In the immediate aftermath of an earthquake, timely and reliable assessment of the impact in terms of the damage sustained by built assets and the associated repair and replacement costs assumes a crucial role in the strategic organization and prioritization of the response and recovery efforts. Selection of a particular approach for damage and loss evaluation typically involves a trade-off between the level of details collected and the timeliness of evaluation.

Within the public sector, estimation of damage and loss after a natural disaster is typically performed through field missions using paper forms or, more recently, data capture tools on tablet and mobile devices . This approach requires the mobilization of technical experts to the affected areas. While field surveys still remain the most accurate solution, they can be highly resource intensive and time-consuming for disasters affecting large areas. Lack of available and experienced personnel can also be an issue especially for large-scale events. Furthermore, for some disasters it might not even be possible to access some of the affected regions due to disruption of the transportation network. Completing such detailed damage assessments for moderate to large events can take weeks, if not months . For instance, collection and compilation of building-level damage data after the 2015 Mw 7.8 Gorkha earthquake and landslides in Nepal continued for over 7 months .

Remote-sensing data offer advantages over ground inspection in terms of its collection speed and spatial coverage, particularly when the emphasis is less about producing a detailed and highly accurate building-level inventory of damage but on providing rapid information about the potential locations, extents, and severity of damage. In recent years the improvement of the level of detail, spatial resolution, and reduced latency of Earth-observation (EO) data has encouraged a variety of applications in disaster damage assessment. , , , and provide thorough reviews of current methods for earthquake-induced building damage detection using EO data. Optical EO data are currently available at sub-meter spatial resolution from several satellites, making them particularly appealing for building-level damage detection methods. Optical imagery is also conducive for applications involving visual interpretation, such as crowd-sourced damage labeling. There are presently a few rapid damage mapping services in operation which use primarily optical EO data as the basis for damage assessment. These include the UNITAR Operational Satellite Applications Programme (UNOSAT) of the United Nations and the Copernicus rapid damage mapping service supported by the European Commission . These damage assessments by UNITAR and Copernicus need significant manual effort to scan the raw optical imagery covering the affected area for collapsed buildings, signs of debris or other visible damage, and cracks in bridges and other infrastructure elements.

As an alternative, radar EO data are provided by active airborne or spaceborne radio detection and ranging sensors. These sensors emit pulses of microwave radiation towards a target on the Earth's surface, which reflects back some of the emitted energy. Synthetic aperture radar (SAR) is a technology used to exploit the continuous transmission and reception of radar pulses to and from a radar imaging system mounted on a moving platform, such as an airplane or a satellite. Processing of the signals from the multiple pulses received from the same target but at different relative locations of the sensor can in effect help create a larger synthetic aperture that can allow capturing images at a much higher resolution than a similar stationary antenna. Light detection and ranging (lidar) sensors work in a manner similar to radar sensors, using laser pulses instead of radio pulses. For damage detection applications, lidar data are typically obtained through airborne sensors that can collect 3D data in the form of point clouds. Pre-event lidar data are typically not available for most events, making it challenging to attempt change detection. Thus, nearly all studies involving damage detection using lidar data involve only the post-earthquake data.

A significant advantage offered by SAR over optical EO and lidar is that SAR data can be obtained even in poor-light conditions, including at night; can penetrate through ground-level obstacles; and is independent of cloud cover. Heretofore, most of the SAR-based methods described in the comprehensive review undertaken by attempt to infer changes and predict damage at the block level rather than at the building level. One reason for this has been the limited availability of very-high-resolution SAR data. However, meter-level spatial resolution is now offered by several SAR satellites, including ALOS-2, COSMO-SkyMed, TerraSAR-X, and TanDEM-X, which makes building-level damage detection promising, especially when ancillary datasets such as digitized building footprint layers are available for use in conjunction with high-resolution SAR data (e.g., ).

Meanwhile, recent advances in machine learning (ML), with high-performance open-source libraries for training and evaluating models, have led to an increasing body of work that aims to use learning algorithms to predict damage following earthquakes. Broadly, these efforts can be classified into the following four categories: (i) ML models using building attributes and geophysical features alone (e.g., ), (ii) ML models using optical EO data (e.g., ; see , for an extensive review), (iii) ML models using SAR data alone (e.g., ), and (iv) ML models using SAR EO data in conjunction with building attributes and geophysical features (e.g., ).

evaluated the performance of various ML classification algorithms to classify building damage for the 2017 Puebla, Mexico, earthquake, based on input features including structural attributes of the buildings and seismic demand in terms of maximum spectral acceleration. The random forest model was found to have the highest relative accuracy amongst the models evaluated, being able to correctly identify 78 % of the damaged buildings in the test set. also evaluated the performance of different ML classification algorithms to classify building damage from the 2014 South Napa earthquake and indicated that the random forest algorithm provides the best performance amongst the evaluated techniques. The input features included distance from the fault, spectral acceleration, and structural attributes of the buildings. However, even the random forest algorithm was correctly able to identify only 12.5 % of the red-tagged buildings, though it was able to correctly classify 79 % of the yellow-tagged buildings. provided a comprehensive summary of the state of the art on earthquake building damage detection using deep-learning methods, mostly based on convolutional neural networks (CNNs), with optical imagery obtained by remote-sensing satellite or airborne sensors.

