Earthquake building damage classification based on full suite of Sentinel-1 features

Xiaoran Lv; Guichun Luo; Lifu Zheng; Bozhi Zhang; Chen Zhang

doi:10.1515/geo-2025-0882

Article Open Access

Earthquake building damage classification based on full suite of Sentinel-1 features

Xiaoran Lv , Guichun Luo , Lifu Zheng , Bozhi Zhang and Chen Zhang

Published/Copyright: September 9, 2025

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Open Geosciences Volume 17 Issue 1

Abstract

This study investigates the capabilities of the proposed approach that fully utilizes the intensity, phase, and polarimetry features of Sentinel-1 for multi-class and binary classification of building damage, aiming to address the limitations of existing studies, which typically only utilize partial Sentinel-1 features or integrate them with an ancillary dataset for building damage classification. Using the Puerto Rico earthquake and Turkey earthquake as case studies, multi-class classification achieved F1 score of 0.50/0.58, an overall accuracy of 0.52/0.60, and receiver operating characteristic – area under the ROC curve (AUC) value of 0.59/0.69, respectively. For binary classification under varying proportions of collapsed buildings and no damaged buildings, overall accuracy ranged from 0.631 to 0.813 and F1 score from 0.630 to 0.806 across the two earthquakes. The performance for both classification tasks is comparable to previous studies that employed other synthetic aperture radar data or integrated one or two Sentinel-1 features with ancillary dataset, and the performance is better than that of studies using only one or two Sentinel-1 features for the destroyed vs undamaged classification. These quantitative comparisons highlight that our proposed approach, which solely rely on Sentinel-1, exploiting all features of Sentinel-1 data and using a readily applied machine learning model, is suitable for the timely, rapid, and microregion-level building damage detection in the local region after the earthquake.

Keywords: earthquake; Sentinel-1; building damage; multi-class classification

1 Introduction

Earthquakes are among the most catastrophic and unpredictable disasters that cause serious economic losses and casualties [1]. Rapid relief and accurate assessment of economic losses after earthquakes are highly important, especially for urban areas [2]. The degree and spatial distribution of building damage are crucial information for rescue operations and damage assessment.

On-site surveys are a typical method for assessing the spatial distribution and damage level of building damage. However, conducting comprehensive field surveys across entire affected areas immediately after an earthquake is challenging. This difficulty conflicts with the critical golden 48-h window [3] for rescuing individuals trapped in the collapsed buildings. Furthermore, disrupted transportation and communication systems impede the on-site survey. In addition, field investigators face significant danger from aftershocks or secondary disasters.

Remote sensing (RS) technology can provide data with wide spatial coverage and non-contact observation capabilities [4]. It also enables the acquisition of data across various electromagnetic spectrum, such as optical, synthetic aperture radar (SAR) data. As a result, RS serves as an effective and efficient data source for building damage classification following earthquakes.

RS-based approaches for building damage classification can be divided into two categories [5,6,7]: change detection and post-earthquake data-only detection. Change detection methods utilize both the pre- and post-earthquake data, while the post-earthquake data-only detection relies exclusively on post-earthquake data.

Post-earthquake data-only detection, which utilizes only post-earthquake RS data, encompasses relatively fewer approaches [6]. Based on the electromagnetic spectrum used, these methods can be categorized into three types: those utilizing high-resolution optical data, those utilizing SAR data, and those fusing optical with SAR data. For optical data-based methods, they can be further categorized into visual interpretation [8,9], texture-based algorithms [4,10], and algorithms integrating spectral and textural features [11,12]. For SAR data-based methods, classifications include polarimetry feature-based algorithms [13,14,15,16,17,18], texture feature-based algorithms [19,20,21], and algorithms combining multiple features [22,23].

Change detection, utilizing both pre- and post-earthquake data, encompasses significantly more diverse approaches than post-earthquake data-only detection. Although numerous, these methods can be categorized into four types. The first one is optical data-based algorithms, which can be further classified into visual interpretation (which is laborious but widely used) [24,25], image enhancement techniques (e.g. change highlighting via image subtraction) [26,27,28], post-classification comparison (often adopting deep learning) [29,30,31], and other methods, such as using building roofs [32] or 3D geometric changes [33] as indicators for damage extraction. Additionally, LiDAR [34,35], a vertical optical data, is also employed for damaged building classification. The secondary SAR data-based algorithms, which can be further classified into approaches utilizing intensity information [36,37], phase information [38,39], polarimetry information [40,41], and integrated methods employing part or all of above information types [42,43,44]. The third category, integrating optical and SAR data, [45,46], combines the complementary advantages of optical and SAR data. The fourth category includes methods fusing the RS data with ancillary datasets, such as peak ground acceleration (PGA), building inventories [46,47].

Compared with post-earthquake data-only detection, change detection adopts pre-earthquake data, providing significantly richer information to support building damage classification and is consequently more widely used. Moreover, while optical data offer ease of interpretation and typically feature sub-meter resolution, they are passive remote sensing data and thus highly dependent on favorable weather conditions (e.g. no clouds, no rain). Conversely, SAR data can be acquired regardless of nighttime or harsh weather conditions [48]. Consequently, SAR data is considered more flexible for building damage classification, particularly during the critical early phase following the earthquake.

As described above, SAR-based change detection methods use intensity, phase, and polarimetry features. For intensity-based change detection, common approaches involve calculating intensity differences, correlation coefficients, or texture disparities between pre- and post-earthquake data to assess building damage. For instance, Matsuoka et al. [9,49] applied this method to the 1995 Kobe earthquake using C-band and L-band SAR data, respectively. Cui et al. [50] utilized intensity data to quantify damaged building distributions in the 2010 Yushu earthquake. Kim et al. [37] employed novel textural features to detect damaged buildings in the 2016 Kumamoto earthquake. For phase-based change detection, the core principle involves calculating the difference or ratio between pre-seismic coherence and the co-seismic coherence. Ito et al. [38,51] applied the ratio of pre- and co-seismic coherence to detect damaged buildings following the 1995 Kobe and 1999 Kocaeli earthquakes. Sharma et al. [52] employed the difference between pre- and co-seismic coherence for building damage classification after the 2015 Nepal earthquake. To mitigate interference from factors such as spatial baseline variations and background noise, researchers have utilized histogram-matched pre-seismic coherence time series alongside co-seismic coherence. Liu et al. [39] utilized coherence time series to identify damaged buildings after the 2022 Afghanistan earthquake. Yang et al. [53] combined coherence time series with an artificial neural network (ANN) algorithm to detect damaged buildings following the 2023 Turkey earthquake. For polarimetry-based change detection, methods can be categorized into two groups according to polarization mode. In the dual-polarization mode, only one primary polarization decomposition technique, H-alpha decomposition, is typically applied. Consequently, studies often employ intensity, textures, or coherence across different polarizations. For example, Karimzadeh and Mastuoka [40] used dual-polarized SAR intensity combined with coherence data to generate the binary damage map for the 2016 Amatrice earthquake. Oxoli et al. [44] adopted coherence under different polarizations to assess the building damage in the 2016 Central Italy earthquake. In the full-polarization mode, the technique provides richer information for analyzing building backscattering mechanisms. Studies focus on developing polarimetry decomposition algorithms and examining the relationship between polarimetry indices (e.g. average scattering angle, volume scattering) with building damage states (intact vs damaged), as these indices are sensitive to the physical properties, geometry structures, and orientation of buildings. For instance, Yamaguchi [54] used three polarimetry indices to construct the pre- and co-earthquake RGB composites, demonstrating that the spectral difference could identify damaged buildings. Chen et al. [41] calculated the ratio of dominant double-bounce scattering to the standard deviation of polarization orientation angle differences to map damaged buildings in the 2011 Tohoku earthquake.

However, acquiring full-polarization SAR data in a timely manner remains challenging, particularly for undeveloped regions, as most such datasets are neither freely available nor inexpensive. Consequently, numerous building damage classification studies continue to focus on dual-polarization SAR data, combining intensity, coherence, and/or polarimetry features. Sentinel-1, a twin-satellite constellation launched by the European Space Agency (ESA), provides freely accessible C-band dual-polarization SAR data. Its revisit time are 12 days for a single satellite and 6 days for the constellation. Therefore, despite its medium spatial resolution (approximately 10m for GRD products and 5 m × 20 m for the SLC products), Sentinel-1 enables the acquisition of timely and cost-free data, especially for the pre-earthquake data. This underscores the crucial importance of fully exploring the achievable accuracy and diverse application scenarios of Sentinel-1 SAR data for building damage classification. For example, Akhmadiya et al. [42] employed texture features from two polarimetric modes and phase information of Sentinel-1 to classify damaged vs intact buildings using a maximum likelihood algorithm for the 2016 Italy earthquake. Aimaiti et al. [45] utilized Sentinel-1 intensity data and Sentinel-2 texture features to perform binary classification via threshold analysis in the Ukraine region. Sandhini Putri et al. [46] integrated Sentinel-1 intensity and texture features with Sentinel-2 spectral data to classify building damage into medium, severe, and collapse categories using a random forest (RF) algorithm for the 2018 Lombok earthquake. Liu et al. [55] applied Sentinel-1 phase data for binary classification in the 2023 Turkey earthquake. Rao et al. [47] combined Sentinel-1 phase data, building inventories, and PGA to conduct binary and multi-class classification via RF algorithm across five earthquake-affected regions. Liu et al. [39] used Sentinel-1 phase data to distinguish damaged from intact buildings in the 2022 Afghanistan earthquake, and Yang et al. [53] employed an ANN algorithm with Sentinel-1 phase data for binary classification in the 2023 Turkey earthquake. Du et al. [56] used surface deformation derived from the Sentinel-1 SLC data to calculate the direction of earthquake parameter for detecting damaged buildings in the 2023 Turkey earthquake.

However, the existing studies (Table 1) primarily utilize only one or two features derived from Sentinel-1 data or fusing some of the features with ancillary data. Furthermore, most focus on assessing Sentinel-1’s performance for binary building damage classification. Relatively few studies have deeply explored the potential of building damage classification solely relying on Sentinel-1 data, particularly for multi-class classification, by exploiting its intensity, phase, and polarimetry features.

