Ukraine data collection
For our analyses of building destruction during the Ukraine war, we draw on three existing datasets.
War damage data
We use war-related damage data from ref. 9, which identifies probable damage across Ukrainian human settlements between March 2022 and October 2023. This dataset was generated using a coherent change detection framework applied to Sentinel-1 SAR data75.
Building footprints
Building footprint data were sourced from Overture Maps, providing a comprehensive dataset of 26,714,652 building footprints in Ukraine. The primary sources for these footprints include Microsoft ML Building Footprints (72.8% of buildings) and OpenStreetMap (27.2% of buildings).
Tessellated geometries
The unit of analysis is tessellated geometries around populated places. We use the polygons provided by the VIINA dataset73, which are based on the GeoNames gazetteer and include 33,141 tessellated geometries in Ukraine.
Boundaries
Spatial data on admin1 and admin2 boundaries, and settlement boundaries for Mariupol and Sievierodonetsk, are from State Scientific Production Enterprise ‘Kartographia’, as provided by the Humanitarian Data Exchange platform74.
Areas of control
For areas of control, we used the daily data provided by Violent Incident Information from News Articles (VIINA)73, which draws on four sources: VIINA event reports, DeepStateMap, the Institute for the Study of War, and Wikipedia.
Pre-war population data
Pre-war population statistics for Mariupol and Sievierodonetsk are taken from the State Statistics Service of Ukraine (2021, Number of Present Population of Ukraine).
Conflict event data
Information on fatalities is taken from the UCDP GED v25.11,14.
Ukraine data analysis
Building-level damage data
We disaggregate damage data based on ref. 9 to the building level using Overture Maps data. The original damage data are provided as 10-m projected pixel grids corresponding to built-up area identified by Global Human Settlement Layer76. We attribute damage to Overture building footprint data if a building footprint fully (\( > \)99%) overlaps with pixel-wise damage data. This resulted in 246,777 likely damaged Overture Maps building footprints.
Improvement section
To examine differences in building destruction before and after territorial control changes between Russian and Ukrainian gains (Fig. 3), we first identify changes in control at the level of tessellated geometries and then link these to the war damage data for the corresponding month–geometry combinations.
We aggregate control patterns to a monthly timestep to align with the monthly timestep of damage data. We classify a given month as under full control by one actor if that actor maintained control throughout the entire month; as transitioning from actor A to actor B if A controlled the territory on the first day of the month and B on the last day; and as ‘other’ for all remaining patterns, including contested status at the beginning or end of the month, or control by actor A at both the beginning and end but back and forth in control in between. This yields 569,606 month–geometry combinations under full Ukrainian control, 107,094 under full Russian control, 5,731 for Russian gains, 4,372 for Ukrainian gains, and 19,261 for contested or mixed control during the period of analysis (February 2022 to October 2023).
Next, we identify first stable transitions, defined as the first instance in which a territory transitions from Ukraine to Russia (Russian gains) or from Russia back to Ukraine (Ukrainian gains). A transition is considered stable if it was followed (and for Ukraine regaining territory, also preceded) by stable control. For Ukrainian gains, the territory had to be under full Russian control for at least two months before and under full Ukrainian control for at least two months after the transition. For Russian gains, we require at least two months of full Russian control following the transition. We do not require Russian gains to be preceded by two months of full Ukrainian control because the VIINA control data are only available from February 2022 onward. We are confident that Russian gains in the analysis had been under Ukrainian control prior to the transition because, by construction, month–geometry combinations for Russian gains must begin with Ukrainian control, which excludes territories already under Russian control in February 2022. To control for any buildings damaged prior to our analysis window, we only assess newly damaged buildings across the comparison window and further normalize per 1,000 undamaged buildings at the start of each month.
By focusing on clear and stable transitions, excluding areas that fluctuated between the two sides, and controlling for earlier building damage, we can more confidently attribute damage before and after control changes to Russian and Ukrainian gains separately. A total of 2,438 transitions meet our criteria for first stable Russian gains, of which 2,130 occurred in the initial phase of the invasion (February and March 2022), 108 in April 2022, and 200 between May 2022 and October 2023. For Ukraine, we identify 967 first stable transitions that meet our criteria, all of which fall in the period between May 2022 and October 2023. For the box plots in Fig. 3b and Extended Data Figs. 3 and 4, and the statistical analyses presented in Extended Data Tables 1 and 2 and Supplementary Table 1 in Supplementary Information, we restrict the analyses to populated areas, removing geometries with no buildings in our building layer. For Russian gains, this leaves 273 populated areas in February and March 2022, 29 in April 2022, and 95 between May 2022 and October 2023. For Ukraine, it leaves 148 populated areas between May 2022 and October 2023.
