Building Segmentation and Classification for Ger Districts of Ulaanbaatar, Mongolia

Abstract: Gers are portable round tents found all over the city of Ulaanbaatar, Mongolia, and are a large source of air and soil pollution in the city. Efforts to locate and map gers are expensive and time consuming. A machine learning approach to mapping and identifying gers was developed and a random forest classifier was trained on morphometric building features to correctly classify 98% of ger buildings on an out-of-sample test set. Our approach is semi-supervised as not all buildings in our data are correctly labeled as ger. We use our model to predict these unlabeled ger buildings and find our model performs at 99.8% accuracy after a manual review of 500 predicted ger buildings. This model adds 27,808 gers to the hand labeled 36,004 for a total of 63,812 gers. We conclude with a discussion on policy implications and how this method can enable further studies while reducing data acquisition costs.