The use of machine-learning algorithms in conjunction with SAR data for earthquake damage detection is a recent development, and relatively fewer studies are found in the literature compared to optical-EO-data-based studies that employ machine learning. evaluated the application of a support vector machine (SVM) classifier to identify changes in single- and multi-temporal X- and L-band SAR images from TerraSAR-X and ALOS PALSAR, and they used their classifier to detect damage from the 2011 Tohoku earthquake and tsunami. While the single-image approach that uses only the post-image SAR data yielded reasonable results, the multi-temporal approach demonstrated a greater performance. The authors report that the SVM classifier performs well for binary classification, but performance degrades when trying to classify damage into multiple grades. in their research tried to shed some light on the difference in the building damage mapping performance when using multi-temporal or only post-event SAR images in the framework of machine learning. The k-nearest-neighbor learning algorithm was selected as the preferred classifier, as it showed the best performance in evaluation. Using the 2016 Kumamoto earthquake as a case study, the authors indicate a prediction accuracy of 40.1 % for damaged buildings when only post-event SAR images were used and a 38.9 % accuracy for identifying damaged buildings when multi-temporal SAR images were used.

proposed a new approach for the classification of collapsed buildings in the aftermath of a disaster based on SAR imagery, the spatial distribution of hazards, and a set of fragility functions. The method was applied to the 2011 Tohoku earthquake and tsunami for collapsed building detection using two TerraSAR-X images (one pre-event image and one post-event image) of the coastal area of Miyagi Prefecture. Their binary classification into collapsed and non-collapsed buildings was compared with field survey damage data. Depending on the fragility function used, the method was able to correctly classify between 80.4 %–92.7 % of the buildings that were completely washed away by the tsunami, whereas 61.2 %–69.8 % of the buildings categorized as collapsed were actually washed away.

In this study, we build upon the aforementioned work and propose a supervised ML framework for building-level earthquake damage assessment using high-resolution SAR data, which we combine with high-resolution building inventory datasets and earthquake ground shaking intensity maps to classify buildings into different damage states. We attempt both binary damage classification as well as multi-class damage classification for four recent earthquakes using ensemble models in a machine-learning approach. Multi-class damage grade classification using SAR EO data has rarely been attempted, as evidenced by the previously presented literature. Thus, the case studies presented in this report represent one of the first such attempts using SAR data. Comparing the predicted damage states with ground truth data obtained through field surveys allows us to assess the accuracy of the ML model for the different case study earthquakes.

The key elements of the framework include the SAR-derived damage proxy map (DPM) for the event, a ground shaking intensity map such as the ShakeMap published by the US Geological Survey (USGS), and a building inventory layer which includes at minimum the building footprints in the affected area. Where available, additional building attributes can also be incorporated into the framework. Finally, ground truth damage grade labels for the buildings in the affected area are needed to train and test the machine-learning model for damage classification. The proposed approach is designed to be compatible with both detailed exposure information (building-by-building data) or proxy exposure information mapped on to a building footprint layer (for regions where building-specific data are not available), although its prediction accuracy is enhanced when building-specific data are used. Whereas the illustrative application case studies in this paper focus on implementing and testing this framework for earthquake-related damage, the proposed framework adopts a modular approach, so that it can be extended to other natural hazards by tailoring the hazard-specific parts whilst using the core building blocks that would be common to all hazards. The various input datasets used by the framework and the processing steps are described in this section.

We evaluate the best-fit model on both the training and test subsets using a few different performance metrics, including the precision and recall scores, the F1 score, and the balanced accuracy score . These metrics are described in brief below.

Precision. This is defined as the ratio (true positives) / (true positives + false positives). The precision is intuitively the ability of the classifier not to label as positive a sample that is negative.

The results for the four case study earthquakes, for both multi-class classification and binary classification of damage, are summarized below in Tables  through . When we attempt a multi-class damage grade classification, the trained model exhibits a balanced accuracy score ranging from 0.23 for the 2017 Puebla earthquake, to 0.36 for the 2015 Gorkha earthquake and 0.40 for both the 2020 Puerto Rico and Zagreb earthquakes. The balanced accuracy scores improve significantly to 0.65 for the 2017 Puebla and 2020 Zagreb earthquakes, to 0.72 for the 2020 Puerto Rico earthquake, and to 0.82 for the 2015 Gorkha earthquake when we switch to a binary damage grade classification, i.e., attempting only to separate the damaged buildings from the undamaged buildings.