For the map of territorial control change (Fig. 3a), we visualize all first stable transitions that occurred during the frontline war (May 2022 to October 2023). In addition, we also visualize all tessellated areas that experienced at least one monthly change in territorial control during the study period. For the latter, we include areas that were (1) not under full Russian control in February 2022 but were under full Russian control for at least one month between March 2022 and October 2023, or (2) under full Russian control in February 2022 but were under full Ukrainian control for at least one month between March 2022 and October 2023.
To visualize the distribution of damage, we normalize the timelines by assigning month 0 to the month of the control change (first stable transition), with a five-month window spanning two months before to two months after (m−2 to m+2). Using the tessellated month–geometry combinations with normalized timelines, we attribute damage to Russian versus Ukrainian gains and visualize the distributions using box plots (Fig. 3b). Each box aggregates the newly damaged-building counts across the cells with at least one newly damaged building in that specific relative month. Cells with zero new damage in a given month do not contribute to that month’s box but may contribute to others.
To assess the difference in targeting patterns, we conduct a two-step statistical analysis. First, we compare, at each relative month in the five-month window (m−2 to m+2), the proportion of populated areas that experienced damage around control change between areas gained by Russia and those retaken by Ukraine (two-sided Pearson chi-squared test, results reported in the main text and Extended Data Table 1). Second, we conduct two-sided Kolmogorov–Smirnov and Mann–Whitney tests to compare levels of damage associated with Russian and Ukrainian gains, conditional on at least one newly damaged building in month m (reported in Fig. 3 and Extended Data Table 2). The results of the two-step analysis show that, in the months around the transition, Russian territorial gains are more often associated with building damage than Ukrainian gains (Extended Data Table 1), and the median number of newly damaged buildings per 1,000 intact buildings is roughly twice as high for Russian gains, a significant difference (Extended Data Table 2). To check the independence assumption underlying the per-month tests, we compute Moran’s I on log1p(newly damaged buildings) using a spatio-temporal weights matrix (full results and interpretation in Supplementary Information, section E).
The war damage data are available from March 2022 to October 2023. For the main analysis, we include only control changes from May 2022 onward, which is approximately when the frontline war began. Many sources date the beginning of the ‘Battle of Donbas’ to 18 April 2022; however, because our war damage data are at monthly resolution, we use May 2022 as the starting point. This approach has two advantages: it aligns with the temporal resolution of our damage data, and it ensures that damage data for the two months preceding control changes are available (as the damage data only extend back to March 2022). In the extended data, we present the same analysis for the full period covered by the war damage data (Extended Data Fig. 3) and further disaggregate Russian gains into three phases: the initial multi-front invasion, the reorientation phase and the frontline war (Extended Data Fig. 4).
Enrichment section
For the enrichment analysis presented in Fig. 5, we use the month-settlement unit as the common unit of analysis to compare fatalities and war damage over time. To identify relevant conflict events, we perform a spatial intersection between the UCDP GED event coordinates and the settlement boundaries. While the majority of fatalities were recorded with a temporal precision of one month or finer, some events involving high fatality counts were recorded with a lower temporal precision level (level 5), indicating a range exceeding one month. Excluding these data points would have distorted the overall level of fatalities, particularly in Mariupol. To maintain the month-settlement unit, we disaggregate these entries by distributing fatalities uniformly across the days within the reported date range.
Regarding spatial precision, our primary estimates include only events coded at the settlement level (the highest precision level). Given the documented urban bias in conflict event data, we are confident that most events occurring within Mariupol and Sievierodonetsk are captured at this level. However, to account for events occurring within settlement boundaries that may have been coded with lower spatial precision, we also provide a broader estimate using spatial precision level 2 (event occurred within 25 km), represented by a dotted line. All fatality figures are based on the UCDP GED ‘best’ estimates. In the Extended Data Fig. 8, we also provide ‘high’ estimates reported by UCDP GED (events coded at settlement-level precision only), which show even more pronounced contrasts in fatality levels between the two cities. To facilitate comparison between cities of different size, we report cumulative fatalities as a percentage of pre-war inhabitants, using 2021 population baselines.
The war damage data, originally reported at monthly intervals, required only spatial aggregation via intersection with the settlement boundaries. We report these results as the percentage of the total buildings affected, providing both detected damage and adjusted estimates based on reported true positive rates9, capped at 100%.
Myanmar data collection
For our analyses of civilian targeting of the Rohingya population in Myanmar, we draw on five existing datasets.
Conflict event data
Information on violent events, including involved actors and reported fatalities, is sourced from the UCDP Georeferenced Event Dataset (GED) v25.11,14. In Extended Data Fig. 7 we replicate our enrichment analyses using ACLED15.
