1. Introduction
The provisions of the 1954 Hague Convention for the Protection of Cultural Property in the Event of Armed Conflict clearly state that cultural assets must not be the target of hostilities [
1]. Emphasizing that the humanitarian catastrophe because of an armed conflict is much more serious, the protection of cultural heritage, historical monuments and arts is of great importance, since they are part of a nation’s identity and represent the historical development of a country. The ongoing armed conflict between Russia and Ukraine, which started in February 2022, threatens some centuries-old cultural property. In addition to seven (cultural) world heritage sites, Ukraine is home to a large number of historic city centers and cultural treasures. Their exposure to damage makes it crucial to provide evidence of the sites’ condition, to be ready for recovery, or to investigate allegations of war crimes [
2]. Earth observation (EO) imagery is particularly effective for monitoring and assessing the state of cultural property in hostile situations where the locations are not accessible and on-site inspection is inhibited [
3].
The United Nations Educational, Scientific and Cultural Organization (UNESCO) is committed, among other tasks, to safeguarding and preserving humanity’s cultural heritage. Since the beginning of the Ukrainian war, UNESCO has regularly published figures on destroyed cultural property throughout the country [
4]. Damages are recorded by elaborately cross-referencing reported incidents with information from multiple sources. The United Nations Satellite Centre (UNOSAT), as part of the United Nations Institute for Training and Research (UNITAR), engaged in damage assessment based on EO data by visual interpretation of very high resolution (VHR) optical imagery [
5]. Although the analysis of VHR optical data plays a fundamental role in the delineation of the impact caused by armed conflict, manual damage mapping takes a considerable amount of time and resources and may be limited in area. Furthermore, the availability of cloud-free VHR optical data after an event can cause limitations for visual analysis. Support for this vitally important task may arise from an effective change-detection (CD) methodology as an additional component in damage assessment related to armed conflict. A combination of CD with semi-automated data processing provides a pre-evaluation of the area of interest, giving important clues as to where to complement the optical investigation. CD methodology is particularly suited in the context of urban disaster monitoring, which could help analysts prioritize their activities and empower damage assessment procedures from optical imagery and the evaluation of incident reports.
CD with EO synthetic aperture radar (SAR) imagery has been investigated for rapid damage assessment in urban environments caused by natural hazards or military confrontation [
6,
7,
8,
9]. The main objective of such applications is the comparison of imagery collected over the same geographical location at different dates to discriminate changes to the built-up area, indicating destruction. Since active SAR-based satellites operate independently of weather conditions and time of day, they are advantageous over optical data in disaster-related applications, when timely monitoring is required [
10]. Both incoherent and coherent methodologies using SAR data were described [
11]. While incoherent methods typically compare the SAR backscattering intensity (amplitude) to identify scene changes, coherent change detection (CCD) utilizes both the amplitude and phase of the imagery. Coherence describes the similarity of the reflection properties between two SAR scenes acquired at different points in time [
12]. Changes in, for example, surface conditions result in a decorrelation between the images, which is expressed by a decrease in coherence values [
13]. Since coherent-based CD has the potential to identify even the smallest changes in the land surface through time, the technique promises high usability for damage assessment in urban areas affected by disastrous events. Some studies have successfully applied the difference of coherence between pre- and post-event image acquisitions as an indicator for delineating the areas affected by natural hazards [
14,
15] or found the level of coherence decreased related to the severity of building damages [
16].
SAR-based damage mapping in urban areas has predominantly relied on VHR imagery with a spatial resolution to the sub-meter level [
17,
18,
19,
20]. Although such images enable a fine-grained analysis of building features as the primary target of investigation, on-demand high-resolution data usually comes at high costs and small area coverage, confining the approaches to single pre- and post or only post-event data. The European Space Agency’s (ESA) Copernicus Sentinel-1 mission provides freely accessible and timely available SAR images in a considerable spatial and temporal frequency and coverage. Several studies have showcased the benefits of using Sentinel-1 data over commercial VHR imagery to monitor building damage caused by armed conflict due to their capacity to map on a regular, short time interval [
21,
22,
23]. The availability of a consistent and frequent time series of image acquisitions offers an opportunity to apply multi-pair CD approaches for the systematic, repeated monitoring of building structures and cultural assets. Temporal analysis could visualize a chronological sequence of impact locations and provide evidence of whether damage occurred within a specific time window [
21].
