Intelligent assessment of building damage of 2023 Turkey-Syria Earthquake by multiple remote sensing approaches

The Turkey–Syria region lies at the junction of the African, Anatolian, and Arabian plates within the Alpine–Himalayan seismic belt and is highly tectonically active. On February 6, 2023, a Mw7.8 mainshock along the East Anatolian Fault (EAF) and a Mw7.5 event nine hours later caused widespread destruction, >44,000 fatalities, and >160,000 damaged or collapsed buildings. This study addresses the need for rapid, objective, large-scale, and building-level assessment of earthquake-induced building damage using widely available satellite data. The research question is whether synergizing multi-sensor remote sensing indicators with seismic shaking (PGA) in an AI-based framework can improve multi-class building damage detection (no, slight, serious) relative to conventional single-index approaches (e.g., DP alone). The purpose is to deliver a quantitative, automatic, and scalable model to support emergency response and prioritization of rescue efforts.

Remote sensing is a standard tool for post-earthquake damage assessment due to its wide coverage and all-weather capability. Optical imagery (e.g., Sentinel-2, Landsat) provides spectral information for surface change detection and indices such as the normalized difference built-up index (NDBI) to highlight impervious surfaces, though clouds/snow can hinder acquisition. SAR enables observations independent of cloud cover and provides two commonly used damage indicators from InSAR stacks: damage proxy (DP, differential coherence) and amplitude dispersion index (ADI), both sensitive to abrupt surface changes. Prior studies have applied SAR and optical data for damage mapping and explored AI/ML in natural hazards, including supervised and unsupervised learning for event detection and impact estimation. However, a gap remains in differentiating damage severity at the individual-building scale using combined multi-source features, motivating the proposed MCDD approach.

Data and study area: The study focuses on the 2023 Turkey–Syria earthquake sequence, with detailed analysis in Kahramanmaraş. Inputs include SAR and optical satellite data and PGA. Sentinel-1A/B C-band SAR scenes: 7 frames across 4 paths (ascending Path 14 frames 114, 119; ascending Path 116 frame 114; descending Path 21 frames 465, 471; descending Path 123 frames 466, 471). Temporal interval: 12 days; three scenes per frame (two pre- and one co-/post-seismic, Feb 9–17, 2023). ALOS-2 PALSAR-2 L-band SAR: descending Path 78 Row 2860, images from 04/07/2021, 04/06/2022, 02/08/2023 (10 m). Sentinel-2 optical: frame 37SCB over Kahramanmaraş on 01/20/2023 (pre), 02/09/2023 and 02/14/2023 (post). PGA: USGS published contours interpolated to continuous fields. Feature extraction: - DP (differential coherence) from SAR/InSAR: Coherence computed using GMTSAR with multilooking (4 range × 1 azimuth) yielding ~20 m spacing. Histogram matching (imhistmatch in MATLAB) calibrated co-seismic coherence to pre-seismic distributions to reduce systematic bias across frames. Differential coherence (post minus pre) used as damage proxy; negative values set to zero (causality constraint); maps normalized. Generated from both Sentinel-1 and ALOS-2. - ADI (Amplitude Dispersion Index) from Sentinel-1 RTC amplitude: Used ASF HyP3 Radiometrically Terrain Corrected products (10 m), terrain-corrected with SRTM DEM, speckle filtering and radiometric calibration (sigma0). Computed ADI from six amplitude scenes (five pre-event, one post-event). Higher ADI indicates greater disturbance. - Differential NDBI from Sentinel-2: Level-2A surface reflectance bands resampled to 10 m; DN converted to reflectance using scale factors; outliers clipped. NDBI = (SWIR − NIR) / (SWIR + NIR). Snow/ice and clouds masked; differential NDBI computed for 02/09/2023 − 01/20/2023 and validated by 02/14/2023 − 01/20/2023. Zonal statistics extracted features per building footprint/pixel. Ground truth labels: Ministry of Environment and Urbanization of Turkey building damage assessments (https://hasar.6subatdepremi.org/) with five original classes (no, slight, heavy, to be demolished, collapsed). For modeling, simplified to three: 0 no damage; 1 slight damage; 2 serious damage (heavy + to be demolished + collapsed). Class imbalance noted: 64% no damage, 24% slight, 8% heavy, 2% to be demolished, 2% collapsed (in 24,352 building pixels in Kahramanmaraş). Preprocessing for ML: All features (Sentinel-1 ADI, Sentinel-1 DP, ALOS-2 DP, differential NDBI, PGA) resampled onto a common grid of ~0.00027° × 0.00027° (~30 m) via nearest neighbor. Dataset split: 80% training, 20% testing. Imbalanced-learn undersampling applied to majority (no-damage) class to mitigate bias. Model and evaluation: Multiclass classification implemented via OneVsRestClassifier with Random Forest Classifier (scikit-learn), fixing random_state for reproducibility. Trained three binary classifiers (each class vs all others). Performance evaluated using ROC-AUC (area under the ROC curve), considering sensitivity and specificity. Additional experiments (n=100) compared the full-feature model vs DP-only baseline. Correlation analyses: Spatial comparisons of DP, ADI, and PGA conducted, including a profile (AA′) crossing Kahramanmaraş and Gaziantep to examine variations relative to the EAF and Engizek Fault zones. Histograms of PGA and ADI conditioned on official damage levels examined. High-priority site inspection: A building cluster near Highway 835 in Kahramanmaraş analyzed with Google Earth imagery and overlaid DP (ALOS-2, Sentinel-1), ADI (Sentinel-1), and differential NDBI (Sentinel-2) to qualitatively verify indicator sensitivity at building scale.