War damage
To capture building destruction, we rely on damage assessment data produced by United Nations Institute for Training and Research (UNITAR)–United Nations Satellite Centre (UNOSAT), which record satellite-detected destroyed and damaged settlements in the Buthidaung, Maungdaw and Rathedaung townships of Rakhine State71. These data are based on a manual analysis of satellite imagery collected on multiple dates between 31 August and 11 October 2017.
Thermal anomalies
To identify thermal anomalies, we used data from the near-real time VIIRS 375 m Active Fire product (VNP14IMGT)77.
Populated areas
To distinguish populated from unpopulated areas, we use raster data from WorldPop70, drawing on the constrained estimates of total population per grid square at 3 arcsec resolution (approximately 100 m at the equator), R2025A v1. We use estimates for 2016—the year preceding the 2017 violent campaign—because violence in Rakhine State led to large-scale population displacement. To select thermal anomalies detected in or near populated areas, we also use village point locations from the Myanmar Information Management Unit (MIMU) v9.678.
Administrative units
Spatial data on the administrative boundaries of Myanmar are taken from MIMU v01, via the Humanitarian Data Exchange platform72. We use administrative levels one (state), two (district) and three (township).
Myanmar data analysis
Improvement section, settlement level
We conduct local analyses at the level of individual settlements (villages or towns) across multiple locations in northern Rakhine State. In a first step, we use text-based conflict event data (UCDP GED) to identify settlements in which government violence against civilians involving killings was reported during the initial phase of the 2017 campaign. The majority of lethal conflict events occurred between 25 August and 9 September 2017. Our analysis focuses on the window of 25 August to 16 September 2017, representing the narrowest temporal interval that captures the peak violence while also aligning with the capture dates of satellite-derived war damage data. We restrict the conflict event data to observations with the highest level of spatial precision—that is, events for which the affected settlement was recorded.
For each settlement, we define a settlement-specific spatial window based on the visually interpreted settlement extent. We then integrate gridded population data at 100 m resolution from WorldPop, clipped to each settlement’s spatial window, to delineate inhabited space prior to violence. Population cells with zero values are excluded, and the remaining cells are converted to polygons to enable spatial intersection with satellite-derived damage assessments. Prior to this intersection with the more spatially precise satellite-derived damage data, the entire populated polygon is assumed to be affected by violence.
To further refine this area, we incorporate UNITAR-UNOSAT damage polygons and restrict these data to observations dated on or before the second available satellite image acquisition date (16 September 2017). We spatially intersect the damage polygons with each settlement’s populated polygon to classify each populated image pixel within the spatial window as damaged or undamaged. This enables us to distinguish between settlements that experienced physical damage and those that did not within a tightly bounded local context, thereby refining the spatial footprint implied by settlement-level conflict event reporting.
The article presents the results of this analysis for one such settlement, Chut Pyin (Fig. 2), where a large-scale massacre occurred during the analysis period56. In Extended Data Figs. 1 and 2, we present the same analyses for additional settlements, and in Supplementary Information, section D we discuss how they can be conducted at scale. Study locations were selected based on the availability of satellite-derived damage assessments in an otherwise information-poor conflict environment.
Improvement section, state level
To illustrate how satellite-derived data can be used to improve spatial precision in cases where reported events are located at coarse administrative levels, we again focus on the 2017 violence in Rakhine State, Myanmar.
We restrict the UCDP GED conflict event data to events recorded in Rakhine State. Of the 54 events reported in 2017, 42 (78%) occurred within a two-week period between 25 August and 9 September 2017; we therefore limited our analysis to this time frame.
For approximately half of these events, we know within which village or town they occurred. All these events with high spatial precision occurred in Maungdaw and Sittwe districts in the northern part of Rakhine State. For an additional third of events, the district is known; in all cases, this is Maungdaw district. The remaining events—accounting for 281 reported fatalities (best estimate)—are recorded as occurring in Rakhine State without any sub-state location information.
To narrow down the areas likely affected by the violent campaign, we analysed active fire detections from the NRT VIIRS 375 m product (VNP14IMGT) between 25 August and 9 September 2017. These detections are point features, which we restricted to the Rakhine State boundary using administrative level-one shapefiles. We further intersected the fire detections with a 500-m radial buffer around settlement locations to differentiate potential village burning from wildfires. To further distinguish these events from non-conflict-related events, such as agricultural burning or industrial heat emissions, we conducted visual verification using publicly available historical very high-resolution imagery available through Google Earth and optical imagery from the Copernicus Sentinel-2 mission79. Fires were positively attributed to the campaign only when thermal detections co-occurred with visible evidence of settlement destruction, such as scorch marks, burned structures, or collapsed roofs.
