← Back to Sessions

How to Quantify Uncertainties for Logistic-Regression-Based Geospatial Natural Hazard Models?

Description: 

Global natural hazard models are data-driven models that take advantage of broadly available geospatial proxies to predict natural hazard phenomena and are wildly used in both pre-disaster planning and post-disaster responses. It remains less explored on how to quantify uncertainties for such geospatial natural hazard models. Taking the logistic regression based global geospatial liquefaction model (GGLM) (Zhu et al., 2017) as an example, we propose an uncertainty quantification (UQ) framework that consists of uncertainty source characterization, global sensitivity analysis, and forward uncertainty quantification. The GGLM predicts the liquefaction probability using the ShakeMap's PGV, slope-based V_S30, distance to water body, water table depth, and annual precipitation as explanatory variables. First, we characterize three sources of uncertainties: parametric uncertainty, model error, and geospatial input uncertainty. We use Bayesian inference to quantify the posterior distribution of model parameters and their variations with the sample size, and find that the parametric uncertainties are small when a large dataset is used to learn the model parameters. Model errors are calibrated as a normal distribution based on a novel classification residual analysis. The geospatial input uncertainties are characterized using the literature and expert judgement. Second, we use the Sobol’s method to investigate the sensitivity of model output to different uncertain inputs and find that the variance of model output is largely contributed by the geospatial input uncertainties and model errors. Last, we propose an approximation-based forward uncertainty propagation (FWD) method, which shows consistent results as the Monte-Carlo-Simulation-based method but much better computational efficiency. Given that the logistic regression models are wildly used in the earth science filed, we conclude this UQ framework could also be applied to other geospatial modeling problems.

Session: Coseismic Ground Failure: Advances in Modeling, Impacts and Communication [Poster]

Type: Poster

Date: 4/20/2023

Presentation Time: 08:00 AM (local time)

Presenting Author: Weiwei Zhan

Student Presenter: No

Invited Presentation: 

Authors