Table 1

Previous building damage classification studies based on Sentinel-1

Methods	Data used	Classification type	References
Decision-tree	Sentinel-1 polarimetry	Multi-class	[57]
Discriminant analysis	Sentinel-1 phase and LiDAR		[58]
Random Forest	Sentinel-1 phase, Alos-2 phase, Sentinel-2		[59]
Random Forest	Sentinel-1 intensity, polarimetry and Sentinel-2		[46]
Discriminant analysis	Sentinel-1 intensity and phase	Binary class	[40]
RGB Visualization	Sentinel-1 phase and polarimetry		[60]
Decision-tree	Sentinel-1 polarimetry		[57]
Multivariate alteration detection (MAD) and naive Bayes (NB)	Sentinel-1 phase and intensity		[61]
Discriminant analysis	Sentinel-1 phase and intensity		[62]
Maximum Likelihood and Mahalanobis	Sentinel-1 polarimetry and phase		[42]
Discriminant analysis	Sentinel-1 intensity and Sentinel-2		[45]
Discriminant analysis	Sentinel-1 phase		[39]
Recurrent Neural Network (RNN)	Sentinel-1 phase		[53]

Therefore, this study aims to explore the classification capabilities of integrating all suite of Sentinel-1 features (intensity, polarimetry, and phase) for multi-class and binary classification of building damage. This approach seeks to address the limitations of existing studies in unleashing the potential of Sentinel-1 data, which typically only combines one or two Sentinel-1 features or integrates them with ancillary data (e.g., optical data, PGA, or high spatial resolution data) to perform building damage classification. Secondly, this study constructs distinct binary classification scenarios to explore the optimal binary classification grouping approach for building damage severity levels tailored to Sentinel-1 data. Finally, this study further investigates whether the pre-trained model for destroyed vs undamaged classification generated by the proposed approach supports the cross-region classification. Two earthquake-affected regions with detailed on-site multi-class damaged buildings classifications data were selected as study areas, and the RF algorithms were employed for the analysis.

2 Methods

2.1 Study area

2.1.1 The 2020 Puerto Rico earthquake

The Mw 6.4 Puerto Rico earthquake occurred on 7 January 2020. The epicenter was located at the southwest coast, and the focal mechanism indicates that the earthquake ruptured a normal fault. The economic loss and casualties caused by the earthquake are comparable to those of the 1918 San Fermin Mw7.1 earthquake [47]. Field investigations [63] were carried out by the Federal Emergency Management Agency (FEMA) from 12 January to 15 January 2020. The survey indicated that about 335 buildings were damaged, including 77 completely destroyed buildings. The damaged buildings were mainly distributed in five provinces, namely Guanica, Yauco, Guayanilla, Penuelas, and Ponce. Therefore, buildings in these regions were considered for the multi-class classification in this study (Figure 1a).

Figure 1

Study area and the building damage field survey results. (a), study area and field survey results for the Puerto Rico earthquake, and (b), study area and field survey results for the Turkey earthquake. The red polygons in (a) indicate the administration boundaries of the Guanica, Yauco, Guayanilla, Penuelas, and Ponce provinces.

2.1.2 The 2023 Turkey earthquake

The Mw 7.8 and the Mw 7.5 strike-slip earthquakes occurred on 6 February 2023 in Turkey. They caused serious casualties and economic losses, affecting about 44,000 people and resulting in a loss of about $34 billion. More than 160,000 buildings, including 520,000 apartments, collapsed partially or entirely [53]. These earthquakes almost destroyed the whole Kahramanmaras city, which was one of the worst-hit areas in Turkey. Thus, we selected Kahramanmaras city for research in this study (Figure 1b).

2.2 Data and the preprocessing

2.2.1 SAR intensity and the gray-level co-occurrence matrix (GLCM)

We used the Sentinel-1 ground range detected at high resolution (GRDH) of ascending data, which contain the dual-polarized (VV and VH) information in this study (Table 2). The intensity and the GLCM metrics (Table 3) were calculated using SNAP software [64] based on the pre-event and post-event images. In total, we obtained four intensity images and forty GLCM index images. Then, the log ratio [45] of intensity before and after the event was calculated as the final intensity index. Finally, 42 features were obtained.

I ratio = 10 log 10 I pre − event I post − event .

Table 2

Sentinel-1 ascending data used in this study

Data	Date	Earthquake
Sentinel-1 IW GRDH	20200104, 20200116	Puerto Rico
Sentinel-1 IW GRDH	20230128, 20230209	Turkey
Sentinel-1 IW SLC	20190109-20200116 (30)	Puerto Rico
Sentinel-1 IW SLC	20220202-20230221 (32)	Turkey

Table 3

Features used for damage level classification

Feature	Note
GLCM Mean	Texture feature (VV/VH)
GLCM Correlation
GLCM Variance
Contrast
Dissimilarity
Energy
Entropy
Angular Second Moment (ASM)
Homogeneity
Maximum Probability (MAX)
Intensity log ratio	Intensity feature (VV/VH)
Coherence differences	Coherence feature

2.2.2 Coherence differences calculation

Coherence is a byproduct of interferogram generation [65], which represents the similarity of two repeat-pass SAR signals. Higher coherence indicates smaller changes between two images, and the precise definition of the coherence is given by,

γ = 〈 s 1 s 2 * 〉〈 s 1 s 1 * 〉〈 s 2 s 2 * 〉 ,

where, the s 1 and s 2 are the complex signal of SAR, * means the conjugate, and the 〈〉 denotes the ensemble average.

We used a 1-year pre-earthquake [39] and one closest post-earthquake Sentinel-1 ascending SLC IW images (Table 2) to calculate the coherence differences. The coherence stack was calculated using the JPL/Caltech ISCE stack processor [66,67] and MintPy [68]. We used the histogram matching [39,69] and one standard deviation [39,70] value to remove the spatial baselines' influence and the background effects, respectively. Coherence differences were then calculated by subtracting coseismic coherence from the closest preseismic coherence. Finally, the coherence differences were resampled to the 10 m × 10 m pixel spacing as the GRDH data. The standard deviation [39,71] of the 1-year pre-earthquake coherence stack was not used to mask out unstable regions because this index could not satisfactorily distinguish between the unstable (such as vegetation) and stable (such as buildings) regions for the medium-solution Sentinel-1 data.

2.2.3 Field survey data

After the Puerto Rico earthquake, the FEMA Hazus program [63] carried out the field survey from 12 January 2020 to 15 January 2020. They identified a total of 775 buildings with three levels of damage, namely destroyed, major, and minor (Table 4). However, they did not provide a clear definition of each damage type. In this study, the destroyed type was combined with the major type as one category termed “serious damage”, the minor type was renamed as “slight damage”, and the no damage type was labeled as “undamage.”

Table 4

Statistics characteristics of damaged buildings

Damage type	Number
Puerto Rico earthquake
Destroyed	41
Major	178
Minor	396
Turkey earthquake
Collapse	583
Need to be demolished	372
Heavily damage	2,480
Slightly damage	9,823

After the Turkey earthquake, Yazilim and Cizenler [72] conducted a field survey and investigated a total of 200,000 buildings. Four types of building damage were established, namely collapsed, need to be demolished, heavily damaged and slightly damaged. According to their data, damaged buildings in Kahramanmaras city were counted and shown in Table 4. Yazilim defined the “collapsed buildings” as buildings in ruins, “need to be demolished” buildings as buildings that are partially collapsed, “heavily damaged” buildings as buildings experiencing concrete construction and structure destroyed, and “slightly damaged” buildings as buildings that were generally repairable. In addition, in order to compare our multi-class classification results with other studies, we adopt Yu’s (59) and Yang’s (53) method and grouped the collapsed buildings, need to be demolished buildings, and heavily damaged buildings as “serious damage” [53,59]. The other two types are the “slight damage” and “undamage,” respectively.

2.2.4 Building footprint data

For the Puerto Rico earthquake, we utilized building footprint data obtained from the OpenStreetMap (OSM), and 72,416 buildings were identified in the study area. For the Turkey earthquake, we used building footprint data extracted from Xu’s et al. [73] work, which contains more building vectors than the OSM. They obtained building footprints based on the Sentinel-2 Normalized Difference Built-up Index. Then, the building vector of Kahramanmaras was extracted (36,676 buildings).

The building vector data were used as the mask to extract the corresponding coherence differences, intensity ratio, and GLCM metrics. For instance, with a multi-grid in one building footprint, the average value was calculated as the final property. Finally, 43 features (Table 3) were obtained for each building vector.

2.3 Methods

In this study, a supervised machine learning algorithm was employed to evaluate the potential of the medium-resolution Sentinel-1 SAR data in the prediction of building damage types. Considering that the dataset has a medium size and the dataset is actually in tabular form with each row representing a build unit, the RF algorithm [47,59,74] was applied for the classification.

Hyperparameters [75,76], which directly affect the generalization and complexity of the model, cannot be learned during the training phase. Thus, the hyperparameter tuning process must be set prior to the training process. Probst et al. [76] provided a detailed overview of hyperparameter tuning strategies for RF and pointed out that the main hyperparameters of RF are the number of trees in the forest, the node size, the number of features sampled when looking for the best split for a node, and the samples used by each tree. In this study, the optimal values of these hyperparameters were determined using the Bayes method [77] offered in the Python software package hyperopt, and the RF classifier provided by the Python software package scikit-learn was used.

A unique feature of the damaged building dataset is the heavy imbalance [47,59], which implies that the vast majority of buildings are undamaged. This imbalance also exists in the dataset of different damage types. The imbalance tends to encourage the trained model to classify most buildings in the test dataset into the undamaged category. Therefore, we also applied the imblearn module from scikit-learn to randomly undersample the majority, such as the “undamaged.” Then, we randomly selected 80% of the dataset as the training dataset and used the remaining 20% as the testing dataset.

We selected the precision, recall, F1 score, overall accuracy, and ROC (receiver operating characteristic)-AUC (area under the ROC curve) metrics as the evaluation metrics to determine the optimum classification model. Additionally, a modified F1 score (79–81), which is the summed average of the three categories’ F1 scores, was calculated as follows:

F 1 d = 3 ∑ i = 0 2 1 F 1 i .

Among them, F 1 d , overall accuracy and ROC-AUC were utilized as the main reference index [59] because they consider both the model sensitivity and specificity and are widely used in classification models. Moreover, we conducted 100 additional experiments to quantitatively evaluate the effectiveness of our method combining the phase, intensity, and polarimetry information of Sentinel-1. In each experiment, the trained model was used to classify all buildings into three categories, and the mode of 100 classification results for each building footprint was used as the final category. Then, the evaluation metrics were calculated.