We then aggregated confirmed thermal anomalies to townships by intersecting detections with third-level administrative polygons. We constructed a binary indicator for each township denoting whether it contained at least one confirmed thermal anomaly during the analysis period. This produced a state-wide map of administrative units with physical evidence consistent with widespread settlement burning. We detected thermal anomalies exclusively in the townships in which spatially precisely coded killings took place. No thermal anomalies were detected elsewhere in Rakhine State during the period under investigation.
Using this exclusion-based logic, we infer that the imprecisely coded conflict events are likely to have occurred in the same northern part of the state as the precisely located events, rather than elsewhere in Rakhine. Importantly, this process does not reassign individual events to specific locations. Rather, it demonstrates how satellite-derived data can be used to narrow the plausible geography of coarse event locations for descriptive and statistical downstream analyses in cases where lethal violence and the burning of structures spatially and temporally co-occur. Because VIIRS detections capture thermal activity rather than violence per se, and because fires may go undetected due to cloud cover or timing, the absence of detections should not be interpreted as definitive evidence of the absence of violence.
Enrichment section
For the enrichment analyses presented in Fig. 4 and Extended Data Fig. 6, we integrate text-based conflict event data from UCDP GED with satellite-derived damage data from UNITAR-UNOSAT into a single tabular dataset. The spatial unit of analysis is the township. We focus on three townships in northern Rakhine State—Maungdaw, Buthidaung and Rathedaung—for which UNOSAT data are available.
To compare fatalities and war damage over time, we first align the damage data and the conflict event data temporally. Damage assessments are tied to specific image acquisition dates, which are irregular over time. Conflict event data, by contrast, are temporally more fine-grained and, in the large majority of cases, report the exact date of an event. Because the UNOSAT data have the coarser temporal resolution, they define the temporal unit of analysis. In our case, the UNOSAT data contained 22 distinct acquisition dates, with intervals between acquisitions ranging from 1 to 35 days. For interpretability of the time series, we collapse closely spaced acquisition dates, resulting in 17 analysis periods with intervals between 5 and 35 days. These periods constitute the temporal unit of analysis used throughout the enrichment example.
To aggregate the conflict event data to these periods, we first subset the data to events that are known to have occurred in Rakhine State. For approximately 80% of events, the precise date on which they occurred is reported (temporal precision 1). The rest of events are recorded with lower temporal precision and information about the period within which they occurred. To allow alignment with the UNOSAT periods, we distribute fatalities uniformly across the days within the reported date range—a common procedure in conflict research for temporally disaggregating event data80. After constructing the daily conflict event data, we restrict the sample to the period of UNOSAT damage data availability (ending 18 March 2018). Although the violent campaign began on 25 August 2017, the first UNOSAT image was not acquired until 31 August 2017. Consequently, damage detected between the first two image acquisitions is attributed to the full period from the campaign’s onset to the date of the second image. We assign each daily observation to the corresponding UNOSAT acquisition period, enabling direct aggregation of fatalities to the satellite-defined temporal units.
Next, we spatially intersect UNOSAT damage polygons with township boundaries (administrative level 3) to assign each observation to a specific administrative unit. In instances where a single damage polygon spanned multiple townships, we retain only the fragment with the largest overlap to ensure unique assignment. We apply a similar filtering process to the UCDP GED conflict data, restricting the sample to events with spatial precision levels 1 (exact location) and 2 (known location within a 25 km radius). Events at precision level 3 (second-level unit known) are excluded as they are too spatially imprecise for our township-level analysis. This results in the exclusion of over half of the recorded events, underscoring how improvements in spatial precision (such as those demonstrated in Fig. 2) facilitate more advanced forms of data integration.
We then aggregate both datasets to a common township-period unit of analysis. For the damage data, we calculate the total destroyed or damaged area in m2 and convert to km2. For the conflict event data, we compute the sum of the best-estimate fatalities. To enable consistent comparisons over time, we complete both panels by explicitly adding missing township–period combinations and setting damaged area and fatalities to zero where no observations were recorded for a given township-period.
Finally, we merge the two datasets into a single township-by-period panel, which forms the basis for the township-level time series shown in Fig. 4 and Extended Data Figs. 6 and 7. Given the highly skewed distribution of both variables, we visualize both series using a log10(x + 1) transformation. This approach preserves zero values while facilitating comparison across periods of varying magnitudes. The main text presents the comparative plots for Maungdaw and Buthidaung townships. In Extended Data Fig. 6, we present the corresponding visualizations for Rathedaung, and in Extended Data Fig. 7 the fatality counts using ACLED data instead of UCDP GED.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
%20SOURCE%20Julian%20Chokkattu(2).jpg "The Fairphone (Gen 6+) Review: Finally in the US")
