Enhanced prediction of highway flood inundation through Bayesian generalized linear geostatistical models

Transportation infrastructure facilitates the mobility of goods and humans. Following flooding, blocked road access would prevent vulnerable communities from accessing essential services and disaster relief resources. To reduce the impact of damaged transportation infrastructure on community lifelines, efficient infrastructure restoration is desired. Conventionally, damage identification is often performed via field inspections. However, due to the blocked road access and safety concerns, a limited amount of damage inspection data can be collected immediately following flooding. Aimed at providing a quick prediction of highway inundation status, this research proposes a novel approach that integrates geospatial correlation to address the data sparsity issue. At the core of this approach is a Bayesian generalized linear geostatistical model (BGLGM) that measures (1) correlations between highway inundation status and the associated geospatial variables (e.g., road elevation and flood depth), (2) spatially correlated residuals that cannot be explained by the geospatial variables, and (3) parametric uncertainties. To verify and validate the proposed approach, a case study on highway flood inundation in Harris County, Texas, following Hurricane Harvey was conducted. A sensitivity analysis of the model performance to the availability of damage inspection data was conducted. The results show that the proposed approach is capable of providing accurate highway inundation prediction using limited damage inspection data, which validates the concept of integrating geospatial correlation for more accurate highway inundation representation and prediction. In addition to supporting rapid damage inspection, the validation opens up possibilities for integrating geospatial correlation into machine learning and deep learning models to enhance model performance. The region-specific geospatial correlation also has the potential to recalibrate pre-trained models, improving their generalizability to other regions.