3 Results

3.1 Multi-class classification

3.1.1 Puerto Rico earthquake

The ROC–AUC histogram of the 100 experiments is shown in Figure 2a–d. The peak ROC–AUC was 0.58, 0.59, and 0.66 for the “serious damage,” “slight damage,” and “undamage” categories, respectively. The peak value of the average ROC-AUC was 0.59. Based on the truth data, the confusion matrix and evaluation metrics are shown in Figure 3a and Table 5, respectively. For the three categories, the recall was 0.64, 0.51, and 0.43, and the F1 score was 0.44, 0.52, and 0.53, respectively. The overall accuracy and F 1 d was 0.52 and 0.50, respectively.

Figure 2

Model performance through the ROC-AUC index for the two earthquakes. We conducted 100 multi-class classification experiments to quantitatively evaluate the effectiveness of our method. The ROC-AUC histogram of each damage type was calculated. Then, the histogram of the average ROC–AUC value for all damage types is calculated. (a)–(d) The histograms for the Puerto Rico earthquake, and (e)–(h) the histograms for the Turkey earthquake. For the Puerto Rico earthquake, the peak ROC–AUC was 0.58, 0.59, and 0.66 for the serious damage, slight damage, and undamaged categories, respectively. The peak value of the average ROC–AUC was 0.59. For the Turkey earthquake, the peak ROC-AUC was 0.69, 0.67, and 0.69 for the serious damage, slight damage, and undamaged categories, respectively. The peak value of the average ROC–AUC was 0.69.

Figure 3

Confusion matrix of multi-class classification for the two earthquakes. We conducted 100 multi-class classification experiments. And we used the mode of 100 classification results for each building footprint as the final category. (a) Confusion matrix for the Puerto Rico earthquake, and (b) confusion matrix for the Turkey earthquake.

Table 5

Performance metrics of the multi-class classification for Puerto Rico earthquake

	Precision	Recall	F1 score	F 1 d	Overall accuracy
Serious damage	0.34	0.64	0.44	0.50	0.52
Slight damage	0.52	0.51	0.52
Undamage	0.70	0.43	0.53

3.1.2 Turkey earthquake

The ROC–AUC histogram of the 100 experiments and evaluation metrics is shown in Figure 2e–h. The peak ROC-AUC was 0.69, 0.67, and 0.69 for the “serious damage,” “slight damage” and “undamage” categories, respectively. The peak value of the average ROC-AUC was 0.69. The confusion matrix and evaluation metrics for the three categories are shown in Figure 3b and Table 6, respectively. The recall was 0.89, 0.73, and 0.51, and the F1 score was 0.50, 0.59, and 0.65, respectively. The overall accuracy and F 1 d was 0.60 and 0.58, respectively.

Table 6

Performance metrics of the multi-class classification for Turkey earthquake

	Precision	Recall	F1 score	F 1 d	Overall accuracy
Serious damage	0.34	0.89	0.50	0.58	0.60
Slight damage	0.49	0.73	0.59
Undamage	0.90	0.51	0.65

3.2 Binary classification

In addition to multi-class classification, we also evaluated our approach for binary classification. Four scenarios for Puerto Rico and six for the Turkey earthquake were established (Table 7), each involving different proportions of collapsed/destroyed and no damaged buildings.

Table 7

Scenarios for binary classification

Scenario	Definition of “damage” type	Definition of “undamage” type
Puerto Rico earthquake
ScenarioP1	Destroyed and major damaged buildings	No damaged buildings
ScenarioP2	Destroyed buildings	No damaged buildings
ScenarioP3	Destroyed and major damaged buildings	Minor and no damaged buildings
ScenarioP4	Destroyed buildings	Minor and no damaged buildings
Turkey earthquake
ScenarioT1	Collapsed, need to be demolished, and heavily damaged buildings	No damaged buildings
ScenarioT2	Collapsed, need to be demolished buildings
ScenarioT3	Collapsed buildings
ScenarioT4	Collapsed, need to be demolished, and heavily damaged buildings	Slightly and no damaged buildings
ScenarioT5	Collapsed, need to be demolished buildings
ScenarioT6	Collapsed buildings

The performance metrics for binary classification under various scenarios in the Puerto Rico and Turkey earthquakes are detailed in Table 8. Our approach achieved superior performance in binary classification compared to multi-class classification. The overall accuracy ranged from 0.631 to 0.813, and the F1 score from 0.630 to 0.806 across the two earthquakes. For both earthquakes, a higher proportion of collapsed/destroyed buildings generally corresponded to better performance, given the “undamage” building type.

Table 8

Performance metrics of binary classification

	ROC–AUC	Precision	Recall	F1 score	Overall accuracy
Puerto Rico earthquake
ScenarioP1	0.725	0.648	0.648	0.630	0.631
ScenarioP2	0.873	0.688	0.718	0.676	0.687
ScenarioP3	0.722	0.707	0.708	0.695	0.696
ScenarioP4	0.891	0.812	0.864	0.806	0.813
Turkey earthquake
ScenarioT1	0.744	0.665	0.661	0.658	0.659
ScenarioT2	0.816	0.717	0.716	0.717	0.718
ScenarioT3	0.854	0.787	0.782	0.772	0.771
ScenarioT4	0.701	0.663	0.657	0.653	0.655
ScenarioT5	0.802	0.764	0.761	0.761	0.762
ScenarioT6	0.814	0.767	0.759	0.754	0.755

4 Discussion

4.1 Comparison with other studies

4.1.1 Multi-class classification performance comparison

To evaluate our approach’s performance, we compared performance metrics with existing studies (Table 9, Figure 4a), focusing on F 1 d score and overall accuracy as they are widely adopted as benchmarks. Optical data-based multi-class classification generally outperforms SAR or LiDAR data-based approaches, attributable to the high spatial resolution (often sub-meter) of the optical images, which enables the individual building-level analysis. In contrast, spaceborne SAR resolution is typically limited to meter-scale (e.g., Sentinel-1’s 10 m resolution). Given this resolution constraint, we first compared with studies performing multi-class classification at the street-block unit level. Our results show lower metrics than street-block unit level [22,81–84] (overall accuracy: 0.73–0.85), primarily due to their statistical methodology: assuming building structure homogeneity within street-blocks, they computed the ratio of damaged building pixel number to total pixel number and applied thresholds to assign multi-level damage classifications. Conversely, when compared to pixel-block unit level SAR analyses [39,47,53,57–59,85], our model’s performance is comparable. The previous studies report an F 1 d score ranging from 0.23 to 0.53, overall accuracy from 0.52 to 0.65, and the ROC-AUC of approximately 0.69. In our study, the corresponding metrics for the two earthquakes are 0.50/0.58 ( F 1 d score), 0.52/0.60 (overall accuracy), and 0.59/0.69 (ROC-AUC), respectively. These results demonstrate that our approach achieves a precision level comparable to these methods. Additionally, this study exhibits performance similar to or better than studies [35,86] that employ LiDAR dataset.

Table 9

Performance metrics for multi-class classification in previous studies

Data	Method	Classes	Precision	References
Street block-unit
Airborne PolSAR	Discriminant Analysis	Slight, Median, Serious	Overall accuracy: 0.79, Kappa coefficient: 0.68	[81]
RadarSat-2	Discriminant Analysis	Slight, Moderate, Serious	Overall accuracy: 0.73	[82]
Radarsat-2/ALOS-1	Discriminant Analysis	Slight, Moderate, Serious	Overall accuracy: 0.73/0.83	[83]
Airborne SAR	Discriminant Analysis	Slight, Moderate, Serious	Overall accuracy: 0.85	[22]
Radarsat-2	Discriminant threshold	Slight, Moderate destroy	F 1 d : 0.72	[84]
pixel-block unit
Sentinel-1 polarimetry	Decision-tree classifier	No damage, Minor, Major destroyed	Overall accuracy: 0.52	[57]
Sentinel-1 and LiDAR	Change detection	No damage, Minor, Major destroyed	overall accuracy: 0.63	[58]
ENVISAT	PCA change detection	No damage, Moderate destroyed	Overall accuracy: 0.65	[85]
AIAR DMP, PGA and Building inventors	Random forest	No damage, Slight, Moderate, Substantial, Very heavy destruction	F 1 d : 0.34, Precision: 0.33, Recall: 0.36	[47]
		No damage, Slight, Intermediate, Heavy, Partial collapse, Total collapse	F 1 d : 0.53, Precision: 0.51, Recall: 0.56
		No damage, Slight, Moderate, Heavy	F 1 d : 0.23, Precision: 25, Recall: 0.40
		No damage, Slight moderate, Heavy	F 1 d : 0.47, Precision: 0.50, Recall: 0.72
Sentinel-1, Alos-2, and Sentinel-2	Random forest Classifier	No damage, Slight, Serious	ROC-AUC: 0.68	[59]
Sentinel-1 and Sentinel-2	Random forest	Moderate, Severe destory	Overall accuracy: 0.62	[46]
Individual building level
CASMAR	Random forest	Slight, Moderate, Serious	Overall accuracy: 0.82	[87]
xBD	Deep learning, Encoder-Decoder backbones	No damage, Minor damage, Major damage	F 1 d : 0.68–0.78	[88]
xBD	Deep learning, Cross-directional fusion model	No damage, Minor damage, Major damage	F 1 d : 0.77	[89]
xBD	Deep learning	No damage, Minor, Major destroyed	F 1 d : 0.73	[78]
Optical images	Deep learning	No damage, Minor, Major destroyed	Precision: 0.80, Accuracy: 0.86, F 1 d : 0.83	[90]
xBD	Deep learning bitemporal attention trans former module	No damage, Minor, Major destroyed	F 1 d : 0.78	[91]
xBD	Deep learning: transformers	No damage, Minor, Major destroyed	F 1 d : 0.79, IoU: 0.87	[92]
xBD	Deep learning	No damage, Major destroyed	F 1 d : 0.86	[93]
Optical images	Deep learning	No damage, Light, Heavy collapse	Overall accuracy: 0.64	[94]
xBD	Deep learning	No damage, Minor, Major destroyed	Pixel-based F 1 d : 0.71, Object-based F 1 d : 0.75	[79]
Optical images	Deep learning: CNN	No damage, Minor, Moderate, Severe destroyed	Accuracy: 0.80	[95]
xBD	Deep learning: Transformer learning	No damage, Minor, Major destroyed	F 1 d : 0.76, MioU: 0.62	[80]
xBD	Deep learning: ResNet	No damage, Minor, Major destroyed	F 1 d : 0.37	[96]
xBD	Deep learning: semi-Siamese network	No damage, Minor, Major destroyed	F 1 d : 0.61	[97]
xBD	U-Net	No damage, Minor, Major destroyed	F 1 d : 0.75	[98]
LiDAR	Deep learning: FCSNN	Slight, Moderate, Substantial to heavy, Very heavy destruction	Accuracy: 0.42, F 1 d : 0.41, Precision: 0.39, Recall: 0.47	[35]
LiDAR	Mechanic learning	No damage, Moderately, Highly destroyed	Overall accuracy: 0.60	[86]

