Urban multiple land use change (LUC) modelling enables the realistic simulation of LUC processes in complex urban systems; however, such modelling suffers from technical challenges posed by complicated transition rules and high spatial heterogeneity when predicting the LUC of a highly developed area. Tree-based methods are powerful tools for addressing this task, but their predictive capabilities need further examination. This study integrates tree-based methods and cellular automata to simulate multiple LUC processes in the Greater Tokyo Area. We examine the predictive capability of 4 tree-based models–bagged trees, random forests, extremely randomised trees (ERT) and bagged gradient boosting decision trees (bagged GBDT)–on transition probability prediction for 18 land use transitions derived from 8 land use types. We compare the predictive power of a tree-based model with multi-layer perceptron (MLP) and among themselves. The results show that tree-based models generally perform better than MLP, and ERT significantly outperforms the three other tree-based models. The outstanding predictive performance of ERT demonstrates the advantages of introducing bagging ensemble and a high degree of randomisation into transition probability modelling. In addition, through variable importance evaluation, we found the strongest explanatory powers of neighbourhood characteristics for all land use transitions; however, the size of the impacts depends on the neighbourhood land use type and the neighbourhood size. Furthermore, socio-economic and policy factors play important roles in transitions ending with high-rise buildings and transitions related to industrial areas.
|Number of pages||26|
|Journal||International Journal of Geographical Information Science|
|Publication status||Published - Apr 3 2018|
All Science Journal Classification (ASJC) codes
- Information Systems
- Geography, Planning and Development
- Library and Information Sciences