Enhancing the calibration of an urban growth model using a memetic algorithm

At present, many approaches and models have been developed to perform spatially explicit simulations that mimic observed land use and land cover changes (LULC) for a given area. Calibration of such models is often performed using comparatively standard 'off-the-shelf' machine-learning algorithms that are not necessarily suited to perform effectively within the model's implementation. This method becomes problematic when the computational costs of applying an evaluation function to determine the goodness-of-fit are high; calibration using 'standard' algorithms often requires many iterations to achieve satisfactory outcomes. Furthermore, in some cases, future LULC projections manifest significant changes in trends, particularly when increasing the number of LULC classes in the simulation and the number of associated transition rules. This study presents an adapted machine-learning algorithm to optimize a parameter set applied in a Dinamica-EGO-based LULC change model. A sequentially applied memetic algorithm is applied to optimize a vast parameter set by extending a genetic algorithm with a local search function. To achieve consistent long-term projections, a 2-stage approach is applied in which the expansion of the urban extent and diversification of urban LULC classes are calculated sequentially. The outcomes repeatedly show a much faster convergence toward a high goodness-of-fit; significantly fewer iterations and a smaller population size can be used to attain a similar performance level than when using a standard GA-enhanced calibration. Furthermore, the observed spatial trends are maintained for long-term projections using 5-year intervals. In the current application, the model is applied to the rapidly growing metropolitan area of Beijing, China.