$Figure 4 Performance metrics comparison between this study and other studies (a) for multi-class classification and (b) for binary classification. For both multi-class and binary classification tasks, while the performance metrics of our work is lower compared to those using optical data, our results exhibit close similarity to those of numerous studies relying on SAR data. For multi-class classification, when compared to pixel-block unit-level SAR studies, our model’s performance is comparable. The previous studies report an F 1 d {F1}_{d} score ranging from 0.23 to 0.53, overall accuracy from 0.52 to 0.65, and the ROC–AUC of approximately 0.69. In our study, the corresponding metrics for the two earthquakes are 0.50/0.58 ( F 1 d {F1}_{d} score), 0.52/0.60 (overall accuracy), and 0.59/0.69 (ROC-AUC), respectively. For the binary classification, the overall accuracy in these studies ranges from 0.58 to 0.86, with Kim’s study reporting an F1 score of 0.80. In our research, the overall accuracy for these two earthquakes under different scenarios ranges from 0.631 to 0.813, and the F1 score ranges from 0.630 to 0.806.$

Figure 4

Performance metrics comparison between this study and other studies (a) for multi-class classification and (b) for binary classification. For both multi-class and binary classification tasks, while the performance metrics of our work is lower compared to those using optical data, our results exhibit close similarity to those of numerous studies relying on SAR data. For multi-class classification, when compared to pixel-block unit-level SAR studies, our model’s performance is comparable. The previous studies report an F 1 d score ranging from 0.23 to 0.53, overall accuracy from 0.52 to 0.65, and the ROC–AUC of approximately 0.69. In our study, the corresponding metrics for the two earthquakes are 0.50/0.58 ( F 1 d score), 0.52/0.60 (overall accuracy), and 0.59/0.69 (ROC-AUC), respectively. For the binary classification, the overall accuracy in these studies ranges from 0.58 to 0.86, with Kim’s study reporting an F1 score of 0.80. In our research, the overall accuracy for these two earthquakes under different scenarios ranges from 0.631 to 0.813, and the F1 score ranges from 0.630 to 0.806.

Comparison results with other studies, adopting diverse datasets and operating at different analytical levels, indicates that while making full use of the intensity, phase, and polarimetry features of medium-resolution Sentinel-1 data does not achieve the highest precision in the building damage multi-class classification compared to optical images, its performance is nearly equivalent to researches using other commercial SAR data or fusing Sentinel-1 with ancillary datasets. This suggests that following an earthquake, freely available large-area Sentinel-1 SAR data can be used effectively and efficiently to conduct a coarse-grained multi-class classification of building damage. This approach is particularly valuable for bad-weather conditions or undeveloped regions, where it serves as a critical first-response tool.

4.1.2 Binary classification performance comparison

This study further compares the binary classification performance metrics with those of other studies (Table 10, Figure 4b). Analogous to the multi-class classification results, studies employing optical data generally exhibit a higher F1 score or overall accuracy compared to those utilizing SAR data. The optical data-based studies report an F1 score ranging from 0.69 to 0.90 and overall accuracy from 0.84 to 0.90. When comparing our binary classification outcomes with prior studies using SAR data, our study demonstrates comparable performance. The overall accuracy in these studies ranges from 0.58 to 0.86, with Kim et al.’s study [37] reporting an F1 score of 0.80. In our research, the overall accuracy for these two earthquakes under different scenarios ranges from 0.631 to 0.813, and the F1 score ranges from 0.630 to 0.806. For destroyed vs undamaged classification, the overall accuracy of our approach is higher than that of [42,45,62] using only one or two Sentinel-1 features. And for scenario T1, our AUC (0.744) is comparable to that of Yang et al.’s [53] result (0.761) using RNN to dig phase information. Finally, our F1 score is remarkably close to that of Kim’s study, and our overall accuracy falls within the range reported by other studies.

Table 10

Performance metrics for binary classification in previous studies

Data	Method	Building damage type	Precision	References
ALOS-2	SVM	(1) Damaged: Destroyed, severely damaged	Overall accuracy: 0.80, Kappa: 0.61	[99]
ALOS-2	SVM	(2) Undamaged: No damage	Overall accuracy: 0.80, Kappa: 0.61	[99]
Sentinel-1/Alos-2	Discriminant analysis	(1) Damaged: Destruction	Accuracy: 0.84/0.76	[40]
Sentinel-1/Alos-2	Discriminant analysis	(2) Undamaged: Slight damage, Moderate damage, Heavy damage, Very heavy damage	Accuracy: 0.84/0.76	[40]
TerraSAR-X	Random forest	(1) Damaged: Destruction	Overall accuracy: 0.72	[100]
TerraSAR-X	Random forest	2) Undamaged: Slight damage, Moderate damage, Extensive damage	Overall accuracy: 0.72	[100]
TerraSAR-X	Discriminant analysis	(1) Damaged: Collapse	Overall accuracy: 0.74	[101]
TerraSAR-X	Discriminant analysis	(2) Undamaged: Negligible damage, Moderate damage, Severe damage	Overall accuracy: 0.74	[101]
Capella SAR	Machine learning	(1) Damaged: Collapse	Overall accuracy: 0.72	[19]
Capella SAR	Machine learning	(2) Undamaged: No damage, Minor damage, Moderate damage	Overall accuracy: 0.72	[19]
Kompsat-5	Discriminant analysis	(1) Damaged	F1: 0.80	[37]
Kompsat-5	Discriminant analysis	(2) Undamaged, No damage	F1: 0.80	[37]
Sentinel-1	Earthquake damage visualization (EDV) technique	(1) Damaged: Severely damaged, Collapse	Overall accuracy: 0.85	[60]
Sentinel-1	Earthquake damage visualization (EDV) technique	(2) Undamaged, No damage	Overall accuracy: 0.85	[60]
Sentinel-1	Decision-tree classifier	(1) Damaged: Very heavy damage destruction	Overall accuracy: 0.71	[57]
Sentinel-1	Decision-tree classifier	(2) Undamaged: Negligible to slight damage, Moderate damage, Substantial to heavy damage	Overall accuracy: 0.71	[57]
Sentinel-1	Multivariate alteration detection (MAD) and naive Bayes (NB)	(1) Damaged: Damaged destroyed	Overall accuracy: 0.67	[61]
Sentinel-1		(2) Undamaged: Intact, Slightly damaged	Overall accuracy: 0.67	[61]
Sentinel-1	Discriminant analysis	(1) Damaged	Overall accuracy: 0.68	[62]
Sentinel-1	Discriminant analysis	(2) Undamaged	Overall accuracy: 0.68	[62]
ALOS2	Random forest	(1) Damaged: Destruction	Overall accuracy: 0.86	[102]
ALOS2	Random forest	(2) Undamaged: Negligible to slight damage, Moderate damage, Substantial to heavy damage, Very heavy damage	Overall accuracy: 0.86	[102]
PALSAR-2	Discriminant analysis	(1) Damaged	Kappa: 0.69	[103]
PALSAR-2	Discriminant analysis	(2) Undamaged	Kappa: 0.69	[103]
Sentinel-1	Discriminant analysis	(1) Damaged: Destroyed	Overall accuracy: 0.77	[42]
Sentinel-1	Discriminant analysis	(2) Undamaged: No damage	Overall accuracy: 0.77	[42]
Sentinel-1 and Sentinel-2	Discriminant analysis	(1) Damaged: Destroyed	Accuracy: 0.58	[45]
Sentinel-1 and Sentinel-2	Discriminant analysis	(2) Undamaged	Accuracy: 0.58	[45]
Sentinel-1	Coherence time series	(1) Damaged	CI: 0.8	[39]
Sentinel-1	Coherence time series	(2) Undamaged	CI: 0.8	[39]
Sentinel-1	RNN	(1) Damaged: Need to be demolished, Heavily damaged, Collapsed	AUC: 0.76	[53]
Sentinel-1	RNN	(2) Undamaged: Slightly damaged, No damaged	AUC: 0.76	[53]
xBD	Machine learning	(1) Damaged: Severe damage destroyed	AUC: 0.83	[104]
xBD	Machine learning	(2) Undamaged: No damage	AUC: 0.83	[104]
Optical and SAR	Mechanical learning	(1) Damaged	Overall accuracy: 0.9	[105]
Optical and SAR	Mechanical learning	(2) Undamaged	Overall accuracy: 0.9	[105]
xBD	UNET	(1) Damaged: Destroyed	F1: 0.83, Overall accuracy: 0.84	[106]
xBD	UNET	(2) Undamaged: No damage	F1: 0.83, Overall accuracy: 0.84	[106]
xBD	CNN	(1) Damaged: Destroyed, Major damaged	F1: 0.71	[107]
xBD	CNN	(2) Undamaged: No damage, Minor damage	F1: 0.71	[107]
xBD	Deep learning	(1) Damaged: Destroyed	Overall accuracy: 0.95, F1: 0.90	[108]
xBD	Deep learning	(2) Undamaged: No damage	Overall accuracy: 0.95, F1: 0.90	[108]
GF1 and Google Earth images	CNN	(1) Damaged: Destroyed	F1: 0.69, Overall accuracy: 0.84, Precision: 0.58, IoU: 0.522	[109]
GF1 and Google Earth images	CNN	(2) Undamaged: No damage		[109]
xBD	SVM	(1) Damaged: Destroyed, Major damaged	Overall accuracy: 0.97	[110]
xBD	SVM	(2) Undamaged: No damage, Minor damaged	Overall accuracy: 0.97	[110]
LiDAR	SVM	(1) Damaged: Debris	Overall accuracy: 0.91	[34]
LiDAR	SVM	(2) Undamaged: Intact	Overall accuracy: 0.91	[34]
LiDAR	FCSNN	(1) Damaged: Collapsed	Accuracy: 0.87, F1: 0.79, Precision: 0.78, Recall: 0.80	[35]
		(2) Undamaged: Intact	Accuracy: 0.87, F1: 0.79, Precision: 0.78, Recall: 0.80
		(1) Damaged: Collapse	Accuracy: 0.95, F1: 0.54, Precision: 0.70, Recall: 0.43
		(2) Undamaged: Non-collapsed	Accuracy: 0.95, F1: 0.54, Precision: 0.70, Recall: 0.43

The comparison analysis demonstrates that fully utilizing Sentinel-1’s intensity, phase, and polarimetry features achieves performance comparable to other research studies, reaffirming that freely available Sentinel-1 SAR data constitute a valuable resource for post-earthquake building damage classification, particularly for the collapsed buildings.

