Interpretable data-driven model for Climate-Induced Disaster damage prediction: The first step in community resilience planning

Abstract

The frequency and magnitude of Climate-Induced Disasters (CID) have been increasing consistently over the past few decades. Alleviating the impacts of such disasters is thus critically important. Hence, a systematic data-driven framework is developed to predict CID-related damages. The framework encompasses four phases: (1) Data Collection and Fusion, where spatial interpolation methods are employed to integrate data from multiple sources, (2) Feature Selection, which aims at comparing several filter, wrapper, and embedded bbmethods to select relevant features, (3) Model Development, where machine learning techniques are employed to develop the prediction model, and (4) Result Analysis and Model Interpretation, where the black-box nature of the model is decoded using different interpretability techniques. To demonstrate the framework's utility, wind disaster property damages were linked to hazard, climate, land cover, social, housing, demographic and economic data recorded in the state of New York from 2010 to 2018. Features that are significantly important for property damage prediction were selected, and a set of machine learning models were subsequently developed. The best performing model was a random forest-based regression tree, and yielded a coefficient of determination of 0.79. Expectedly, property damages were found to depend on the complex interplay between disaster, climate, socioeconomic, housing, and demographic conditions rather than on the hazard characteristics only. The developed framework is the first step in community resilience planning, where CID recovery time and associated costs, together with the predicted damage extent and the specificities of the affected communities, can be linked to quantify community resilience under future climate-induced hazards.