Urban expansion modeling using machine learning algorithms

UNCG Author/Contributor (non-UNCG co-authors, if there are any, appear on document)
Firoozeh Karimi (Creator)
The University of North Carolina at Greensboro (UNCG )
Web Site: http://library.uncg.edu/
Selima Sultana

Abstract: Modeling and simulating urban expansion is required for assessing and predicting the consequences of the current urban growth patterns. Given the dynamic and convoluted nature of the urban expansion process and the necessity of handling continuous and categorical variables, non-normal distributed data, and non-linear relationships, urban expansion modeling is challenging. It is also critically important to find an appropriate method for modeling and simulating urban expansion in order to meticulously identify spatiotemporal variables and predicting the direction of land use/land cover (LULC) changes. To handle these issues effectively and enhance the quality of urban expansion prediction, the capabilities of machine learning methods are explored in this dissertation. Machine learning methods are relatively unknown in urban expansion modeling and have not been evaluated thoroughly in the current literature. The machine learning methods allow the exploration of a variety of data sampling strategies, predictor variables, and model configurations to enhance the accuracy and predictability of urban expansion modeling. The models are calibrated using spatiotemporal data of 2001-2016 and are applied to simulate future urban developments for two urbanized counties—Guilford and Mecklenburg in NC, USA. The accuracy and reliability of the models are evaluated by apposite evaluation metrics. Distance to highways is recognized as the most important predictor variable in both study areas, however, the importance of the predictor variables varies in different geographic contexts and with different methods. A comparative study on machine learning methods demonstrated that the random forest (RF) model is a fast, high-performance, and accurate model with low uncertainty; therefore, it can be effectively utilized to evaluate a wide range of urban development scenarios and support decision-making to accomplish the goal of implementing environmentally sustainable development. Sustainable urban growth management in addition to sophisticated and elaborative models requires different urban growth scenarios. An integration of random forest and cellular automata (RF-CA) is proposed to simulate urban development under three urban growth scenarios, including current trends, controlled urban development, and environmentally sustainable urban development. While current trends allow the urban fringe to be uncontrollably developed, the controlled and environmentally sustainable urban development scenarios constrain future developments and reduce the environmental implications. The results show that the current urban development in the study area for 2021 and 2026 will appear near current or newly built urban clusters or adjacent to the major roads, however, the controlled and environmentally sustainable urban development scenarios are much higher compact and minimize environmental costs.

Additional Information

Language: English
Date: 2021
Decision tree, Environmentally sustainable development, Machine learning, Random forest, Support vector machine, Urban expansion
Sustainable urban development $z North Carolina $z Guilford County (N.C.)
Sustainable urban development $z North Carolina $z Mecklenburg County (N.C.)
Cities and towns $x Growth
Machine learning
Decision trees
Support vector machines

Email this document to