite of it requiring a multi-phase process and hundreds of CPU hours. This chapter reports on the use of a new alternative calibration method, in which the brute force method is replaced with a genetic algorithm (GA). A version of the model code that executes the GA calibration has been written and made public. The GA calibration process populates a “chromosome” with a set of parameter combinations (genes), of which five are required by the model, each with ranges from 0 to 100. These combinations are then used for model calibration runs, and the most successful (as measured by the Optimal SLEUTH metric) are selected for mutation (recombination of their values), while the least successful are replaced with new randomly selected values. Critical values that must be provided are the population size of the chromosome, the number of iterations or generations over which evolution will continue, the evolution mutation rate, and the number of offspring and replacements in each generation. To select suitable default values for these rates, two SLEUTH applications were used at the extremes of the model’s calibration performance success. These were for San Diego, California where the model fit was very strong, and Andijan, Uzbekistan, where the model was most hard pressed to capture the complex growth process. In both cases, full model calibrations were completed using brute force calibration, followed by calibrations using the GA. It was found necessary to hold the GA parameters constant while repeatedly recalibrating the model using different values for the GA settings. In all cases, the GA model performed as well as the brute force method, but used vastly less computation time. There were also subtle but minor differences in the best SLEUTH forecasts that were explored by mapping the differences among results. The optimal values for GA calibration are given and set as the defaults for SLEUTH-GA, a new version of the SLEUTH model.