A clustering-classification approach in categorizing vulnerability of roads and bridges using public assistance big data

Abstract

In the US, the state, local, territorial, and tribal governments seek funding from the federal government to repair, restore, reconstruct, or replace the disaster damaged infrastructures through Public Assistance program. This research has utilized historical public assistance funded projects’ data in an unsupervised machine learning algorithm: agglomerative hierarchical clustering and uncovered three distinct levels of vulnerability among the roads and bridges in different counties under different natural hazard conditions. The paper shows the spatial distribution of the different levels of vulnerability. The difference in the extent and severity of infrastructure failure among different vulnerability levels was found to be statistically significant. In the second part of the research, a random forest classification model is utilized to explore whether the number of disasters, the area and population of a county, and the existing condition of roads and bridges in a county can predict the vulnerability level of the county as derived from the clustering model to cross-validate them. It has been found that the classification model can predict the vulnerability levels with 85% accuracy. It was also found that the differences in the condition of roads and bridges among the risk profiles are statistically significant and higher risk profile counties have worse roads and bridges. The segmentation of roads and bridges into different vulnerability levels can help the decision makers in planning for asymmetric initiatives for roads and bridges in different counties. Moreover, it has been established that improving the existing conditions of infrastructures can help in reducing the cost of public assistance in a county.