A Probabilistic Flood Loss Model for Adaptation Planning in Ho Chi Minh City

Transforming rural-urban-systems such as Ho Chi Minh City, Vietnam, are facing exacerbating flood risk due to climatic and socio-economic changes, necessitating effective adaptation solutions. Risk-based adaptation planning requires plausible and accurate flood loss estimation. However, state-of-the-art flood loss models for the region that take into account the multi-causality of flood damage and convey information about predictive uncertainty are lacking.

This study presents a Bayesian network for flood loss estimation for the residential sector in Ho Chi Minh City. We developed the graphical probabilistic model based on new object-level survey data with flood-affected households (n=1530), which cover the topics of flood intensity, household characteristics, warning and emergency, private precaution, and damages. An analysis of the survey data concerning the explanatory power for flood damage revealed a subset of relevant variables, which we used for model elicitation. Using a systematic learning procedure, we identified a robust Bayesian network structure that reflects the local circumstances of flood damage processes at the study site. That is, the resulting model takes into account flood intensity variables such as water depth but also vulnerability variables such as households’ flood experience or adaptive behavior. We confirmed the identified damage influencing variables by comparisons to other established statistical and machine learning methods (i.e., random forest and grid search cross-validation with multivariable regression). A prediction exercise with repeated cross-validation demonstrated that the developed Bayesians network model is capable of estimating building loss accurately. However, similar to previous studies in the field, we observed considerable predictive errors for severe loss cases for which data records are scarce. In addition, we show that the predictive skill of the Bayesian network is competitive to non-parametric modeling alternatives such as random forest.

Our validated Bayesian network loss model exhibits high practical value for applications at the city-scale since it enables loss estimation even when information about the predictor variables is only partially available. Moreover, the inclusion of vulnerability variables as predictors in the loss model facilitates the consideration of adaptive behavior in loss and risk assessment. Ultimately, the fully probabilistic model design inherently quantifies predictive uncertainty, which fosters the uncertainty propagation to subsequent elements of flood risk assessment and well-informed decision-making.