However, damage detection from SAR data alone can be challenging, since limitations arise due to, for example, the side-looking geometry of SAR sensors and seasonal changes or snow cover on the ground, which may produce a high degree of false positives in the results [
14]. Several authors have sought to achieve improvements in damage detection related to armed conflict by integrating Sentinel-1 data with Sentinel-2 optical sensors [
24,
25,
26]. A promising approach was pursued by [
27], who visualized the structural damage caused by missile strikes by analyzing Sentinel-2 and the near-infrared (NIR) range of the spectrum. Ref. [
6] combined fire indices derived from the Sentinel-2 NIR band for mapping burnt areas with the CCD of urban buildings. Despite the high revisit frequency of a single Sentinel-2 satellite of 10 days and, with two satellites, the combined constellation of 5 days, the optical images were compromised by dense cloud cover throughout the weeks after the Russian invasion. This prevented a full investigation of the selected areas of interest and the detection of potential damage related to the time windows chosen for this study. In contrast, the availability of Sentinel-1 data with a temporal resolution of at least 12 days allowed for analyzing changes based on a consistent time series of regular intervals both pre- and co-conflict.
The aim of this study is to evaluate the applicability of medium-resolution Sentinel-1 SAR images to detect and assess the damage to buildings and cultural property in urban areas during an ongoing armed conflict using CCD methodology. We investigate if multi-temporal coherence change analysis can highlight significant impacts on building structures and cultural assets and enable approximating overall damage. We address the question of whether coherence change can serve as a proxy variable for building damage analysis. We applied logistic regression as a widely used method for binary classification [
28]. It enables the evaluation of the functional relationship between explanatory variables and a categorical outcome, such as “yes” versus “no” or “damaged” versus “undamaged” and uses this relationship to estimate class probabilities [
29]. We create a logistic regression model (LRM) to determine the likelihood of damage or collapse of buildings and cultural property as a function of the observed loss of coherence as an independent spatial proxy variable. The damage observations to be used as a training dataset to fit the LRM were obtained by visual inspection and manual tagging of building samples based on Google satellite imagery with high spatial resolution showing the extent of damage on 9 May 2022. Although high spatial resolution imagery is essential for effectively detecting changes on the building level in dense urban environments, the study aims to demonstrate the capability, strengths, and limitations of medium spatial resolution Sentinel-1 SAR imagery with broad spatial coverage (250 km swath) and high temporal resolution for keeping track of dynamic changes caused by armed conflicts. To counter limitations from the spatial resolution of the Sentinel-1 data and to increase the potential of urban area monitoring, we include free OpenStreetMap (OSM) data to relate changes that occurred at different points in time explicitly to buildings and heritage sites [
13,
24,
30]. The estimated damages are compared with independent damage assessment data published by UNOSAT and with UNESCO-verified damaged cultural sites to assess the usability and reliability of Sentinel-1 SAR coherent time-series analysis to distinguish impact areas over time and approximate total damage.
4. Logistic Regression Analysis
In the present study, a key assumption is that the presence or absence of building damage is related to the amount of coherence loss. The choice of data analysis technique should allow for finding the relationship between these two data factors. The aforementioned logistic regression is a common method used to estimate class probabilities on a binary range from zero to one. It is used to determine the likelihood of buildings being damaged or undamaged, providing meaningful and interpretable analysis outputs.
Several authors, like [
50,
51], successfully used LRM to assess the damage vulnerability of urban structures towards natural hazards. Ref. [
52] predicted the level of heritage building decay depending on various building properties. The principle of logistic regression is to model a binary target variable
where
can only take the values of 0 or 1 as the two possible states (e.g., 0: “undamaged” and 1: “damaged”) [
53]. A logistic model is based on the logistic regression function. We are interested in the probability P of an outcome being true
given a set of independent variables
, expressed as in Equations (4) and (5):
where
is the damage probability, z is the so-called logit, which represents a linear regression model with
as the independent variables,
as the regression beta coefficients, and
as the error value. A positive
indicates that increasing the value of
is associated with an increasing probability of
. With logistic regression, an s-shaped function curve is fit to the data, following a sigmoid function and running asymptotically towards
and
. The values of the logistic function are interpreted as the likelihood that a building was damaged or destroyed.
4.1. Sample Data
Fitting an LRM requires training data and spatially independent test data to evaluate model performance.
