A Comparative Study of Machine Learning Techniques to Simulate Land Use Changes

Abstract

Design and development of a practical land use change (LUC) model require both a high prediction accuracy, to predict the future changes, and a well-fitted model reflecting and monitoring real world. In this regard, many models follow the three phases: training, testing and validating, in the modelling process to maximise both the accuracy and fitness. Therefore, the choice of model for different applications is still a valid and important question. This paper applies and compares three widely used data mining models: classification and regression tree (CART), multivariate adaptive regression spline (MARS) and random forest (RF), to simulate urban LUCs of Shirgah in Iran. The results of these three phases for the three models: CART, MARS and RF, for the study area of Shirgah, in the north of Iran, verify that having the highest accuracy in the testing run does not necessarily guarantee the highest accuracy in the validating run. And so, with respect to the purpose of each project, such as modelling the current situation or predicting the future, the best model with the highest accuracy at the relevant phase or a combination of some/all should be selected. For example, in this study, MARS can provide the best accuracy in the validating run while the lowest level of accuracy in the testing run. RF provides the highest accuracy in the testing run and the lowest level of accuracy in the validating run.