H211-05
Identifying Key Damage Drivers of Atmospheric River-Induced Flooding in Northern California

Wednesday, 16 December 2020: 17:46
Virtual
Corinne Bowers1, Katherine Serafin2 and Jack W Baker1, (1)Stanford University, Stanford, CA, United States, (2)University of Florida, Department of Geography, Ft Walton Beach, FL, United States
Abstract:
Atmospheric rivers (ARs) are long, narrow bands of water vapor carrying moisture from the tropics to the midlatitudes, responsible for over three-quarters of extreme precipitation events in California and over 90% of the state’s record floods. AR intensity metrics alone are insufficient to capture on-the-ground effects; while AR severity has recently been categorized into a scale using intensity and duration, there is still wide variability within categories in how much damage an AR may cause. Similarly, traditional techniques to evaluate damage such as depth-damage curves use only one or two input variables and neglect critical information specific to ARs (e.g. amount of precipitable water) and specific to the affected location (e.g. resident flood experience).

Here we present a method that combines all available hazard, exposure, and vulnerability information to capture AR-induced flooding in northern California and determine which factors are the strongest predictors of flood damage at the community level. We do so with ensemble decision tree regression models, using flood insurance claims from the National Flood Insurance Program (NFIP) as the response variable. This bottom-up approach to risk analysis, starting with a measure of flood damage and working backwards to identify the factors that contribute the most to community flood loss, is a fundamentally novel approach to AR damage estimation.

Multiple functional forms are tested, including Random Forest, Multiple Adaptive Regression Splines, and boosted trees. The outputs of the final best-fit model are metrics of variable explanatory power, significant interaction terms, and threshold values where damage increases sharply at a “tipping point” for a given predictor. These decision tree-based models are robust to outliers and collinearity, and the data-driven approach proposed here can capture complex relationships between inputs, including metrics of social vulnerability, that would be computationally intractable to account for in a physics-based simulation. Cities and counties that experience repetitive AR-driven flooding, such as Sonoma and Sacramento, can use these results as a tool to identify areas of high risk within their communities and guide future mitigation investments.