The effect of sampling methods on urban cellular automata models

Over recent decades, data-driven models have been extensively employed in the field of urban CA modeling studies. However, upon reviewing literature, we found that despite having the common goal of extracting driving mechanisms of urban growth from samples, different sampling methods were employed. The variability in sampling methods can pose potential risks to knowledge sharing within this field. Specifically, three distinct sampling methods have been utilized. Following the CA modeling process, we evaluate the impact of alterations in sampling methods on each stage of the process. The driving mechanisms of urban growth extracted from samples obtained through different sampling methods exhibit significant differences, leading to spatially distinct development suitability maps. Notably, the samples obtained from the second sampling method (M-2) contain pronounced non-linear characteristics, resulting in substantial differences between the training results of linear and non-linear models. We found that the “land use change rules” extracted from two-period maps are more beneficial for achieving high-precision simulations compared to the “land class distribution rules” derived from a single map. M-2 may represent a more robust approach when employing powerful non-linear models because it incorporates invariant urban land as negative samples, providing a more comprehensive set of non-linear land use change characteristics.