Modeling Dynamics of Community Resilience to Extreme Events with Explainable Deep Learning

Community resilience provides a paradigm guiding communities' preparedness and mitigation efforts to counter the increasing risks of extreme events (EEs). Knowledge regarding how communities perform during and after EEs can inform proactive resilience enhancement practices. However, the EE impacts on impacted communities are influenced by multiple variables and their implicit interactions, which are difficult to be understood and modeled with conventional mathematical or statistical models. Thus, we calibrate a spatio-temporal deep learning model to capture the dynamics of impacted communities and use explainable artificial intelligence (XAI) approach to interpret the influence of communities' social and physical properties on EE impacts. Specifically, we couple graph convolutional neural network (GCN) and long short-term memory (LSTM) to model the mobility dynamics for 666 census tracts (i.e., communities) in 14 medium-sized US cities affected by recent hurricanes. The model takes both static variables characterizing communities' preexisting conditions and dynamic variables depicting the changing hazard exposure. The interpretation of the model predictions based on DeepLIFT shows that community resilience, inferred from the perturbation of human mobility in this research, is highly event and geography dependent. Variables including walkability, green spaces, and civil participation generally contribute to fewer mobility perturbations, i.e., resilience contributive, while automobile-oriented accessibility and social vulnerability lead to more mobility perturbations when hazard conditions are controlled. This study empirically validates the mediative role of communities' social and physical properties on EE impacts and promotes more data-driven approaches for understanding and anticipating complex, dynamic, and place-specific community resilience.