Using the ART-MMAP neural network to model and predict urban growth: a spatiotemporal data mining approach

Predicting patterns of urban growth will be a major challenge for policy makers and environmental scientists in the 21st century. How cities growötheir shape and sizeöwill have enormous implications for environmental sustainability and infrastructure needs. This paper presents a spatiotemporal ART-MMAP neural method to simulate and predict urban growth. Factors that affect urban growthöthat is, transportation routes, land use, and topographyöwere directly used as inputs to the neural network model for model calibration. The calibrated network was then applied to a study siteöSt Louis, Missouriöto predict future urban growth and to examine future land development scenarios. This paper also introduces an effective and straightforward method for model validation and accuracy assessment, the prediction error matrix, which has been used in the pattern recognition field for several decades. In order to assess the performance of the neural network model, an in-depth accuracy assessment was conducted in which the model results were compared against a null model, an alternative na|« ve model, and two random models. The neural network model consistently outperformed the na|« ve model and two random models, and produced similar or better results than the null model. Furthermore, we evaluated the models' performance at different spatial resolutions. The prediction accuracy increases when spatial resolution becomes coarser. One particularly interesting result is that when the results are aggregated to 1 km spatial resolution, there is 100% accuracy of urban growth predicted by the neural network model versus actual urban growth. doi:10.1068/b3312 ô Alternative address: Center for Environmental Science and Policy, Freeman Spogli Institute for International Studies, Stanford University, Stanford, CA 94305, USA. develop land-use and development policy, (2) understand the past and predict the future, and (3) create scenarios for policy makers and planners (Fragkias and Seto, 2007). In this paper we present an ART-MMAP neural network method with which to predict urban growth through the mining of historical urban growth data, and we introduce an effective method for model validation and accuracy assessment. 2 Urban growth models Computer-based urban system simulation models have been employed to forecast and evaluate land-use change for well over a decade (Batty and Xie, 1994; Engelen et al, 1995; Landis, 1994). These spatially dynamic modeling approaches enable urban planners to visualize and assess the impacts of their decisions and policies prior to implementation. Computer simulations and forecasts can help to improve our fundamental understanding and communication of the dynamics of land-use transformation and the complex interactions between urban change and sustainable systems (Deal, 2001). There exist two basic types of land-use change models: regression-based models and transition-based models. Regression-based models aim to establish functional relationships between land-use change and a set of predictor variables that are used to explain the locations of future land-use change on the landscape. By including historical land-use data, these models establish functional relationships that can be used to predict the probability of future land-use change. These models may adopt different methods to build prediction functions, including linear or logistic regression (Arai and Akiyama, 2003; Fragkias and Seto, 2007; Seto and Kaufmann, 2003; Theobald and Hobbs, 1998), and artificial neural networks (Pijanowski et al, 2002). One advantage of regression models is that they can explicitly express the effects of each predictor (spatial variables) on future land-use change (Pijanowski et al, 2002). Spatial transition models are characterized by transition rules, neighborhoods, and decision makersöoften called agentsöwithin a cell. They differ from regression models in that a group of simple rules that express land-use change patterns and neighborhood effects (eg spatial adjacency) can be incorporated to drive the prediction of land-use change. Over the last decade cellular automata (CA) and agent-based models have become increasingly popular tools for modeling urban growth (Batty and Xie, 1994; Clarke and Gaydos, 1998; Clarke et al, 1997; White and Engelen, 1997). Although a large number of models have been proposed and developed over the last twenty years, there remain a number of limitations. First, many models have significant data input requirements, limiting their utility in developing countries, where data are usually sparse (Fragkias and Seto, 2007). Second, CA-based modeling techniques are still far from being mature. Despite their flexibility, there are many limitations to CA models (Torrens and O'Sullivan, 2001). The hypothetical urban forms emerging from CA models with surprisingly simple local transition rules are certainly plausible. However, in reality, the evolution of urban systems is significantly more complex. One criticism of current CA-based urban models is that they are too simple to capture the richness of urban systems. Consequently, very few CA models are operational and used as a productive tool to support regional planning practice. To build useful models researchers have extended the concept of CA and integrated a diversity of models, such as traditional regional social-economic models (White and Engelen, 1997; Wu and Martin, 2002), neural network models (Li and Yeh, 2002) and decision trees (Li and Yeh, 2004). In this paper we present a new type of regression model: an ART-MMAP neural network model. The model uses historical urban growth data to `learn' urban growth patterns and then predicts urban growth. The paper also introduces a new method for model validation and presents in-depth accuracy analysis. 2 W Liu, K C Seto