Integrating Ensemble-Urban Cellular Automata Model with an Uncertainty Map to Improve the Performance of a Single Model.
Xuecao Li,Xiaoping Liu,Peng Gong
DOI: https://doi.org/10.1080/13658816.2014.997237
2015-01-01
International Journal of Geographical Information Science
Abstract:Transition rules are the core of urban cellular automata CA models. Although the logistic cellular automata Logistic-CA is commonly used for rules extraction, it cannot always achieve satisfactory performance because of the spatial heterogeneity and the inherent complexity of urban expansion. This article presents an ensemble-urban cellular automata Ensemble-CA model to achieve better transition rules. First, an uncertainty map that assesses the performance of transition rules spatially was achieved. Then, two auxiliary models i.e. classification and regression tree, CART; and artificial neural network, ANN, both of which have been stabilized with a Bagging algorithm, were prepared for integration using a proposed self-adaptive -nearest neighbors -NN combination algorithm. Thereafter, those unconfident sites were replaced with the ensemble output. This model was applied to Guangzhou, China, for an urban growth simulation from 2003 to 2008. Static validation confirmed that this ensemble framework i.e. without substitution of uncertain sites can achieve better performance 0.87 in terms of receiver operating characteristic ROC statistics area under the curve, AUC, and outperformed the best single model ANN, 0.82 and other common strategies e.g. weighted average, 0.83. After the substitution of unconfident sites, the AUC of Logistic-CA was elevated from 0.78 to 0.81. Subsequently, two urban growth mechanisms i.e. pixel-and patch-based were implemented separately based on the integrated transition rules. Experimental results revealed that the accuracy obtained from simulation of the Ensemble-CA increased considerably. The obtained kappa outperformed the single model, with improvements of 1.74% and 2.76% for pixel-and patch-based approaches, respectively. Correspondingly, landscape similarity index LSI improvements of these two mechanisms were 4.24% and 1.82%.