Comparison of metaheuristic cellular automata models: A case study of dynamic land use simulation in the Yangtze River Delta

Cellular automata (CA) is a bottom-up modeling framework that has increasingly been applied to simulate land use change by capturing its dynamics. Metaheuristics such as particle swarm optimization (PSO), generalized simulated annealing (GSA) and genetic algorithm (GA) have widely been incorporated into CA modeling to generate more realistic simulation patterns. We present a comparative study of four CA models incorporating logistic regression (LR) and the three metaheuristics respectively to simulate land use change in the Yangtze River Delta from 2005 to 2015. The metaheuristic processes are guided by an objective function that represents the root-mean-square error (RMSE) of the transition rules, which can then automatically search for suboptimal CA coefficients. The three metaheuristics are substantially different in terms of the algorithm mechanism, optimization iteration, and computational time. The land conversion potentials from the metaheuristics are similar in global patterns but marginally different in local regions, which substantially differ from that calculated using LR. All three metaheuristic CA models simulated slightly less than the reference change while the LR-CA model simulated substantially more than the reference change, however all models allocated the change to similar places. Our study shows that the three metaheuristics can achieve similar outcomes in the optimization of CA transition rules and land use simulation, albeit with different sensitivities to their intrinsic control parameters. We suggest that any of the three metaheuristics could be used to construct land use CA models, if the algorithm complexity and computational time are not highly concerned.