Figure 5 shows the workflow for the logistic regression analysis.
Figure 5 left shows the overlay of the OSM building outlines and the pixel-wise coherence difference ΔCoh
tot (16 February 2022 to 23 May 2022) to derive the mean coherence difference per building as an independent variable of the LRM. The light-to-dark red color indicates an increasing coherence difference and, thus, a loss of coherence.
Figure 5 right shows the locations of the manually tagged building samples. Points in red color represent buildings categorized as damaged. Points in black color represent buildings categorized as undamaged.
We used an updated Google satellite image available for the Mariupol area to generate our own set of damage reference data as input for model development. Since the mentioned satellite images depict the extent of damage on 9 May 2022 according to Google Earth, we used coherence difference ΔCoh
tot calculated from Mariupol pre-conflict coherence Coh
1 and co-conflict coherence map Coh
tot for model training. We created a random sample of 1000 buildings spread across the entire city area of Mariupol by overlaying the OSM building outlines with the Google satellite imagery. After the visual inspection, we mapped the damaged or undamaged state of the sampled buildings. The damage to the buildings was determined on the basis of debris around the building; partial collapse of the roof, indicated by dark spots on the rooftops; or complete collapse, i.e., the building structure was no longer recognizable (
Figure 6a–c). We categorized buildings as undamaged if they appeared to be structurally intact, e.g., if the roof appeared intact and no debris was visible (
Figure 6d). However, the buildings may have suffered damage that could not be recognized on vertical satellite imagery. The sample dataset contained 396 points classified as damaged and 604 points classified as undamaged buildings. Slightly degraded optical images due to light clouds or other limiting factors could have influenced the visual damage assessment. As a result, the reference data may not be as accurate as the damage data obtained from on-site inspection. To determine the reference damage data associated with the estimated coherence loss, we performed zonal statistics in QGIS software version 3.10.8-A Coruña to calculate the average coherence difference for each building footprint. For cultural assets that were only carried as point locations, the average coherence difference was calculated with a 25 m buffer area, considering Sentinel-1 SLC pixel size.
4.2. Model Calibration
The statistical analysis was carried out in RStudio version 2022.02.3+492. We partitioned the reference dataset, choosing a split ratio of 70:30, meaning we randomly sampled 70% of the dataset for training and 30% for testing purposes, respectively. Logistic regression is a specific form of the generalized linear model (GLM), which can accommodate for a wide range of distributions, such as binomial, and is implemented in the R package Stats as a glm() function [
53,
54,
55] as follows:
where the response variable is the binary target variable (damaged = 1, undamaged = 0), and the so-called predictor is the independent variable (mean coherence difference). In the present case, we built the logistic model with a single predictor (explanatory variable).
Table 4 presents the model summary output, including the confidence levels for the estimated model parameters. The positive sign of the predictor’s coefficient, as well as the magnitude, implies a strong positive correlation, meaning that the likelihood of observing damage significantly increases with an increase in coherence loss.
The model was applied to the test dataset to estimate the probability that a building is damaged at a given difference in coherence and, hence, coherence loss.
Figure 7 visualizes the results as predictions across the range of coherence difference values derived from the test dataset. The observations of undamaged or damaged buildings are shown along the y = 0 and y = 1 lines. The grey lines identify the pointwise 95% confidence interval.
4.3. Model Performance Evaluation
The predictive performance of the model was evaluated by the area under the receiver operating characteristic curve (AUC) using the R package pROC [
56]. We chose AUC as a standard measure because it is independent of a previously selected classification threshold and summarizes overall model performance over all possible thresholds [
57]. AUC values range on a scale of 0 to 1, with a value of 0.8 or higher attributing a generally good performance, whilst a value of 0.5 is considered a classifier worse than a random estimate. A comparison of AUC for the training and test datasets showed the logistic model achieved a value of 0.80 for the training dataset and slightly decreased for the test dataset, to 0.78, which seems an overall acceptable discrimination of the data. We classified the predicted values based on an optimal 0.49 cut point, which produced the highest AUC for the test dataset.
