Training-free post-disaster damage estimation with incomplete satellite images by integrating open data and fractional hot deck imputation

Abstract

Analyzing satellite imagery is an essential step after a severe natural disaster for monitoring the level of damage and making the best decision. The current study proposes a fast post-disaster damage framework incorporating the remote sensing of imagery, entropy calculation, object detection using an advanced machine learning (ML) model, and missing data curation. The present framework includes a combination of open-source data for high-resolution imagery, digital elevation model (DEM), building footprint, and dual-polarization synthetic aperture radar (SAR) components. For damage estimation at the building level, the first phase of the proposed framework is the entropy difference calculation between the pre-disaster and post-disaster imagery. Followed by feature collection from open-source platforms that have been used for predicting real-world scenarios where data may be missed, for example, by cloud cover after the disaster, which prevents the estimation of the damage level by satellite. In the current study, the missing values were randomly assumed in the target variable $$\Delta H$$ , followed by curing and imputation methods for predicting the missing values. The imputation methods used in this study are the Fractional Hot Deck Imputation (FHDI), Fully Efficient Fractional Imputation (FEFI), and for validation Naïve method, and a deep learning (DL) model. The selected study area is Lake Charles, Louisiana, which was impacted by the severe Hurricane Laura. In addition, the research compares the performance of $$\Delta H$$ against Kullback-Leibler divergence ( $$D_{KL}$$ ) and synthetic aperture radar (SAR) channels in terms of accuracy in damage detection. The results show that SAR polarization channels are appropriate for flood mapping purposes, while $$\Delta H$$ demonstrated higher accuracy in damage detection than $$D_{KL}$$ , and proved to be robust to changes in spatial resolution and urban density, matching Federal Emergency Management Agency (FEMA) damage classification. The results of this study highlight the effectiveness of both FHDI and FEFI in curing and predicting the missing target variable $$\Delta H$$ , reducing the error by approximately 14%, and 10% respectively, for a high missing rate of 50% compared to the Naïve method, and a DL model, respectively. The results highlight the ability of the proposed methodology for fast, training-free damage assessment and the adaptability of the framework to other natural disasters such as wildfires.