Rules versus layers: which side wins the battle of model calibration?

Continuous surface of urbanization suitability, as an input to many urban growth models (UGM), has a significant role on a proper calibration process. The present study evaluates and compares the simulation success of the Cellular Automata-Markov Chain (CA-MC) model through multiple methods. For this, a series of mapping algorithms are applied ranging from empirical methods such as multi-criteria evaluation (MCE) to statistical algorithms without spatially explicit suitability mapping rules such as logistic regression (LR) and multi-layer perceptron (MLP) neural network and finally statistical and spatially explicit rule-based methods such as SLEUTH-Genetic Algorithm (SLEUTH-GA) model. The CA-MC model was calibrated in three study locations including Azadshahr, Gonbad, and Gorgan cities in northeastern Iran. Applying Kappa-based indices (Kappa, K ₗₒcₐₜᵢₒₙ, K Sᵢₘᵤₗₐₜᵢₒₙ, and K Tᵣₐₙₛₗₒc) and computing relative error (RE) values of landscape metrics, performance of the model was quantified and compared across the three study sites. The MCE and SLEUTH-GA methods, as the most data-demanding and the most computationally complex methods, respectively, yielded approximately similar results (especially in case of Kappa-based indices) and these methods were less successful compared to LR and MLP models. LR and MLP models were less data-demanding, while they produced approximately equal results. This study concludes that, when historical growth patterns feed an urbanization suitability mapping process, neither rules (SLEUTH-GA) nor layers (MCE) are effectively efficient when applied in a separated manner. Instead, methods with statistical rules and least-correlated input layers (LR and MLP) provide better simulation outputs. In contrast, methods such as MCE are more applicable when a non-path-dependent mapping procedure is desired since this method does not require training data (dependent variable) and the provided flexibilities in urbanization suitability mapping under various scenarios can improve the functionality of land-use change prediction algorithms into innovative land allocation tools.