Robust disaster impact assessment with synthetic control modeling framework and daily nighttime light time series images

Abstract

Remotely sensed nighttime light (NTL) has been acknowledged as an ideal proxy of the extent and intensity of human activity. One of its main NTL-based applications is to assess disaster impacts; nevertheless, the full potential of NTL-based disaster impact assessment has been largely constrained due to the uncertainties in estimating business-as-usual (BAU) NTL intensity (i.e., the counterfactual condition with no disaster occurrence) and hurdles in isolating the disaster impact from other cocontributing factors of NTL changes. To address these issues, we adopted the synthetic control (SC) modeling framework to construct a robust estimation of BAU NTL with daily NTL images from NASA’s Black Marble VIIRS product. We further improved the traditional SC model by optimizing donor selection with the dynamic time warping algorithm (DTW) and incorporating random forest regression to better capture target-donor relationships. Applying our model to 20 severe disasters across geographies, types, magnitudes, and socioeconomic contexts, our model significantly outperformed existing approaches, with an average correlation coefficient of 0.94 against reference and a 0.47% difference of covariates. Besides, our model showed a robust performance in detecting disaster impacts with a low impact intensity and short-term impact duration, which were largely under-detected by existing approaches. The resulting disaster impact assessment metrics, including impact duration, impact intensity, and impact severity, provided further insights into the substantial heterogeneity in disaster coping capability and socioeconomic resilience across regions. Our proposed model holds a broad significance in supporting not only strategic and effective disaster relief but also achieving ambitious climate resilience and sustainability goals.