Towards automatic calibration of neighbourhood influence in cellular automata land-use models

Cellular Automata (CA) land-use models are widely used for understanding processes of land-use change. However, calibration of these models is a knowledge-intensive and time-consuming process. Although calibration of common driving factors such as accessibility (A), or suitability (S) is a relatively straightforward task, calibrating the neighbourhood dynamics (N), which controls the key model behaviour, is often very challenging. Here, building on the SIMLANDER modelling framework, we develop an automatic rule detection (ARD) procedure to automate the calibration of N. To demonstrate the performance of the tool, we simulated 15 years of urban growth in Ahvaz, Iran (2000-2015) using a wide range of different rule-sets. We evaluated calibration goodness-of-fit for each rule-set against a reference map by means of cross-comparison of multiple metrics using a ranking procedure. The ARD procedure can be implemented in two ways: 1) by random sampling of the parameter space, a user-defined number of times, or 2) through a stepwise "grid search" approach for a user-defined number of rule combinations. Both approaches were found to produce successful rule combinations according to the goodness-of-fit metrics applied. Grid search performed slightly better, but at the cost of a fivefold increase in computation time. The ARD approach facilitates model calibration by allowing rapid identification of the optimum ruleset from a wide range of possible parameter settings, while the ranking procedure facilitates comparison of simulations using multiple metrics. The approach we present also helps to improve simulation accuracy with respect to manual calibration methods, and increases understanding of neighbourhood dynamics for the urban area studied. To encourage repeatability and transparency, our approach uses only open data and Free-and-Open Source Software (RStudio environment) and we provide our ARD scripts as supplementary material