Uncertainty‐aware convolutional neural network for explainable artificial intelligence‐assisted disaster damage assessment

Abstract

Accurate damage assessment is a critical step in post-disaster risk assessment, mitigation, and recovery. Current practices performed by experts and reconnaissance teams in the form of field evaluation require considerable time and resources. Recent advances in remote sensing imagery, artificial intelligence (AI), and computer vision have enhanced automated and rapid disaster damage assessment. Recent literature has shown promising progress in AI-assisted aerial damage assessment. However, accounting for the uncertainty in the outcome for improved quantification of confidence and enhanced model explainability for human decision-makers remains one of the key challenges. Overlooking uncertainty can lead to erroneous decisions, especially in highly-consequential tasks such as damage assessment. The aim of this study is to develop uncertainty-aware deep learning models for the assessment of post-disaster damage using aerial imaging. Within the framework of variational Bayesian inference, Monte Carlo dropout sampling technique is used to propagate epistemic uncertainty in model predictions. With this stochastic setting, the model produces damage prediction labels with softmax as random variables, which helps quantify confidence in the model outcome using appropriate measures of uncertainty. Two networks are implemented and trained separately on two different disaster damage datasets consisting of unmanned aerial vehicle building footage as well as satellite-captured post-disaster imagery. The first network attains 59.4% accuracy in building classification, and the second network gives an accuracy of 55.1%. Results from uncertainty analysis, model confidence quantification, and analyzing model attention zone can lead to more explainable and risk-informed automated damage assessment outcomes using AI technology.