We then produced a confusion matrix using R package caret [
58], computing several performance metrics. The overall classification accuracy achieved about 76%. Among the buildings that appeared to have remained undamaged, 150 out of a total of 171 were predicted correctly, which corresponds to a specificity of about 88%. This results in a low false-positive rate of about 12%, representing the proportion of buildings incorrectly classified as damaged although they seemed intact. For the buildings where damage was recorded, 77 out of a total of 129 were correctly classified, which corresponds to about 60% correct predictions, i.e., sensitivity. It is worth mentioning that, in the context of damage detection and monitoring of potential threats to buildings and, in particular, cultural properties, we may be more concerned with improved sensitivity, hence, a higher true positive rate, to miss out on less damage. Therefore, we suggest considering a lower probability threshold, hence, accepting a higher rate of false alarms rather than overlooking potentially damaged buildings and not taking any further action.
4.4. Evaluation Based on UNOSAT Data
We calculated the model output for all OSM building structures identified in Mariupol and Kharkiv and applied the binary threshold classification to distinguish between undamaged and damaged locations. The predicted damage sites were plotted with UNOSAT reference data to assess the overall plausibility and compare the spatial distribution patterns [
21]. We identified a mismatch between the UNOSAT points and the building footprints in all cases, which complicated the comparison of the results. We filtered UNOSAT damage locations by excluding points outside the study areas and only points representing “destroyed”, “severely damaged”, “moderately damaged”, and “possibly damaged” structures were used. Each reference dataset represents the combined damages derived from repeated visual inspection carried out by UNOSAT between 14 March 2022 to 12 May 2022 in Mariupol and 24 April 2022 to 15 June 2022 in Kharkiv, respectively. Whereas UNOSAT damage assessment for Kharkiv covered the entire city area (
Figure 8), the analysis for Mariupol was limited to two residential areas, Livoberezhnyi and Zhovtnevyi district, including the city center (
Figure 9). We, therefore, extracted the OSM building data and model outputs based on the given district boundaries.
The OSM dataset indicates at least 86,871 structures in the Kharkiv area and 17,161 structures in the two selected city districts of Mariupol. Based on the building stock, UNOSAT reported that less than 1%, or 748, of the buildings in Kharkiv were affected, whereas in Mariupol about 32%, or 5647, of the structures sustained visible damage. This compares with our predicted damage rate of about 0.7% (610 buildings) for Kharkiv and about 39% (6974 buildings) for the two Mariupol districts over the same time period. For the entire city area of Mariupol, we classified 17,742 structures as damaged out of a total of 56,785, which represents about 31% of the built-up area.
A visual inspection of the major discrepancies between the results in Mariupol revealed our predictions to be consistent with the building destruction that is clearly visible in the corresponding Google satellite imagery. These deviations are probably due to the time intervals selected by UNOSAT to illustrate damage mapping, where parts of the data only represent the results from an earlier analysis dated 14 March 2022. For the Kharkiv area, our findings quite agree with UNOSAT about the overall extent of the damage, although the results indicate an underestimation compared to the manually tagged damages. The damage is predominantly identified in residential areas towards the northeastern outskirts of the city. The settlement structure in this area is reflected by small buildings lined up along roads and embedded in agricultural and afforested areas. This could influence the results in terms of Sentinel-1 spatial resolution, which makes damage detection at the individual building level more difficult and seems more aimed at delineating affected neighborhoods. However, both analyses detected hotspots throughout the city center, particularly in the vicinity of Freedom Square.
6. Discussion
Our study investigated the applicability of Sentinel-1 SAR multi-temporal CCD for damage assessment in the cities of Kharkiv and Mariupol as targets of the Russian military offensive in Ukraine. SAR images are weather-independent, which makes them highly suitable to complement optical observation for continuous monitoring and damage assessment of urban neighborhoods and embedded cultural assets. Such approaches are often based on very high-resolution SAR sensors that provide on-demand imagery and fine-scale mapping [
17,
20]. However, they are costly and might cover a small area. Therefore, freely available Sentinel-1 data can compensate due to the high temporal and spatial coverage. Since we aimed to identify hotspots of coherent changes, indicating potential impact and correspondingly increased damage risk on cultural assets, Sentinel-1 spatial resolution seems adequate at the required level of detail of preliminary damage maps.
We exploited the difference between two subsequent coherence image pairs as the parameter for discriminating between damaged and undamaged areas. Refs. [
15,
61,
62] proposed normalized coherence difference, as calculated by Equation (6), as an improved indicator for coherent changes showing higher sensitivity in the discrimination of building damage in disasters:
The performance of both parameters was examined based on statistical metrics of the predictive outcome of the logistic model, including cross-tabulation between predicted and observed values, overall model accuracy, and ROC computation. However, the normalized difference could not improve building damage discrimination regarding the statistic criteria. We, therefore, found a traditional coherence difference calculation in the form of simple image subtraction to be a sufficient and straightforward method.
