Predict Urban Growth Patterns Using Machine Learning with ArcGIS Pro and R

The presentation showcases findings from a collaboration between GIS professionals and data scientists to apply machine learning algorithms to predict urban development. The study area (125 by 77 km) is known as the Research Triangle in North Carolina. As population is growing, new areas need to be converted from their current land use types into urban land use. It is crucial for city planners as well as for developers to know which areas would be most suitable for urbanization and how probable urban development in a specific area is. The answers to these questions are critical for government agencies such as planning departments to better understand urban growth in order to make better policies. It is also beneficial for private investors who are searching for locations to make profitable investments. Investors looking for opportunities to invest in real estate and commercial infrastructure such as shops or restaurants can use the model to locate suitable areas. The goal of the project is to predict locations with a high probability of urban development. The machine learning model used is Random Forest, which is a popular supervised learning algorithm based on a multiple of decision trees. To prepare the input predictor variables for the Random Forest model, ArcGIS geoprocessing tools were used to generate various raster layers. These raster layers are factors that affect urban development, including proximity to roads, urban centers, environmental protected areas, flood zones, as well as terrain characteristics and projected population growth. The outcome of the Random Forest model is a categorical variable predicting whether an area is urbanized or not. The related training and testing data were processed using ArcGIS raster analysis tools based on 2001 to 2016 data from the National Land Cover Database. After the model was trained with the Random Forest algorithm in R, the prediction results were visualized using ArcGIS's advanced cartography capability. The predictions showed promising accuracy level. Based on the test dataset that was not used to optimize the Random Forest model 95% of records were predicted overall with balanced sensitivity and specificity. The study applies Esri's GIS software ArcGIS Pro, various R packages in Rstudio, and the R-ArcGIS Bridge an open source R package from Esri that allows the exchange of data between ArcGIS Pro and R. The presentation will also share the project's collaboration experience as an example on how to integrate knowledge and skills from GIS and Data Science fields as well as the experience with the data exchange. Since large datasets were used, a smooth workflow was needed to exchange data between R and ArcGIS Pro. The R-ArcGIS Bridge package that allows R to dynamically access ArcGIS data, and save R results back to ArcGIS datasets provided a seamless workflow. -------------------------------------------------------------------------------------------------------------------------- Follow us on Social Media! Twitter: https://twitter.com/Esri Facebook: https://facebook.com/EsriGIS LinkedIn: https://www.linkedin.com/company/esri Instagram: https://www.instagram.com/esrigram The Science of Where: http://www.esri.com