USING CALL DETAIL RECORDS DATA TO PREDICT POST-EARTHQUAKE EVACUATION WITH A MACHINE LEARNING APPROACH

Abstract

The devastating February 2023 Turkey-Syria earthquakes resulted in over 55,000 death. While current technology can not predict earthquakes, efficient evacuation management can significantly reduce secondary casualties and optimize resource allocation. This study explores the application of Call Detail Records (CDR) in times of crisis, with a particular focus on the consequences of earthquakes. This study focuses on two issues, the prediction of population movements after earthquakes, and the features influence post-earthquake evacuation behavior. We use machine learning models with gravity transformed features to predict population movements immediately after earthquake. The experiments show our model have good ability to predict evacuation flow between different district. Our main findings are that population distribution and earthquake intensity are the primary factors of evacuation patterns. The comparative analysis between Turkish population and Syrian population shows the same feature importance rankings but distinct pattern distributions. These results provide valuable insights for emergency management authorities in resource allocation and evacuation planning, such as the effect of social connectedness.