4.2 Performance constraints in building damage classification

Building upon the aforementioned analysis, two key conclusions can be drawn. First, while the performance metrics of our work is lower compared to those using optical data, our results exhibit close similarity to those of numerous studies relying on SAR data. Second, the performance of multi-class classification shows a relatively more pronounced decline when contrasted with that of binary classification.

The significant disparity in spatial resolution between optical and SAR data represents a primary factor contributing to the poorer performance of SAR data in building damage classification. Sub-meter optical data enables individual building-level damage detection [108–110], as it not only accentuates the geometric, textural, and spectral features of each building, but also provides sufficient pixels within each building footprint for detailed analysis. In contrast, the medium spatial resolution (10 m) of Sentinel-1 data often results in mixed pixels rather than pure building-only pixels [46,47,53]. Consequently, in this pixel-oriented [67] building damage classification study, substantial variations in damage levels among buildings or the variations in surface objects within a single pixel hinder the RF classifier’s ability to effectively learn the characteristics of each damage level, leading to relatively lower classification metrics.

The dataset employed for RF classifier was produced using building footprint data sourced from prior studies and OSM. As illustrated in Figure 5a–c, a subset of these footprints exhibited positional inaccuracies, encompassing regions with mixed surface objects rather than strictly corresponding to the building. This spatial misalignment was exacerbated by the medium resolution of Sentinel-1 data, which predominantly yielded mixed pixels containing heterogeneous surface objects rather than pure building-only pixels [46]. Moreover, the field survey data provided by FEMA Hazus [63] and Yazilim and Cizenler [72] are point vectors, with some points showing inconsistency with the building footprint data or even deviating from actual building positions (Figure 5d–f). This inconsistency introduces labeling errors where damaged buildings may be misclassified as undamaged and vice versa, thereby degrading the quality of the dataset.

Figure 5

Spatial inconsistency between the building footprint, field survey points, and actual building position. (a)–(c) The spatial inconsistency between the building footprint and actual building position, and (d)–(f) the spatial inconsistency between the field survey points and actual building position. The black polygons are the building footprints, and the points are the field survey locations.

Like most studies [19,39,47,53,59], field survey data serves as the reference for constructing the datasets used by the RF classifier. The damage levels assigned to each survey point/polygon are also used as the labels for the machine learning datasets. These field surveys typically employ the widely used EMS-98 (European Macroseismic Scale 1998) damage scales [111], which ranges from “negligible to slight damage” to “destruction”. Dell’Acqua and Gamba [112] and Cotrufo et al. [86] proposed building damage scales adapted from EMS-98 specifically for high-resolution optical data. However, a comparable building damage scale tailored to SAR data remains lacking [47]. Consequently, similar to previous studies, we adopt the building damage scale used in the field survey data. This approach introduces potential errors because the field survey scales are not well-aligned with the information derivable from SAR data. For example, buildings labeled “to be demolished” in Yazilim’s field survey [72] might correspond to buildings that remain standing despite the first floor collapse. Identifying such buildings using SAR intensity or texture features can be challenging. Furthermore, buildings experiencing severe damage to the internal walls or columns without visible external damage or collapse may be classified as “serious damage” in a field survey. These buildings, however, are highly likely to be classified as “undamage” in SAR images. Thus, this discrepancy also contributes to the relatively poor performance observed in the multi-class classification. It also explained why the higher percentages of collapsed buildings in the dataset (Table 7) correlate with improved performance metrics (Table 8) in binary classification.

4.3 Limitations and outlook

Our focus is on evaluating the feasibility and accuracy of the medium-resolution Sentinel-1 SAR data for building damage classification, utilizing its full suite of features (intensity, phase, and polarimetry). Although the performance metrics of this study do not match the performance of optical data-based approaches, our results demonstrate accuracy comparable to numerous studies fusing Sentinel-1 with other data sources. These indicate that freely available Sentinel-1 data can provide an effective and efficient means for timely building damage classification, particularly in undeveloped regions or under conditions where optical data are unavailable.

With the continuous development of building damage classification methods, scholars expect that pre-trained models constructed in a certain region can be directly applied to other regions. The key to achieving this goal lies in continuously enriching the damaged building sample library. xBD [113] is the first building damage classification dataset to date, and the dataset samples are produced using high spatial resolution optical imagery. Since in high spatial resolution optical imagery, different damage levels, especially collapsed buildings, have basically the same features, a pre-trained model based on this dataset can be well applied to new disaster areas. However, there is currently no building damage classification dataset based on SAR imagery. In this study, we attempted to extract collapsed buildings in Turkey using a pre-trained model (collapse vs no damage) obtained from the Puerto Rico dataset and vice versa, but the detection results were not good (Table 11). Then, for collapsed buildings in Turkey and Puerto Rico, three combinations of randomly selected features used by RF classifier were compared (Figure 6). The figures illustrate the distinct feature distribution of collapsed buildings between the two regions for these selected features. The primary cause is attributed to mixed pixels resulting from Sentinel-1's medium spatial resolution, with secondary factors including disparities in surface characteristics (e.g., building density, vegetation density) across the two seismic areas. Therefore, our method makes it relatively difficult to provide a highly robust pre-trained model for other regions. Our approach is more suitable for using Sentinel-1 data with high spatial coverage and machine learning algorithms to complete the extraction of damaged buildings in the entire earthquake area after obtaining early-stage field survey data, providing initial information for emergency rescue and other work, and providing a reference for subsequent fine-scale assessment of building damage using high spatial resolution optical imagery or commercial SAR satellites. It should also be noted that the proposed method is better suited for detecting extensive collapsed buildings, and damage occurred in dense building regions rather than individual smaller buildings.

Table 11

Performance metrics for cross-region classification

Object	Pre-trained
Object	Puerto Rico	Turkey
Turkey	F1 score: 0.43
Turkey	Overall accuracy: 0.53
Puerto Rico		F1 score: 0.33
Puerto Rico		Overall accuracy: 0.51

Figure 6

Distribution of three randomly selected features: (a) and (b) for post-earthquake features and (c) for pre-earthquake features. The figures illustrate the distinct feature distribution of collapsed buildings between the two regions for these selected features.

In this study, we only used the ascending data in the study areas. Different line-of-sight direction observation can obtain additional features of buildings. In the future, descending data can be combined to investigate how the performance of our approach can be improved. Moreover, our approach employs pixel-oriented analysis [59], which not only reduces computational time to meet the timeliness requirements of the emergency response but also imposes no strict hardware demands on computing devices. In the future, the image-oriented machine learning method, such as U-net, can replace the RF classifier to explore whether the performance could be improved.

5 Conclusion

This study presents a new approach solely relying on Sentinel-1 data, exploiting all features of it and using a readily applied machine learning model to develop a rapid, microregion-level building damage classification. Taking the earthquake-affected regions of Puerto Rico and Turkey as examples, this study investigates the capabilities of the proposed approach that fully utilizes the intensity, phase, and polarimetry features of Sentinel-1 for multi-class and binary classification of building damage. Thereby aims to address the limitations of existing studies, which typically only utilize partial Sentinel-1 features or integrate them with an ancillary dataset for building damage classification. The key conclusions are as follows:

For multi-class classification, the proposed approach achieved F 1 d score of 0.50 and 0.58, overall accuracy of 0.52 and 0.60, and ROC-AUC of 0.59 and 0.69 for the two earthquakes, respectively. When comparing with previous studies that employed other SAR data or integrated one or two Sentinel-1 features with an ancillary dataset, our approach achieves a precision level comparable to these methods ( F 1 d score ranging from 0.23 to 0.53, overall accuracy from 0.52 to 0.65, and the ROC–AUC of approximately 0.69).
For binary classification, we evaluated different scenarios characterized by different proportions of collapsed buildings and no damage buildings, respectively. The overall accuracy and F1 score for the two earthquakes across these scenarios ranged from 0.631 to 0.813 and 0.630 to 0.806, respectively. These metrics are also comparable to the previous methods (overall accuracy ranging from 0.58 to 0.86), and for the destroyed vs undamaged classification, the performance of our approach is higher than that of studies using only one or two Sentinel-1 features. Moreover, the binary classification performance under different scenarios indicated that our approach can achieve the highest metrics for destroyed vs undamaged classification.
The result of transferring a pre-trained model across the two earthquake-affected regions is suboptimal. This indicates that medium spatial resolution data is more suitable for the timely building damage detection in the local region after the earthquake. Additionally, high spatial resolution SAR data are necessary in the future to construct the microwave-“xBD” database.

Acknowledgments

We would like to thank the anonymous reviewers for their helpful suggestions, which improved the quality of this manuscript. We also would like to thank the ESA for collecting the Copernicus Sentinel-1 data (https://scihub.copernicus.eu/) and to Yanchen Yang for the assistance in the Turkey validation data collection.

Funding information: This research was funded by the Sciences and Technology Funds of Beijing Earthquake Agency, grant number BJMS-2025005, and the Scientific Research Funds of the Institute of Engineering Mechanics, China Earthquake Administration, grant numbers 2020EEEVL305 and 2022QJGJ04.
Author contributions: Xiaoran Lv developed the model code, processed the data, and wrote the original draft. Guichun Luo did the supervision. Lifu Zheng prepared the manuscript with contributions from all co-authors.
Conflict of interest: The authors declare no conflicts of interest.

