Quantifying earthquake-induced building damage from VIIRS nighttime lights: A multi-scale framework with deep learning–based area validation

Rapid and reliable assessment of earthquake-induced building damage is essential for effective emergency response, yet the quantitative relationship between nighttime light (NTL) variations observed by satellite sensors and physical structural damage remains insufficiently constrained by physically consistent validation. This study addresses this limitation by establishing a multi-scale analytical framework that explicitly links regional-scale VIIRS NTL observations (∼500 m resolution) with local, area-based building damage measurements derived from very high resolution (0.5 m) WorldView-2 imagery and deep learning–based semantic segmentation. Using the 2023 Türkiye earthquake as a case study, we introduce a VIIRS-based Normalized Difference Light Index (NDLI) to mitigate baseline luminosity heterogeneity and evaluate its relationship with two independent damage indicators: (i) the percentage of collapsed buildings obtained through visual interpretation, and (ii) an area-based Building Damage Index (BDI) computed from U-Net–based semantic segmentation of building footprints, which achieved an Intersection over Union (IoU) of 0.7819 and an F1-score of 0.8776 on the test dataset. Regression analysis reveals a distinct non-linear exponential relationship between NDLI and building damage severity, with a critical transition threshold at NDLI ≈ −0.3, beyond which damage indicators increase sharply. The area-based BDI demonstrates a stronger correspondence with NDLI variations (R 2 = 0.640, p < 0.001) than the count-based collapse metric (R 2 = 0.578, p < 0.001), consistent with the radiometric principle that satellite-measured radiance responds primarily to the spatial extent of illuminated structures rather than their numerical count. These results demonstrate that VIIRS-derived NDLI serves as a physically interpretable proxy for rapid, large-scale assessment of earthquake-induced building damage when validated using area-based metrics, while acknowledging that the derived empirical relationships are specific to the study area conditions and require multi-event validation for broader applicability.