Figures  and show the confusion matrices for the four earthquakes for multi-class damage grade classification and binary damage classification, respectively. Table  provides an overall summary of the balanced accuracy scores for all of the cases considered. The recall score for the “damaged” label in the binary damage classification scenario ranges from 0.38 for the 2017 Puebla earthquake and 0.48 for the 2020 Zagreb earthquake to 0.58 for the 2020 Puerto Rico earthquake and 0.73 for the 2015 Gorkha earthquake; i.e., for three out of the four earthquakes studied, the model is able to identify over half or nearly half of the damaged buildings successfully when using binary classification.

We observe that for the 2015 Gorkha earthquake, for which multiple building attributes are available for both damaged and undamaged buildings, the prediction accuracies for both binary and multi-class classification are significantly higher when compared to the earthquakes where fewer or no building attributes are available for use as input features. In order to understand the impact of including the building attributes on the performance of the classifier, we also trained the ML model for this earthquake without using any of the building attributes and limiting the input feature vector to the MMI and DPM values alone. The precision and recall for all damage grades are lower for this reduced model compared to the results reported for the full model in Tables  and for both multi-class classification and binary classification, respectively. The recall score for the “destruction” damage grade drops from 0.47 for the full model to 0.31 for the reduced model in the multi-class classification task and from 0.73 to 0.45 in the binary classification task. Similarly, the balanced accuracy score drops from 0.36 to 0.20 in the multi-class classification task and from 0.82 to 0.59 in the binary classification task. These results clearly demonstrate the importance of including the additional building attributes in the analysis. A partial dependence analysis of the damage grade on the building attribute variables for this event that are not related to geographic location indicates that the building age has an impact on the damage grade, with older buildings being related to higher damage (see the Jupyter Notebook for the analysis).

While four key building attributes were also available for the damaged buildings for the 2017 Puebla earthquake, the non-availability of the same for the undamaged buildings meant that a complete dataset with building attributes could not be used for the training of the ML model. Of the four events considered in this study, the 2017 Puebla event had the smallest building damage dataset available for training the ML model. Only 219 buildings in Mexico City had a complete set of building attributes and damage labels and were also covered by the DPM and ShakeMap layers, as compared to thousands or hundreds of thousands of buildings for the other three events. While the random forest classification model performs well for this event in the training phase, the trained model fails to correctly identify even a single partially collapsed or totally collapsed building in the test set. Further attempts at reducing the potential overfitting of the model to the limited training data subset by adjustments to the model hyperparameters did not lead to any noticeable improvements in prediction accuracy for the event.

From Fig. , we also observe that the true-positive prediction rates for the intermediate damage grades are lower than those for the no-damage and highest damage grades for the 2015 Gorkha earthquake and the 2020 Puerto Rico and Zagreb earthquakes. We believe that this partly stems from the fact that the existing damage scales that are widely used for field surveys of building damage, such as EMS-98 , do not map directly to information available through Earth-observation data, particularly for the lower damage grades. For instance, the first three damage grades for reinforced concrete structures according to EMS-98 involve increasing levels of cracking in the beams and columns or partition and infill walls, as well as buckling of the reinforcement rods. Unless this kind of damage results in debris caused by excessive spalling of the concrete cover or partial collapse of infill walls that is visible outside the structure, these damage levels as defined in EMS-98 may be challenging to identify from EO data alone. and both propose a building damage assessment scale tailored for optical satellite imagery and aerial imagery. However, a similar damage scale tailored for InSAR-based building damage assessment is still lacking and merits further research.

Even for the higher damage grades, a potential limitation of SAR-derived damage assessment at the building level is that it is certainly possible to have significant seismic building damage without observing a significant corresponding change in the ground surface level. While a fully collapsed building is expected to cause higher coherence loss compared to partial collapse, field surveys of damage are likely to mark all such buildings as completely damaged or destroyed. Damage to internal walls or columns that may have severely compromised the structural integrity of the building without causing externally visible damage or collapse may not be detectable through remote sensing, as these will be out of the line of sight of the satellites. Story drifts of 2 % for braced steel structures or concrete shear wall structures may be technically classified as “collapsed” (e.g., ) even though the structure has not physically collapsed at all, but such drift levels might be smaller than the level detectable with the 1–3 m spatial resolution offered by the current generation of SAR sensors. With the present resolution of the DPM being around 30 m × 30 m per pixel, the proposed method is more likely to be useful for detecting large damaged buildings, damaged building aggregates, and damage to dense building blocks than for detecting damage to isolated smaller buildings. With the advent of commercial SAR satellite constellations like Capella Space and ICEYE, sub-meter SAR imagery is becoming available. Digital elevation maps (DEMs), which are used in terrain correction or removal of topographically induced phase variations in the generation of the DPMs, are also now becoming available at meter and sub-meter resolutions. Thus, some of the aforementioned deficiencies should be addressable.

