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

The devastating Mw7.8 earthquake that struck southeast Turkey and northwest Syria on February 6th, 2023, resulted in over 44,000 deaths and the collapse of more than 160,000 buildings. Traditional methods of assessing earthquake-induced building damage are subjective, time-consuming, and often hampered by accessibility issues and a lack of immediate high-resolution imagery. The Turkey-Syria region's location at the intersection of three tectonic plates (African, Anatolian, and Arabian) makes it highly seismically active. The Mw7.8 earthquake, originating from the East Anatolian Fault (EAF), caused widespread destruction. A subsequent Mw7.5 earthquake further exacerbated the damage. This catastrophic event highlights the urgent need for rapid and accurate damage assessment to prioritize rescue efforts and aid in efficient resource allocation. This study addresses this need by developing an advanced, AI-driven method that leverages multiple remote sensing approaches to efficiently and quantitatively assess building damage on both a large-scale (tectonic) and individual-building scale.

Existing research demonstrates the utility of various remote sensing techniques in earthquake damage assessment. Unmanned aerial vehicles (UAVs), Light Detection and Ranging (LiDAR), Synthetic Aperture Radar (SAR), and optical images have been employed to identify damage and surface disturbances. Optical images capture surface changes through electromagnetic radiation, using indices like the Normalized Difference Built-up Index (NDBI) to identify man-made structures. However, optical imagery is susceptible to cloud cover. SAR, on the other hand, penetrates clouds and rain, making it a valuable tool in inclement weather. Interferometric SAR (InSAR) uses multiple SAR images to detect ground deformation with high accuracy. Damage Proxy (DP) and Amplitude Dispersion Index (ADI), derived from SAR data, are effective indicators of land surface changes. Artificial intelligence (AI), particularly machine learning (ML), has emerged as a powerful tool for automated decision-making in Earth science, enabling autonomous learning and analysis of remotely sensed data for various applications including winter storms, landslides, urban flooding, and land cover change. However, there remains a need for more accurate and efficient methods for assessing the severity of damage to individual buildings post-earthquake, a critical factor in optimizing disaster response.

This study introduces a multi-class damage detection (MCDD) model based on supervised machine learning. The model integrates data from multiple remote sensing sources: 1. **SAR Data:** Sentinel-1A/B provided C-band SAR images, processed to generate DP and ADI maps. ALOS-2 PALSAR-2 provided additional L-band SAR data with 10m resolution, contributing higher-resolution DP data. Pre- and post-earthquake images were used to calculate DP (change in SAR coherence) and ADI (temporal variance in SAR amplitude). Histogram matching was used to calibrate the post-seismic coherence map, reducing bias and discontinuities. A causality constraint adjusted negative values in DP to zero. 2. **Optical Data:** Sentinel-2A/B provided multi-spectral optical images, used to calculate differential NDBI (post-seismic minus pre-seismic). Snow and ice areas were masked out. Differential NDBI was calculated using SWIR and NIR bands. 3. **Peak Ground Acceleration (PGA):** PGA data from USGS was incorporated as an independent indicator of ground shaking and potential damage. These indices were then used as features in the MCDD model. All features were resampled to 30m resolution. The model used a Random Forest Classifier within a OneVsRestClassifier framework to handle multi-class classification (no damage, slight damage, serious damage). The dataset was addressed for class imbalance using undersampling. Model performance was evaluated using the ROC-AUC metric.

The results showed that the MCDD model effectively classifies building damage levels. The DP and ADI maps from SAR data clearly highlighted the severely damaged areas in central Kahramanmaraş, western Gaziantep, and other affected regions. These areas were spatially correlated with tectonic faults. The ADI map, particularly with its 10m resolution, accurately identified building damage, showing higher values in collapsed areas than in standing structures. ALOS-2's higher-resolution DP data outperformed Sentinel-1's in identifying building collapses and deformation. Differential NDBI effectively detected deformed roofs, although optical imagery was limited by cloud cover. A strong correlation was observed between building damage levels and PGA, with higher PGA corresponding to more severe damage. The MCDD model, integrating multiple remote sensing indices (ADI, DP, and NDBI) and PGA, achieved a ROC-AUC of 0.7, significantly outperforming the model using only DP by 11.25%. Feature importance analysis showed PGA as the most important feature for regional-scale damage assessment. Spatial correlation analysis revealed a relationship between damage levels and distance from faults, with closer proximity correlating to more significant damage. However, this relationship is not straightforward and can be influenced by factors like topography. The study highlighted potential biases in official damage assessments, underscoring the potential of the MCDD model to improve the accuracy of damage classification.

The findings of this study demonstrate the effectiveness of combining multiple remote sensing approaches and AI for rapid and accurate assessment of building damage after earthquakes. The MCDD model's superior performance compared to traditional methods using DP alone highlights the value of integrating diverse data sources. The incorporation of PGA, a direct measure of ground shaking, added significant value to the model. The spatial correlation between damage and distance to faults confirms existing knowledge, but emphasizes the complexity of seismic hazard assessment, requiring consideration of factors such as topography and soil conditions. The identified biases in official damage assessments highlight the potential of the MCDD model to enhance the accuracy and reliability of post-earthquake damage assessment. This, in turn, can lead to more efficient resource allocation and improved disaster response.

This research successfully developed and validated an AI-powered MCDD model for assessing earthquake-induced building damage. The integration of multiple remote sensing indices and PGA significantly improved the accuracy of damage classification compared to using DP alone. This model has the potential to assist in efficient disaster response by enabling rapid and objective damage assessment. Future work could explore the incorporation of other data sources, such as high-resolution optical imagery and LiDAR, to further refine the model's accuracy and capability. The development of more sophisticated AI models and the integration of data from upcoming satellite missions, such as NISAR and the improved Harmonized Landsat Sentinel-2 (HLS) data, could lead to even more precise and timely damage assessments.

While the study achieved promising results, several limitations should be considered. The accuracy of the model is dependent on the quality and availability of remote sensing data. Cloud cover can affect optical imagery, limiting its use in certain areas and time periods. The 30m resolution resulting from resampling may have reduced the benefits of higher-resolution data. Finally, the accuracy of the model also depends on the accuracy of the ground truth data used for training and validation, which itself might be subjective.