References

[1] Li J, Huang X, Hu Y, Zhang Z, Zhang X, Zhang S, et al. Fusion of optical daily and night-time light remote sensing images for collapsed building detection: A case in Turkey Mw 7.8 earthquake. Geomat Inf Sci Wuhan Univ. 2023;48(10):1706–14.Search in Google Scholar

[2] An LQ, Zhang JF, Ricardo M, Zhang L. A review and prospective research of earthquake damage assessment and remote sensing. Natl Remote Sens Bull. 2024;28(4):860–84.Search in Google Scholar

[3] Karimzadeh S, Miyajima M, Hassanzadeh R, Amiraslanzadeh R, Kamel B. A GIS-based seismic hazard, building vulnerability and human loss assessment for the earthquake scenario in Tabriz. Soil Dyn Earthq Eng. 2014;66:263–80.10.1016/j.soildyn.2014.06.026Search in Google Scholar

[4] Yamazaki M, Matsuoka M. Remote sensing technologies in post-disaster damage assessment. J Earthq Tsunami. 2007;1(3):193–210.10.1142/S1793431107000122Search in Google Scholar

[5] Dong L, Shan J. A comprehensive review of earthquake-induced building damage detection with remote sensing techniques. ISPRS J Photogramm Remote Sens. 2013;84:85–99.10.1016/j.isprsjprs.2013.06.011Search in Google Scholar

[6] Ge P, Gokon H, Meguro K. A review on synthetic aperture radar-based building damage assessment in disasters. Remote Sens Environ. 2020;240:111693.10.1016/j.rse.2020.111693Search in Google Scholar

[7] Xu Y. A comprehensive review of CNN-based methods for earthquake-induced building damage detection using remote sensing imagery. Comput Fraud Secur. 2025;663–79.10.52710/cfs.385Search in Google Scholar

[8] Ogawa N. Photo-interpretation of building damage due to earthquakes using aerial photographs. Proceedings of the 12th World Conference on Earthquake Engineering (12 WCEE), Auckland, New Zealand, 2000.Search in Google Scholar

[9] Kouchi K, Matsuoka M, Kohiyama M, Muraoka N. Damage detection from high-resolution satellite images for the 2003 Boumerdes, Algeria earthquake. Asia Conference on Earthquake Engineering, Manila, Philippines, 2004.Search in Google Scholar

[10] Ishii M, Goto T, Sugiyama T, Saji H, Abe K. Detection of earthquake damaged areas from aerial photographs by using color and edge information. The 5th Asian Conference on Computer Visio, Melbourne, Australia, 2002.Search in Google Scholar

[11] Abdi G, Esfandiari M, Jabari S, eds. Building damage detection in post-event high-resolution imagery using deep transfer learning. 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS; 2021.10.1109/IGARSS47720.2021.9553720Search in Google Scholar

[12] Li X, Yang W, Ao T, Li H, Chen W. An improved approach of information extraction for earthquake-damaged buildings using high-resolution imagery. J Earthq Tsunami. 2011;5(4):389–99.10.1142/S1793431111001157Search in Google Scholar

[13] Li X, Guo H, Zhang L, Chen X, Liang L. A new approach to collapsed building extraction using RADARSAT-2 polarimetric SAR imagery. IEEE Geosci Remote Sens Lett. 2012;9(4):677–81.10.1109/LGRS.2011.2178392Search in Google Scholar

[14] Zhai W, Huang C. Fast building damage mapping using a single post-earthquake PolSAR image: a case study of the 2010 Yushu earthquake. Earth Planets Space. 2016;68:86.10.1186/s40623-016-0469-2Search in Google Scholar

[15] Shan LI, Fengli ZH, Shiying WE, Na LI, Yun SH. SAR image visual optimization and building damage assessment based on polarimetric decomposition combination. J Univ Chin Acad Sci. 2020;37(6):750–9.Search in Google Scholar

[16] Zhai W, Jingfa Z, Xiulai X, Jinghan W, Haoran Z, Xinxin Y, et al. Damaged building extraction from post-earthquake PolSAR data based on the Fourier transform. Remote Sens Lett. 2021;12(6):594–603.10.1080/2150704X.2021.1906975Search in Google Scholar

[17] Chen Q, Nie Y, Li L, Liu X. Buildings damage assessment using texture features of polarization decomposition components. J Remote Sens. 2017;21(6):955–65.10.11834/jrs.20176414Search in Google Scholar

[18] Du Y, Gong L, Li Q, Zhan S, Zhang J. Earthquake induced building damage assessment on SAR multi-texture feature fusion. Remote Sens Technol Appl. 2021;36(4):865–72.Search in Google Scholar

[19] Macchiarulo V, Giardina G, Milillo P, Aktas Y, Whitworth M. Integrating post-event very high resolution SAR imagery and machine learning for building-level earthquake damage assessment. Bull Earthq Eng. 2024;1–27. 10.1007/s10518-024-01877-1.Search in Google Scholar

[20] Zhai W, Shen HF, Huang CL. Collapsed buildings extraction from the PolSAR image based on the analysis of texture feature. Remote Sens Technol Appl. 2016;31(5):975–82.Search in Google Scholar

[21] Gokon H, Meguro K. Building damage assessment using intensity SAR data with different incidence angles and longtime interval. J Disaster Res. 2019;14(3):456–65.Search in Google Scholar

[22] Zhai W, Huang C, Pei W. Building damage assessment based on the fusion of multiple texture features using a single post-earthquake PolSAR image. Remote Sens. 2019;11(8):897.10.3390/rs11080897Search in Google Scholar

[23] Bai Y, Gao C, Singh S, Koch M, Adriano B, Mas E, et al. A framework of rapid regional tsunami damage recognition from post-event TerraSAR-X imagery using deep neural networks. IEEE Geosci Remote Sens Lett. 2018;15(1):43–7.10.1109/LGRS.2017.2772349Search in Google Scholar

[24] Adams B, Mansouri B, Huyck C. Streamlining post-earthquake data collection and damage assessment for the 2003 Bam, Iran, earthquake using VIEWS™ (visualizing impacts of earthquakes with satellites). Earthq Spectra. 2005;21(1_suppl):213–8.10.1193/1.2098588Search in Google Scholar

[25] Yano Y, Matsuoka M. Visual damage interpretation of buildings in Bam city using QuickBird images following the 2003 Bam, Iran, Earthquake. Earthq Spectra. 2005;21(S1):S329–S336.10.1193/1.2101807Search in Google Scholar

[26] Yusuf Y, Matsuoka M. Damage assessment after 2001 Gujarat earthquake using Landsat7 satellite images. J Indian Soc Remote Sens. 2001;29:17–22.10.1007/BF02989909Search in Google Scholar

[27] Rathje E, Crawford M, Woo K, Neuenschwander A. Damage patterns from satellite images of the 2003 Bam, Iran, earthquake. Earthq Spectra. 2005;21(1_suppl):295–307.10.1193/1.2101047Search in Google Scholar

[28] Miura H, Midorikawa S, Soh H. Texture analysis of high-resolution satellite images for damage detection in the 2010 Haiti earthquake. J Jpn Assoc Earthq Eng. 2012;12(6):2–20.10.5610/jaee.12.6_2Search in Google Scholar

[29] Chini M, Cinti F, Stramondo S. Co-seismic surface effects from very high resolution panchromatic images: The case of the 2005 Kashmir (Pakistan) earthquake. Nat Hazards Earth Syst Sci. 2011;11:931–43.10.5194/nhess-11-931-2011Search in Google Scholar

[30] Xie Z, Zhou Z, He X, Fu Y, Gu J, Zhang J. Methodology for object-level change detection in post-earthquake building damage assessment based on remote sensing images: OCD-BDA. Remote Sens. 2024;16(22):4263.10.3390/rs16224263Search in Google Scholar

[31] Qiao W, Li S, Qi W, Quan W, Shiyang T, Li Z. Revolutionizing building damage detection: A novel weakly supervised approach using high-resolution remote sensing images. Int J Digital Earth. 2024;17(1):2298245.10.1080/17538947.2023.2298245Search in Google Scholar

[32] Chesnel AL, Binet R, Wald L, eds. Object oriented assessment of damage due to natural disaster using very high resolution images. 2007 IEEE International Geoscience and Remote Sensing Symposium; 2007.10.1109/IGARSS.2007.4423655Search in Google Scholar

[33] Tong X, Hong Z, Liu S, Zhang X, Xie H, Li Z, et al. Building-damage detection using pre- and post-seismic high-resolution satellite stereo imagery: A case study of the May 2008 Wenchuan earthquake. ISPRS J Photogramm Remote Sens. 2012;68:13–27.10.1016/j.isprsjprs.2011.12.004Search in Google Scholar

[34] Rastiveis H, Eslamizade F, Hosseini-Zirdoo E. Building damage assessment after earthquake using post-event lidar data. Int Arch Photogramm Remote Sens Spat Inf Sci. 2015;40:595–600.10.5194/isprsarchives-XL-1-W5-595-2015Search in Google Scholar

[35] Ramadhan MML, Jati G, Jatmiko W. Building damage assessment using feature concatenated Siamese neural network. IEEE Access. 2024;12:19100–16.10.1109/ACCESS.2024.3361287Search in Google Scholar

[36] Matsuoka M. Use of SAR imagery for monitoring areas damaged due to the 2006 Mid Java, Indonesia earthquake. 4th International Workshop on Remote Sensing for Post-Disaster Response, Cambridge, UK, 2006.Search in Google Scholar

[37] Kim M, Park S-E, Lee S-J. Detection of damaged buildings using temporal SAR data with different observation modes. Remote Sens. 2023;15:308.10.3390/rs15020308Search in Google Scholar

[38] Ito Y, Hosokawa M, Matsuoka M, eds. A degree estimation model of earthquake damage using temporal coherence ratio. IGARSS 2003 2003 IEEE International Geoscience and Remote Sensing Symposium Proceedings (IEEE Cat No03CH37477); 2003.Search in Google Scholar

[39] Liu H, Song C, Li Z, Liu Z, Ta L, Zhang X, et al. A new method for the identification of earthquake-damaged buildings using Sentinel-1 multitemporal coherence optimized by homogeneous SAR pixels and histogram matching. IEEE J Sel Top Appl Earth Obs Remote Sens. 2024;17:7124–43.10.1109/JSTARS.2024.3377218Search in Google Scholar

