Predicting Urban Functional Zones with Twitter Data Using the Space-Time Scan Statistics Method and the Random Forest Classifier

Studies relating to spatiotemporal data clustering in geosocial media data have laid the groundwork for the analysis of clusters over a large study area. This thesis aims to identify whether urban functional zones can be predicted based on a pattern of data clusters that occurs within each functional zone. A framework is presented for the segmentation, annotation, classification, and validation of the functional zones of land segments in an urban study area. The Space-Time Scan Statistics approach was adopted and used to cluster a dataset of tweets into ‘events’. Characteristic attributes of the detected cluster areas were converted into land segment attributes and used as the input in the Random Forest classifier. Labels for each segment were determined based on high-level land use classes. The classifier was trained, validated, and tested on subsets of the study area. The resulting precision, recall, and F1 score were 88.96%, 89.15%, and 88.23%, respectively.