This article describes a framework for semi-automated damage assessment due to earthquake from Earth-observation (EO) data and other supplementary datasets, leveraging upon recent advances in machine-learning algorithms. This framework combines high-resolution building inventory data from OpenStreetMap and other local sources with image-processing algorithms for the detection of earthquake damage using InSAR data generated by the JPL-ARIA initiative, along with supplementary geospatial datasets as inputs to a random forest ML classification model. The ML model is trained using detailed building damage datasets from past events in a supervised learning framework. Both multi-class and binary damage classification are attempted, and we compare the predicted damage labels with ground truth damage grade labels reported in field surveys. Binary damage classification is shown to outperform multi-class classification for the earthquakes studied, and the highest classification accuracy scores are observed for the case where the largest number of building attributes relevant for structural damage are available.

The time span between the acquisition of the pre-event and post-event images can have a considerable impact on the potential false positives. The closer the “before” and “after” bracket the event, the fewer the false positives that are likely to be observed. Multiple SAR satellite missions currently have revisit intervals of a few days. DPMs are already reliably generated by the ARIA team within a few days of major earthquakes. With the planned launches of the NASA-ISRO SAR (NISAR) mission and ALOS-4 and the advent of commercial SAR satellite constellations , post-event image acquisition is expected to become available within 1–2 d after observation or even within a few hours in response to disasters . Earthquake ground shaking intensity maps are also made available by the US Geological Survey within a few hours after an event. The curation of building inventory datasets is a critical step that should ideally be undertaken before the occurrence of a disaster event, and such datasets should be regularly updated. In the absence of a precompiled building inventory dataset for the affected region, building extracts from OSM can be used within the proposed damage detection framework. The training of the machine-learning models would have been undertaken prior to the disaster event, and the trained model can then be deployed for damage detection following an earthquake as soon as the pre-event building inventory, ShakeMap, and DPM become available. Thus, the time frame for obtaining the first results from the proposed damage detection framework is expected to be in the order of 1–7 d following an earthquake.

Machine-learning models for the prediction of post-disaster damage can benefit greatly from having access to labeled and georeferenced building damage data. Cross-regional training datasets will also help greatly improve the performance of these models for earthquakes in new regions previously unseen by the model. By expanding the datasets used to train the ML damage classification models, we can transfer the learning from regions with more damage data availability to data-sparse regions. Cross-regional training is also critical as it will ultimately make such damage classification models more robust as they can be more confidently applied to future disasters, which may affect regions the model has not been trained on. This remains a challenge, however, particularly in the case of multi-class classification of damage grades, due to differences in the damage scales used in the field damage surveys in different events and also because of the subjectivity involved in the assignment of damage grades by the field damage surveyors.

One of the eventual promises of the framework described in this paper is to be able to predict damage using InSAR data even for locations that are not present in the training data. Ideally, region-specific damage detection models could be developed that take into consideration input features that are region-specific. Alternatively, region-specific or location-specific characteristics could be encoded as additional input features to a global remote-sensing-based damage detection model. For instance, one of the inputs in the proposed methodology is the ground shaking intensity map (ShakeMap) generated by the US Geological Survey, which does take into consideration local site conditions, albeit through the proxy measure of Vs30 values. The tectonic setting is also taken into account implicitly in the derivation of the ShakeMap, as the choice of the ground motion model used to predict the ground shaking intensities in the affected area depends on the tectonic region type. If information about building construction types is available, this can be encoded as a categorical input feature, as was done for the 2015 Gorkha and 2017 Puebla examples in this study.

Both SAR and optical EO data have their relative strengths. SAR data can be obtained even in poor-light conditions, at night, and independent of cloud cover. However, while meter-level spatial resolution is now offered by several SAR satellites, optical EO data are currently available at sub-meter spatial resolution from several satellites, making it particularly appealing for building-level damage detection methods. Finer differentiation of damage grades involving detection of cracks in walls or residual drifts is still quite challenging with the 1–3 m spatial resolution offered by the current generation of SAR sensors. Thus, simultaneous use of SAR and optical EO data in a deep-learning workflow could potentially combine the advantages of the two different data types and increase the accuracy of damage detection. The addition of supplementary information, such as hazard intensity data measured at a few locations, local site conditions, and building attribute data, could also help improve rapid post-earthquake damage classification.