[40] Karimzadeh S, Mastuoka M. Building damage assessment using multisensor dual-polarized synthetic aperture radar data for the 2016 M 6.2 Amatrice Earthquake, Italy. Remote Sens. 2017;9(4):330.10.3390/rs9040330Search in Google Scholar

[41] Chen SW, Wang XS, Sato M. Urban damage level mapping based on scattering mechanism investigation using fully polarimetric SAR data for the 3.11 East Japan earthquake. IEEE Trans Geosci Remote Sens. 2016;54(12):6919–29.10.1109/TGRS.2016.2588325Search in Google Scholar

[42] Akhmadiya A, Nabiyev N, Moldamurat K, Dyussekeyev K, Atanov S. Use of Sentinel-1 dual polarization multi-temporal data with gray level co-occurrence matrix textural parameters for building damage assessment. Pattern Recognit Image Anal. 2021;31(2):240–50.10.1134/S1054661821020036Search in Google Scholar

[43] Xuedong Z, Wenxi L, Shuguang H. Urban change detection in TerraSAR image using the difference method and SAR coherence coefficient. J Eng Sci Technol Rev. 2018;11:18–23.10.25103/jestr.113.03Search in Google Scholar

[44] Oxoli D, Boccardo P, Brovelli MA, Molinari ME, Guarnieri AM. Coherent change detecting for repeated-pass interferometric SAR images: an application to earthquake damage assessment on buildings. In The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Istanbul, Turkey: 2018. p. 383–8.10.5194/isprs-archives-XLII-3-W4-383-2018Search in Google Scholar

[45] Aimaiti Y, Sanon C, Koch M, Baise LG, Moaveni B. War related building damage assessment in Kyiv, Ukraine, Using Sentinel-1 Radar and Sentinel-2 optical images. Remote Sens. 2022;14(24):6239.10.3390/rs14246239Search in Google Scholar

[46] Sandhini Putri AF, Widyatmanti W, Umarhadi DA. Sentinel-1 and Sentinel-2 data fusion to distinguish building damage level of the 2018 Lombok Earthquake. Remote Sens Appl: Soc Environ. 2022;26:100724.10.1016/j.rsase.2022.100724Search in Google Scholar

[47] Rao A, Jung J, Silva V, Molinario G, Yun SH. Earthquake building damage detection based on synthetic-aperture-radar imagery and machine learning. Nat Hazards Earth Syst Sci. 2023;23(2):789–807.10.5194/nhess-23-789-2023Search in Google Scholar

[48] Wang C, Zhang H, Wu F, Zhang B, Tang Y, Wu H, et al. Disaster phenomena of Wenchuan earthquake in high resolution airborne synthetic aperture radar images. J Appl Remote Sens. 2009;3(1):031690.10.1117/1.3154558Search in Google Scholar

[49] Matsuoka M, Nojima N. Building damage estimation by integration of seismic intensity information and satellite L-band SAR imagery. Remote Sens. 2010;2(9):2111–26.10.3390/rs2092111Search in Google Scholar

[50] Cui L, Wang X, Dou A, Ding X, High Resolution SAR. Imaging employing geometric features for extracting seismic damage of buildings. ISPRS - Int Arch Photogramm Remote Sens Spat Inf Sci. 2018;42:239–44.10.5194/isprs-archives-XLII-3-239-2018Search in Google Scholar

[51] Ito Y, Hosokawa M, eds. Damage estimation model using temporal coherence ratio. IEEE International Geoscience and Remote Sensing Symposium; 2002.Search in Google Scholar

[52] Sharma RC, Tateishi R, Hara K, Nguyen HT, Gharechelou S, Nguyen LV. Earthquake damage visualization (EDV) technique for the rapid detection of earthquake-induced damages using SAR data. Sensors. 2017;17(2):235.10.3390/s17020235Search in Google Scholar PubMed PubMed Central

[53] Yang Y, Xie C, Tian B, Guo Y, Zhu Y, Yang Y, et al. Large-scale building damage assessment based on recurrent neural networks using SAR coherence time series: A case study of 2023 Turkey–Syria earthquake. Earthq Spectra. 2024;40(4):2285–305.10.1177/87552930241262761Search in Google Scholar

[54] Yamaguchi Y. Disaster monitoring by fully polarimetric SAR data acquired With ALOS-PALSAR. Proc IEEE. 2012;100(10):2851–60.10.1109/JPROC.2012.2195469Search in Google Scholar

[55] Liu Z, Han B, Liu H, Li Z, Neng Y, Chen B, et al. Discussion of the seismogenic fault of the 2023 herat earthquake sequence and its building damage assessment constrained by Radar Interferometry. Geomat Inf Sci Wuhan Univ. 2023;48(11):722–33.Search in Google Scholar

[56] Du Y-N, Feng D-C, Wu G. InSAR-based rapid damage assessment of urban building portfolios following the 2023 Turkey earthquake. Int J Disaster Risk Reduct. 2024;103:104317.10.1016/j.ijdrr.2024.104317Search in Google Scholar

[57] Ferrentino E, Ferdinando N, Christian B, Laura G, Alessandra M, Migliaccio M. Multi-polarization C-band SAR imagery to quantify damage levels due to the Central Italy earthquake. Int J Remote Sens. 2021;42(15):5969–84.10.1080/01431161.2021.1933247Search in Google Scholar

[58] Saganeiti L, Amato F, Nolè G, Vona M, Murgante B. Early estimation of ground displacements and building damage after seismic events using SAR and LiDAR data: The case of the Amatrice earthquake in central Italy, on 24th August 2016. Int J Disaster Risk Reduct. 2020;51:101924.10.1016/j.ijdrr.2020.101924Search in Google Scholar

[59] Yu X, Hu X, Song Y, Xu S, Li X, Song X, et al. Intelligent assessment of building damage of 2023 Turkey-Syria Earthquake by multiple remote sensing approaches. NPJ Nat Hazards. 2024;1(1):3.10.1038/s44304-024-00003-0Search in Google Scholar

[60] Sun X, Chen X, Yang L, Wang W, Zhou X, Wang L, et al. Using InSAR and PolSAR to assess ground displacement and building damage after a seismic event: Case study of the 2021 Baicheng earthquake. Remote Sens. 2022;14(13):3009.10.3390/rs14133009Search in Google Scholar

[61] Karimzadeh S, Matsuoka M. A preliminary damage assessment using dual path synthetic aperture radar analysis for the M 6.4 Petrinja Earthquake (2020), Croatia. Remote Sens. 2021;13(12):2267.10.3390/rs13122267Search in Google Scholar

[62] ElGharbawi T, Zarzoura F. Damage detection using SAR coherence statistical analysis, application to Beirut, Lebanon. ISPRS J Photogramm Remote Sens. 2021;173:1–9.10.1016/j.isprsjprs.2021.01.001Search in Google Scholar

[63] Center FGR. Puerto Rico Mw 6.4 Earthquake preliminary damage assessments dashboard, federal emergency management agency. In: Center FGR, editor. Washington, USA: The Federal Emergency Management Agency (FEMA); 2020.Search in Google Scholar

[64] Filipponi F. Sentinel-1 GRD preprocessing workflow. Proceedings. 2019;18(1):11.10.3390/ECRS-3-06201Search in Google Scholar

[65] Touzi R, Lopes A, Bruniquel J, Vachon PW. Coherence estimation for SAR imagery. IEEE Trans Geosci Remote Sens. 1999;37(1):135–49.10.1109/36.739146Search in Google Scholar

[66] Fattahi H, Agram P, Simons M. A network-based enhanced spectral diversity approach for TOPS time-series analysis. IEEE Trans Geosci Remote Sens. 2017;55(2):777–86.10.1109/TGRS.2016.2614925Search in Google Scholar

[67] Rosen PA, Gurrola E, Sacco GF, Zebker H, eds. The InSAR scientific computing environment. EUSAR 2012; 9th European Conference on Synthetic Aperture Radar; 2012.Search in Google Scholar

[68] Yunjun Z, Fattahi H, Amelung F. Small baseline InSAR time series analysis: Unwrapping error correction and noise reduction. Comput Geosci. 2019;133:104331.10.1016/j.cageo.2019.104331Search in Google Scholar

[69] Lu C-H, Ni C-F, Chang C-P, Yen J-Y, Chuang RY. Coherence difference analysis of Sentinel-1 SAR interferogram to identify earthquake-induced disasters in Urban Areas. Remote Sens. 2018;10(8):1318.10.3390/rs10081318Search in Google Scholar

[70] Ishitsuka K, Tsuji T, Matsuoka T. Detection and mapping of soil liquefaction in the 2011 Tohoku earthquake using SAR interferometry. Earth, Planets Space. 2012;64(12):1267–76.10.5047/eps.2012.11.002Search in Google Scholar

[71] Olen S, Bookhagen B. Mapping damage-affected areas after natural hazard events using Sentinel-1 coherence time series. Remote Sens. 2018;10(8):1272.10.3390/rs10081272Search in Google Scholar

[72] Yazilim G. Cizenler Y. 2023 Turkey earthquakes building damage assessment map. 2023. Available at: https://www.designsafe-ci.org/data/browser/public/designsafe.storage.published/prj-3824v2/#details-942732811040452115-242ac11b-0001-012 (accessed 20 February 2023).Search in Google Scholar

[73] Xu S, Zhao X, Dimasaka J, Noh H. M7.8 2023 Turkey EQ building damage (Last Updated Mar. 9, 2023) building damage map in Turkey derived from damage proxy maps. Abailable at: https://www.arcgis.com/home/item.html?id=86941dbff14e450bbe3f89897373c59a. (accessed 18 March 2023).Search in Google Scholar

[74] Roeslin S, Ma Q, Juárez-Garcia H, Gómez-Bernal A, Wicker J, Wotherspoon L. A machine learning damage prediction model for the 2017 Puebla-Morelos, Mexico, earthquake. Earthq Spectra. 2020;36(2_suppl):314–39.10.1177/8755293020936714Search in Google Scholar

[75] Schratz P, Muenchow J, Iturritxa E, Richter J, Brenning A. Hyperparameter tuning and performance assessment of statistical and machine-learning algorithms using spatial data. Ecol Model. 2019;406:109–20.10.1016/j.ecolmodel.2019.06.002Search in Google Scholar

