A Bayesian-Based Approach for Post-disaster Recovery Estimation Enhancement

Each year, numerous disasters occur, challenging the resilience of societies worldwide. The effective management of the intricate disaster recovery process necessitates well-prepared plans, robust response strategies, and adequate resources. The recovery process can vary drastically among communities. However, the existing loss models and recovery frameworks rely on predefined loss and recovery functions, limiting their ability to accurately estimate recovery. Additionally, these predefined models encounter challenges in characterizing the complex recovery process due to analytical limitations and the unique nature of recovery in various communities. This study explores a data-centric approach to quantitatively estimate post-disaster recovery. It introduces a Bayesian Neural Network (BNN) model capable of providing near real-time evaluations of the recovery process. Synthetic data has been generated for two different reconstruction modes: in-situ and reallocated reconstruction, as adopted in two towns following the 2008 Wenchuan, China earthquake. In this context, Yingxiu town represents in-situ reconstruction, while Yongchang town represents reallocated reconstruction. The BNN model allowed for uncertainty characterization during the estimation and forecasting of recovery for two different communities. The recovery curves generated using BNN yielded confident recovery predictions with only 30–40% of the training data. This underscores the potential of the BNN approach as an effective post-disaster recovery estimation tool, particularly when integrated with other resiliency estimation models.