We tackled the issue of classification, i.e., finding reasonable discrimination thresholds by applying logistic regression as an approach for estimating building damage probability in relation to coherence loss between two time periods. Since the independent input variable was binary, this implies the back-transformation of the model output into a binary form. A building is then either damaged or undamaged. Our intention is to use the full information of the predicted probabilities providing preliminary damage maps that visualize areas with a high likelihood for damage occurrence. We suggest this is an adequate way for users to individually interpret the probability output in a map based on their specific application context. Further work is intended to establish multiple scenarios of building damage probability utilizing the upper and lower confidence intervals of the prediction. Presumably due to the general lack of sufficient reference data, particularly in disaster areas, the applicability of logistic regression analysis for damage classification and vulnerability mapping has been explored by only a few studies, such as [
50,
51,
52,
63]. Our statistical analysis showed a strong correlation between coherence loss and observed building damage. Therefore, the individual coherence difference images could already be used to quickly identify potentially affected areas and be considered as a basis for detailed follow-up investigation.
The qualitative evaluation with the UNOSAT reference data showed a good agreement in the distribution of building damage. For both study sites, Sentinel-1 SAR damage detection identified structural damages that remained undetected in the reference dataset. Here lies the advantage of a semi-automated approach that can detect areas of change more effectively within a broader area in a timely manner. Whereas manual image interpretation of events is subject to experts’ knowledge, and it is time-consuming. However, the results of the semi-automated approach using freely available data can support visual inspection, especially when time is an issue. In addition, the spatial coverage of Sentinel-1 makes it possible to process an entire city area at once, meaning that the change analysis does not have to be limited to smaller sections. However, for the Kharkiv area, the evaluation revealed a potential weakness of Sentinel-1 in terms of spatial resolution compared to VHR visual analysis, which makes damage detection at the individual building level more difficult and is more aimed at delineating affected neighborhoods.
It has already been pointed out in earlier studies that it is preferable to include additional GIS data such as building outlines to mask out coherence variation caused by other effects [
13,
24,
59]. The visual comparison of satellite data, including derived image products and open GIS datasets, showed some limitations with regard to completeness, positioning, and accuracy of fit, especially in the available OSM building footprint. Buffering of building outlines to accommodate different observation conditions and data sources could improve detecting trends of coherence loss on the building level, resulting in an improved model parameterization and performance [
19]. Due to the spatial resolution of Sentinel-1 data, coherence estimation will still be affected by the backscattering intensity of other objects in a target pixel, which increases the degree of uncertainty regarding the damage detection of small buildings, buildings enclosed by dense vegetation [
6,
59], or single damaged objects located within rather unaffected neighborhoods.
The time-series analysis enabled the detection of hotspots of coherent changes and narrowed the time range within which the damages occurred. However, apparent contrasts were found between the severe condition of individual buildings observed in the VHR optical images and the estimated amount of coherence loss and damage probability, where no or only moderate changes seemed to have occurred. The selected locations in Kharkiv presented in
Figure 13 are good examples to demonstrate the difficulties of damage assessment on the single building level. While the inspection of pixel-based coherence difference indicated even the partial damage of larger buildings, the aggregation at the building level (average coherence loss per building outline) might have resulted in less significant coherence difference and, hence, a lower estimated damage probability.
The aim of the developed logistic model is the spatial and temporal transferability of its application to provide accurate predictions in another target area. A model’s capacity for prediction can decline when transferred to another region or time period other than that upon which it is trained. One reason lies within the observed variation of coherence and its contribution to the predictive relation for damage probability estimation on the different sites. Coherence characteristics vary between the training data, that is, the coherence difference data used for model calibration and the data used for model prediction.
We obtained pre-conflict coherence based on two Sentinel-1 SAR scenes acquired as closely as possible before the start of the armed conflict as a representation of the normal coherence distribution without any damage-causing event. By comparing the results obtained from the pre-conflict image pairs for both the Mariupol and Kharkiv built-up areas, it is apparent that Kharkiv is showing lower coherence values than Mariupol (
Table 2,
Figure 4). This is despite the fact that both image pairs have been obtained over the same 12-day temporal interval and similar acquisition dates. The selection of images with a minimum time lag should result in low temporal decorrelation and produce consistent coherence maps [
64]. However, different regions are subject to different rates of change in surface properties that influence coherence variability and the definition of thresholds for differentiation between natural and damage-related coherence loss [
48].
