Simulation of development alternatives using neural networks, cellular automata, and GIS for urban planning

Abstract

This study integrates neural networks and cellular automata (CA) to simulate development alternatives for planning purposes. Most of the existing CA just focus on simulating realistic urban dynamics. This paper demonstrates that development alternatives can be simulated by incorporating planning objectives in CA. It is important to define appropriate parameter values for simulating development alternatives according to the planning objectives of planners and decision makers. Training neural networks can automatically yield the parameter values for urban simulation. GIS and remote sensing provide the training data for calibrating the model. However, the simulation can inherit past land-use problems if the original training data are used to calibrate the model. The original data should be assessed and modified so that the model can remember the past "failure" in land development. Planning objectives can thus be embedded in the model by properly modifying the training data sets. The training is robust because it is based on the well-defined back-propagation algorithm. Experiments were carried out by using the city of Dongguan, China as an example to test the model.