| description abstract | The motivation for this study is the damage caused to over 300 long-span bridges in Vermont in 2011 by Tropical Storm Irene, which is part of the Hurricane Irene system. This study of the effects of this single extreme flooding event uses multiple variables related to these damaged bridges including bridge characteristics, stream characteristics, geographical features, environmental factors, and land-use resulting in a dataset with over 300 features. A new, evolutionary algorithm performed multivariate feature selection on this dataset to circumvent the computational challenges associated with using traditional statistical analysis. This impartial and exhaustive search of feature combinations generated nonlinear models that best predict bridge damage and identify bridges most at risk. Maps of statewide bridge vulnerability showed good correlation to damaged bridges. Features identified as significantly correlated with bridge damage include Irene stream power, and newly identified variables, such as watershed hydrologic soil types and the geographical context of specific watersheds. Channel rating and waterway adequacy rating, variables from the bridge inspection manual, also proved important. This research has applications beyond Vermont, as many of the newly identified variables can be created or monitored using commonly available data, and the analytical framework extends to other infrastructure assessment efforts at a system level. | |
