Neural-network-based cellular automata for realistic and idealized urban simulation

There is rapid development of CA models for simulation of land use patterns and urban systems recently. Traditional methods using multicriteria evaluation (MCE) have limitations because they only use a linear weighted combination of multiple factors for predictions. It cannot explain much of the non-linear variations presented in complex urban systems. It is most attractive that neural networks have the capabilities of nonlinear mapping which is critical for actual urban systems. The study indicates that improvement has been made by using the proposed model to simulate non-linear urban systems. The advantages of using neural networks are apparent. The method can significantly reduce much of the tedious work, such as the requirements for explicit knowledge of identify relevant criteria, assign scores, and determine criteria preference. Furthermore, variables used in spatial decision are always dependent on each other. General MCE methods are not suitable to handle relevant variables. Neural networks can learn and generalize correctly and handle redundant, inaccurate or noise data which are frequently found in land use information. Users don't need to worry about which variable should be selected or not. Knowledge and experiences can be easily learnt and stored for further simulation. General CA models also have problems in obtaining consistent parameters when there are many variables in the prediction. It is very time consuming in finding the proper values of parameters for CA models through general calibration procedures. This paper has demonstrated that neural network can be integrated in CA simulation for solving the problems in finding the values of parameters. Users don't need to pay great efforts in seeking suitable parameters or weights which are difficult to be determined by general CA methods. In the proposed method, the parameters or weights required for CA simulation are automatically determined by the training procedures instead of by users. It is convenient to embed the neural network in the CA simulation model based on the platform of GIS. The model is plausible in forecasting urban growth and formulating idealized development patterns. Different scenarios of development patterns can be easily simulated based on proper training using neural networks. Remote sensing data can be used to prepare training data sets for more realistic simulation. Based on planning objectives and development evaluation, original training data sets can be rationally modified to obtain different sets of adjusted weights through the training procedure of neural networks. These adjusted weights can be applied to the CA model in generating idealized patterns.