[76] Probst P, Wright MN, Boulesteix A-L. Hyperparameters and tuning strategies for random forest. Wiley Interdiscip Reviews: Data Min Knowl Discovery. 2019;9(3):e1301.10.1002/widm.1301Search in Google Scholar

[77] Malkomes G, Garnett R. Bayesian optimization for automated model. 30th Conference on Neural Information Processing Systems; Barcelona Spain; 2016. p. 41–7.Search in Google Scholar

[78] Gupta R, Shah M. RescueNet: Joint building segmentation and damage assessment from satellite imagery. 2020 25th International Conference on Pattern Recognition (ICPR), Milan, Italy, 2021. p. 4405–11.10.1109/ICPR48806.2021.9412295Search in Google Scholar

[79] Zheng Z, Zhong Y, Wang J, Ma A, Zhang L. Building damage assessment for rapid disaster response with a deep object-based semantic change detection framework: From natural disasters to man-made disasters. Remote Sens Environ. 2021;265:112636.10.1016/j.rse.2021.112636Search in Google Scholar

[80] Wang Z, Zhang F, Wu C. A method of disaster-damaged buidling sample library construction based on transfer learning. Int Arch Photogramm Remote Sens Spat Inf Sci. 2022;48:65–71.10.5194/isprs-archives-XLVIII-3-W1-2022-65-2022Search in Google Scholar

[81] Zhao L, Jie Y, Pingxiang L, Liangpei Z, Lei S, Lang F. Damage assessment in urban areas using post-earthquake airborne PolSAR imagery. Int J Remote Sens. 2013;34(24):8952–66.10.1080/01431161.2013.860566Search in Google Scholar

[82] Li L, Liu X, Chen Q, Yang S. Building damage assessment from PolSAR data using texture parameters of statistical model. Comput Geosci. 2018;113:115–26.10.1016/j.cageo.2018.01.018Search in Google Scholar

[83] Chen Q, Yang H, Li L, Liu X. A novel statistical texture feature for SAR building damage assessment in different polarization modes. IEEE J Sel Top Appl Earth Obs Remote Sens. 2020;13:154–65.10.1109/JSTARS.2019.2954292Search in Google Scholar

[84] Liu Y, Li L, Chen Q, Shu M, Zhang Z, Liu X, eds. Building damage assessment of compact polarimetric SAR using statistical model texture parameter. 2017 SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA); 2017.10.1109/BIGSARDATA.2017.8124923Search in Google Scholar

[85] Li Q, Lixia G, Zhang J. A correlation change detection method integrating PCA and multi- texture features of SAR image for building damage detection. Eur J Remote Sens. 2019;52(1):435–47.10.1080/22797254.2019.1630322Search in Google Scholar

[86] Cotrufo S, Constantin S, Fabio GT, Boccardo P. Building damage assessment scale tailored to remote sensing vertical imagery. Eur J Remote Sens. 2018;51(1):991–1005.10.1080/22797254.2018.1527662Search in Google Scholar

[87] Sun W, Shi L, Yang J, Li P. Building collapse assessment in urban areas using texture information from postevent SAR data. IEEE J Sel Top Appl Earth Obs Remote Sens. 2016;9(8):3792–808.10.1109/JSTARS.2016.2580610Search in Google Scholar

[88] Khvedchenya E, Gabruseva T. Fully convolutional Siamese neural networks for buildings damage assessment from satellite images; 2021. arXiv preprint arXiv:2111.00508.Search in Google Scholar

[89] Shen Y, Zhu S, Yang T, Chen C. Cross-directional feature fusion network for building damage assessment from satellite imagery; 2020. arXiv preprint arXiv:2010.14014.Search in Google Scholar

[90] May S, Dupuis A, Lagrange A, De Vieilleville F, Fernandez-Martin C. Building damage assessment with deep learning. Int Arch Photogramm Remote Sens Spat Inf Sci. 2022;43:1133–8.10.5194/isprs-archives-XLIII-B3-2022-1133-2022Search in Google Scholar

[91] Lu W, Wei L, Nguyen M. Bitemporal attention transformer for building change detection and building damage assessment. IEEE J Sel Top Appl Earth Obs Remote Sens. 2024;17:4917–35.10.1109/JSTARS.2024.3354310Search in Google Scholar

[92] Kaur N, Lee C-C, Mostafavi A, Mahdavi-Amiri A. Large-scale building damage assessment using a novel hierarchical transformer architecture on satellite images. Comput Civ Infrastruct Eng. 2023;38(15):2072–91.10.1111/mice.12981Search in Google Scholar

[93] Braik AM, Koliou M. Automated building damage assessment and large-scale mapping by integrating satellite imagery, GIS, and deep learning. Comput Civ Infrastruct Eng. 2024;39(15):2389–404.10.1111/mice.13197Search in Google Scholar

[94] Lin Q, Ci T, Wang L, Mondal SK, Yin H, Wang Y. Transfer learning for improving seismic building damage assessment. Remote Sens. 2022;14(1):201.10.3390/rs14010201Search in Google Scholar

[95] Cheng C-S, Behzadan AH, Noshadravan A. Deep learning for post-hurricane aerial damage assessment of buildings. Comput Civ Infrastruct Eng. 2021;36(6):695–710.10.1111/mice.12658Search in Google Scholar

[96] Ma Y, Zhou F, Wen G, Gen H, Huang R, Liu G, et al. Assessment of building and eletrical facilities damaged by flood and earthquake from satellite imagery. Int Arch Photogramm Remote Sens Spat Inf Sci. 2022;46:133–40.10.5194/isprs-archives-XLVI-3-W1-2022-133-2022Search in Google Scholar

[97] Bai Y, Hu J, Su J, Liu X, Liu H, He X, et al. Pyramid pooling module-based semi-siamese network: A benchmark model for assessing building damage from xBD satellite imagery datasets. Remote Sens. 2020;12(24):4055.10.3390/rs12244055Search in Google Scholar

[98] Deng L, Wang Y. Post-disaster building damage assessment based on improved U-Net. Sci Rep. 2022;12(1):15862.10.1038/s41598-022-20114-wSearch in Google Scholar PubMed PubMed Central

[99] Bai Y, Adriano B, Mas E, Gokon H, Koshimura S. Object-based building damage assessment methodology using only post event ALOS-2/PALSAR-2 dual polarimetric SAR intensity images. J Disaster Res. 2017;12:259–71.10.20965/jdr.2017.p0259Search in Google Scholar

[100] Ge P, Gokon H, Meguro K. Building damage assessment using intensity SAR data with different incidence angles and longtime interval. 2019;14:456–65.10.20965/jdr.2019.p0456Search in Google Scholar

[101] Miura H, Midorikawa S, Matsuoka M. Building damage assessment using high-resolution satellite SAR images of the 2010 Haiti earthquake. Earthq Spectra. 2015;32:150324070149002.10.1193/033014EQS042MSearch in Google Scholar

[102] Hajeb M, Karimzadeh S, Fallahi A. Seismic damage assessment in Sarpole-Zahab town (Iran) using synthetic aperture radar (SAR) images and texture analysis. Nat Hazards. 2020;103(1):347–66.10.1007/s11069-020-03991-0Search in Google Scholar

[103] Park S-E, Jung YT. Detection of earthquake-induced building damages using polarimetric SAR data. Remote Sens. 2020;12(1):137.10.3390/rs12010137Search in Google Scholar

[104] Xu J, Lu W, Li Z, Khaitan P, Zaytseva V. Building damage detection in satellite imagery using convolutional neural networks; 2019. arXiv preprint arXiv:1910.06444.Search in Google Scholar

[105] Brunner D, Lemoine G, Bruzzone L. Earthquake damage assessment of buildings using VHR optical and SAR imagery. IEEE Trans Geosci Remote Sens. 2010;48(5):2403–20.10.1109/TGRS.2009.2038274Search in Google Scholar

[106] Ahmadi SA, Mohammadzadeh A, Yokoya N, Ghorbanian A. BD-SKUNet: Selective-kernel UNets for building damage assessment in high-resolution satellite images. Remote Sens. 2024;16(1):182.10.3390/rs16010182Search in Google Scholar

[107] Bouchard I, Rancourt M-È, Aloise D, Kalaitzis F. On transfer learning for building damage assessment from satellite imagery in emergency contexts. Remote Sens. 2022;14(11):2532.10.3390/rs14112532Search in Google Scholar

[108] Calantropio A, Chiabrando F, Codastefano M, Bourke E. Deep learning for automatic building damage assessment: application in post-disaster scenarios using UAV data. ISPRS Ann Photogramm Remote Sens Spat Inf Sci. 2021;1:113–20.10.5194/isprs-annals-V-1-2021-113-2021Search in Google Scholar

[109] Qing Y, Ming D, Wen Q, Weng Q, Xu L, Chen Y, et al. Operational earthquake-induced building damage assessment using CNN-based direct remote sensing change detection on superpixel level. Int J Appl Earth Obs Geoinf. 2022;112:102899.10.1016/j.jag.2022.102899Search in Google Scholar

[110] Wang Y, Cui L, Zhang C, Chen W, Xu Y, Zhang Q. A two-stage seismic damage assessment method for small, dense, and imbalanced buildings in remote sensing images. Remote Sens. 2022;14(4):1012.10.3390/rs14041012Search in Google Scholar

[111] Grünthal G. European macroseismic scale 1998 (EMS-98). Luxembourg: Conseil de l Europe; 1998.Search in Google Scholar

[112] Dell'Acqua F, Gamba P. Remote sensing and earthquake damage assessment: Experiences, limits, and perspectives. Proc IEEE. 2012;100(10):2876–90.10.1109/JPROC.2012.2196404Search in Google Scholar

[113] Gupta R, Hosfelt R, Sajeev S, Patel N, Goodman B, Doshi J, et al. xBD: A dataset for assessing building damage from satellite imagery. arXiv; 2019. arXiv preprint arXiv:1911.09296.Search in Google Scholar

Received: 2025-02-09

Revised: 2025-07-07

Accepted: 2025-07-11

Published Online: 2025-09-09

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/geo-2025-0882

Keywords for this article

earthquake; Sentinel-1; building damage; multi-class classification

Creative Commons

BY 4.0