Machine Learning-enabled Adaptive Air Traffic Recommendation System for Disaster Evacuation

Extreme weather conditions, such as floods, hurricanes and wildfires, cause large-scale human population movements and evacuations in the world. Taking flights to evacuate the area before a disaster approaches is mostly preferred because of the much higher efficiency than driving or taking trains. However, how to make effective flight plans serving evacuation before these natural disasters occur is a significant challenge because it relies heavily on many dynamic factors including human resources, environments, and traffic loads. In this paper, we proposed an adaptive flight recommendation schema that aims to coordinate available resources such as aircraft and airports to transport people from evacuated areas to safe places. We formulate the evacuation flight dispatch as a weighted graph matching problem where weights indicating the dynamics of the air route network are inferred by machine learning approaches. In particular, the problem is resolved in a two-phase manner: (a) in the offline phase, a predictive model based on random forest and value functions indicating the expected gain of an evacuation flight in a certain state are learned using historical flight data; (b) in the online dispatching phase, several flight candidates are generated with estimated "values" based on predicted air mobility factors from the trained model in the first phase. Kuhn–Munkres (KM) algorithm is then employed to get the best evacuation flight plans (aircraft-route pairs). Our schema is evaluated using data collected during Hurricane Irma, and the experimental results demonstrate its effectiveness in the real-world application that is likely to play a vital role in future disaster evacuation worldwide.