Authors: Firoozeh Karimi*, University of north carolina at greensboro, Selima Sultana, University of north carolina at greensboro
Topics: Land Use and Land Cover Change, Spatial Analysis & Modeling, Geographic Information Science and Systems
Keywords: Urban expansion, GIS, machine learning, decision trees, classification and regression tree
Session Type: Paper
Start / End Time: 5:00 PM / 6:40 PM
Room: Washington 6, Marriott, Exhibition Level
Presentation File: No File Uploaded
Decision tree (DT) algorithms have been applied for classification and change detection in various geospatial studies and more recently for urban expansion and land use/land cover (LULC) change modeling. However, these studies do not elaborate on specifications of these methods with respect to the selection of an appropriate combination of data sampling, predictor variables, model configuration, and model evaluation. In this study, the capabilities of classification and regression tree (CART) DT algorithm is explored to develop a better urban expansion modeling. For this purpose, several balanced and unbalanced sampling methods, various predictor variables, and different configurations of stopping rules are applied to investigate their effects on the performance of CART. Furthermore, due to a low urbanization rate in the study area, the goodness-of-fit metrics are used based on the changed and unchanged land cells in the whole study area to make the accuracy of the model evaluation more realistic. An implementation of the model in Guilford County, NC, between 2001 and 2011 demonstrated that the best performance of the model with the training accuracy of 94% and the testing accuracy of 90%, is achieved using a sampling dataset containing all changed cells and twice of that from unchanged cells, 15 predictor variables, the minimum number of records in a leaf node equal to 25, the minimum number of records in a parent node equal to 50, and the value of 2000 for maximum number of splits.