- SAR-derived indicators: DP maps (Sentinel-1, ALOS-2) and ADI delineate major damaged zones: central Kahramanmaraş, western Gaziantep, northern Adana, southern Kayseri, Hatay (Antakya), northern Aleppo; damage patterns are constrained by faults (EAF, Engizek, Sariz, etc.). Some path-boundary discontinuities exist due to acquisition differences. - Building-scale detection: 10 m ADI pinpoints partial collapses within building blocks; ALOS-2 10 m DP outperforms Sentinel-1 20 m DP for detailed collapse/deformation identification; differential NDBI highlights deformed roofs but is susceptible to cloud/snow effects. - PGA–damage relationship: Histograms of PGA by damage class peak at approximately 0.26 g (no damage), 0.29 g (slight), and 0.34 g (serious), indicating higher PGA correlates with greater damage. ADI histograms peak at normalized ADI ≈ 0.14 (no), 0.17 (slight), 0.20 (serious). Along profile AA′, the largest DP and high PGA coincide near the EAF; ADI and PGA decay away from the fault. - MCDD model performance: Multifeature model achieves ROC-AUC up to 0.69 and exceeds a DP-only model by 11.25% (based on 100 runs). Misclassifications and low-confidence predictions are concentrated in the slight-damage category. - Feature importance: PGA is the most important predictor for regional differentiation of damage levels; the four remote sensing features (Sentinel-1 DP, ALOS-2 DP, ADI, differential NDBI) have similar importance within ~3% of each other, attributable in part to resampling all inputs to ~30 m. - Data characteristics: Official damage inventory shows strong class imbalance (approx. 64% no, 24% slight, 12% serious when aggregating the last three levels), necessitating undersampling of the majority class during training.

The study demonstrates that combining SAR- and optical-derived indices (DP, ADI, differential NDBI) with seismic shaking (PGA) in a supervised multiclass ML framework significantly improves rapid, automated assessment of building damage severity over single-index baselines. The geospatial consistency observed among high PGA, proximity to active faults (e.g., EAF), and elevated DP/ADI supports the physical linkage between shaking intensity and surface/building disturbance. However, shaking intensity and damage are modulated by complex geologic and topographic factors (basin amplification, mountain shielding), rupture dynamics (sub-/super-shear propagation, peak slip), and building practices, meaning distance to faults alone is not a sufficient predictor. At regional scales, 20 m DP effectively prioritizes heavily affected districts; at building scales, 10 m ADI resolves partial collapses. Optical differential NDBI contributes uniquely to detecting roof-level changes but suffers from cloud/snow contamination and cannot sense under-roof structural failures. The model’s slight-damage class is the most challenging, reflecting both intrinsic ambiguity in moderate damage and potential subjectivity/noise in ground-truth labels. The importance of PGA in feature ranking underscores the value of integrating seismic shaking metrics. The near-equal importance among remote sensing features suggests complementary information content after harmonization to ~30 m, though this resampling suppresses the native 10 m advantages of ADI and ALOS-2 DP. Anticipated advances (e.g., NISAR dual-band SAR, HLS for higher revisit optical) will enhance spatial/temporal resolution, reduce cloud gaps, and further improve model reliability and timeliness for emergency response.

This work introduces an AI-enabled Multi-Class Damage Detection (MCDD) approach that fuses Sentinel-1/ALOS-2 SAR-derived DP, Sentinel-1 ADI, Sentinel-2 differential NDBI, and USGS PGA to classify building damage into no, slight, and serious categories across the 2023 Turkey–Syria earthquake zone. The method delivers quantitative, automated mapping from publicly available data, scales from tectonic to building level, and improves performance over DP-only baselines by 11.25% (ROC-AUC up to 0.69). It supports rapid prioritization of rescue efforts and refinement of official damage inventories. Future directions include: leveraging higher-resolution and higher-revisit missions (NISAR, HLS), integrating ultra-high-resolution optical (shadow-based/object-based change detection), LiDAR-derived height changes, and additional geotechnical layers; improving ground-truth quality and class definitions; addressing class imbalance with advanced resampling or cost-sensitive learning; and exploiting native higher-resolution features without downsampling to preserve fine-scale sensitivity.

- Ground truth limitations: Official labels may contain inconsistencies and subjectivity (unclear criteria for “to be demolished”; examples where visually damaged buildings are labeled no-damage), contributing to model confusion, especially for slight damage. - Data imbalance: Strong class imbalance toward no-damage required undersampling, which may reduce training diversity and affect generalization. - Spatial resolution harmonization: Resampling all inputs to ~30 m diminishes the native 10 m advantages of Sentinel-1 ADI and ALOS-2 DP for building-scale detection. - Remote sensing constraints: Optical imagery affected by clouds, shadows, and snow masking uncertainties; NDBI changes can be confounded by snow extent variations. SAR path boundaries can introduce discontinuities; water bodies may yield high ADI unrelated to damage. - Model performance: Moderate ROC-AUC (~0.69) indicates room for improvement; slight-damage class exhibits highest misclassification and lowest confidence. - Observational geometry: Nadir optical observations cannot detect under-roof damage (e.g., pancake collapse). Temporal coverage limited by satellite revisit schedules soon after the event.