Where decorrelation of the pre-conflict coherence map (reference) caused by the background environment, such as meteorological influences (snow coverage) or urban green spaces, becomes dominant, it decreases the expected stability in coherence over the built-up area prior to changes related to the armed conflict. Damage is assessed by detecting changes, i.e., estimating the difference between corresponding coherence maps. Overall low coherence in the reference image results in less significant coherence loss between the pre-conflict and co-conflict coherence maps in areas affected by changes related to the armed conflict. In consequence, lower coherence difference values may result in a lower estimated damage probability, leading to an increase in the false-negative rate.
Thus, a procedure is required to correct the input data for model calibration to improve the predictive performance of our model and to ensure transferability to another target area and different timeframes.
Possible approaches are shown by studies that explored a series of pre-event coherences calculated at regular intervals for the preceding year [
14,
23,
65]. The intention is to identify trends in pre-event coherence for an individual area by calculating the sequential mean and standard deviation per pixel to be used as a baseline. The statistical values are then considered as a threshold to distinguish significant disaster-induced coherence loss. Ref. [
48] trained a deep-learning algorithm based on observed pre-event coherence behavior to forecast normal coherence distribution expected without damage events. However, the workflows were developed for comparison with a single post-event SAR image. Further investigation is needed on whether and how to apply the methodology on a time series of coherences calculated for adjacent image pairs.
Mariupol shows a substantial loss of coherence over the whole investigated time period (Coh
tot,
Table 2). The extent of coherence decrease could be attributed to the temporal baseline of 96 days between the pre- and co-conflict image acquisitions [
64]. In addition, the pre- and co-conflict coherence maps used for calculating coherence loss as input for the model calibration compare winter and early summer images from February and May 2022, representing a relatively wide temperature and precipitation range. Thus, the coherence images could be affected by natural changes in surface properties or seasonal effects in terms of the overall average value that might increase false alarms. Further investigations include a coherence difference analysis based on coherence maps calculated from image acquisitions of the corresponding pre-year period as a reference to derive a baseline for coherent change caused by seasons, not disastrous events. However, this effect would also have to be observed in the Kharkiv study region, which shows a higher similarity in terms of average coherence in the built-up area. From this, it can be concluded that the detected changes, i.e., coherence loss in Mariupol, can be related to the armed conflict and are an indication of the widespread extent of the damage.
7. Conclusions
SAR Sentinel-1 coherent-based change detection highlighted areas of major destruction over time. We provided building damage estimates based on a coherence difference analysis of subsequent pre- and co-event coherence maps and logistic regression. Coherence loss proved to be sufficient as a proxy measure for building damages related to armed conflict, with the drop level indicating the level of damage probability. Given the probability output, users can determine to which degree damage is likely to occur and prioritize further investigations based on individual requirements.
The damage probability maps should be used as guidance, giving a first, timely initial evaluation of potentially damaged areas. The free availability, regular acquisition dates, and weather independence of Sentinel-1 SAR imagery serve for the production of a stringent time series of preliminary damage maps and the detection of major changes with sufficient spatial and temporal accuracy. A semi-automated workflow to extract areas of change, as applied in this study, takes considerably less time and can support further detailed visual inspections using VHR data.
Our analysis complements visual damage verification and supports repeated monitoring that extends over larger areas and allows users to focus on detailed follow-up surveys. However, the analysis is less reliable over vegetated or agricultural areas, leading to false positives. Since we expected the detected changes to be related to the destruction of buildings caused by armed conflict, the results should only be interpreted for the built-up areas. We included additional GIS data to increase accuracy, with a focus on the building stock and heritage sites, along with enhanced usability of the damage assessment at a user-relevant information level.
The presented CCD methodology is sensitive to changes induced by armed conflict and applicable in areas with large-scale destruction. However, small-scale change or moderate structural damage may not be detected due to Sentinel-1’s spatial resolution. Modeling techniques and training data determine the transferability of the developed LRM and could be improved by integrating additional reference data from different areas and timeframes for training and prediction. However, the sampling procedure was limited by the availability of additional VHR optical satellite imagery for damage assessment.