"*" indicates required fields
A practical, applied data studio package for developing and performing building damage assessment (BDA) workflows to support humanitarian emergency preparedness, response, and recovery following disasters and conflict events. It combines satellite imagery, machine learning models, and infrastructure datasets with case studies, dashboards, and operational use cases to support rapid, reproducible assessments.
The studio focuses on remotely sensed BDA using optical and synthetic aperture radar (SAR) imagery, pre/post-event change detection, and manual-to-semi-automated workflows, including strategies for acquiring very-high-resolution (VHR) commercial imagery. It's designed for data teams, NGOs, government analysts, and humanitarian practitioners, supporting both rapid operational assessments and longer-term recovery monitoring.
This data studio has been developed alongside a companion conflict related building damage assessment decision-making guide available here.
What is a BDA? The process of identifying, mapping, and classifying damage to buildings and structures following disasters or armed conflict. BDA's are conducted via field surveys, aerial photography, drones, satellite imagery, or a combination. Remote sensing enables rapid, scalable assessments across large or inaccessible areas.
What a BDA supports: Estimating affected populations and shelter needs, prioritizing search and rescue, supporting emergency logistics and access planning, guiding debris management, informing recovery planning, and supporting human rights investigations.
Common causes of building damage: Earthquakes, flooding, tropical cyclones/severe storms, landslides, wildfires, volcanic eruptions, and explosions/armed conflict. Damage severity depends on event intensity, building design, local conditions, and repeated exposure. The appropriate analysis method varies by hazard type.
Before an emergency occurs
Immediately following or during an event/conflict
This catalog aggregates self-contained visualizers that pair pre and post event satellite imagery with analytical overlays to surface impact signals early. By unifying imagery sources (PlanetScope, SkySat, Maxar imagery, Esri basemaps) and derived layers (e.g., damage inference, flood extent, change masks), it accelerates situational awareness for responders, analysts, and decision-makers.
Building damage assessment performed by Microsoft AI for Good following an earthquake in Turkey on February 6th 2023. Analysis was performed using very high resolution (VHR) imagery and convolutional neural networks (CNNs) to assess building damage.
https://github.com/microsoft/satellite-imagery-labeling-tool
This case study develops a scalable machine learning algorithm to map and identify damaged and destroyed buildings across Ukraine. The machine learning algorithm was trained on Synthetic Aperture Radar (SAR) imagery from Copernicus Sentinel-1 SAR imagery and Sentinel-2 optical imagery. It also includes an interactive damage explorer hosted by Google Earth Engine.
https://olidietrich.users.earthengine.app/view/ukraine-damage-explorer
Near real-time satellite-derived structural damage assessments produced using Sentinel-1 synthetic aperture radar coherent change detection. Data are made available for research, humanitarian, and journalistic use only. We are actively expanding coverage to additional regions affected by armed conflict for near real-time updates and historical data.
Oregon State Conflict Ecology Lab - https://conflict-ecology.org/
Decentralized Damage Mapping Group - https://conflict-damage.org/
This dashboard shows satellite imagery-based comprehensive damage assessment across the Gaza strip based on satellite images from as recent as July 2025, compared to satellite images retrieved throughout 2023 and 2024. Results can be filtered by damage severity, municipality, and/or facility type (i.e., essential infrastructure).
Data download: https://gaza-unosat.hub.arcgis.com/pages/data
Gaza Emergency Response Data Hub: https://gaza-unosat.hub.arcgis.com/
This dashboard produced by UNITAR/UNOSAT combines satellite analysis from the UNOSAT team, Copernicus, and SERTIT with ground-based data collection from the UN ASIGN data collection app.
Sentinel-2 is a European mission that provides multiple optical bands at varying resolutions. Its twin satellite design provides a high revisit time of 5 days. Sentinel-2 data is provided free of charge to the public and can be downloaded through NASA’s Earthdata access portal or other satellite data repositories, such as Google Earth Engine. Sentinel-2 imagery is available at 10m resolution (bands 1-4), 20m resolution (bands 4-8A, 11-12), and 60m resolution (bands 9-10).
Define an AOI by drawing a polygon, selecting a point, or by uploading a .kmz/.kml file.
Filter by time of day and intended date for imagery acquisition (Hint: the best time to obtain satellite data is between 10:30AM and 12:00 PM local time!)
Sentinel-1 is a European mission that consists of 3 satellites, Sentinel-1A, -1B, and -1C. Sentinel-1 imagery is generated from a C-band Synthetic Aperture Radar (SAR) instrument that penetrates clouds and provides day-and-night coverage.
Define an AOI by drawing a polygon, selecting a point, or by uploading a .kmz/.kml file.
Filter by time of day and intended date for imagery acquisition.
Planet makes available select imagery for major disaster events, including major earthquakes, floods, storms, wildfires, and human-made disasters. Individuals and groups affiliated with NGOs, government authorities, and international organizations assisting with response efforts are eligible for access. To request access and see if you qualify, one can fill out the form and Planet’s Crisis Response Program will be in touch as soon as possible.
Fill out the form provided through the link, describing your organization and the use of requested data, and Planet will be in touch.
Maxar provides very high resolution (VHR) satellite imagery of sites before and after disaster to assist in humanitarian response, primarily from the Worldview-1, -2, and -3 satellites. Following a disaster, Maxar releases imagery packages of before and after the event under a Creative Commons 4.0 license, allowing for rapid humanitarian and community response efforts. This program only makes imagery for certain disasters available; coverage is limited.
Navigate to Maxar’s Open Data Program page using the left sidebar in the map portal. Open Data events will be listed by year; often with new releases highlighted. You must register for a Maxar Geospatial Program (MGP) account before you can access and download images.
The OpenTopography Data Catalog is an open-source resource that catalogs community contributed, open-source, and high resolution Digital Elevation Models (DEMs) and assorted topographical data. Global and regional datasets can be explored and downloaded. Data catalog can be sorted by data type and location.
Access data catalog and select intended dataset. An OpenTopography account is required for most downloads.
Google’s Open Buildings dataset is a global building footprint dataset designed for humanitarian applications. The polygon data is comprised of a set of .csv files, which can be downloaded by individual cells on a global map. Similarly, point data can be downloaded as well. Google also provides a Colab notebook on how to download building footprint data for a specific country or geographic area.
Select a parcel to download; or, follow the instructions in Google Colab to filter data by country or geographic area. Downloaded .csv files contain georeferenced polygon or point data; project and display data using a GIS software of your choice and proceed with analysis!
The Advanced Rapid Imaging and Analysis (ARIA) project provides Damage Proxy Maps (DPMs) and standard science products derived from SAR and InSAR imagery and deformation modeling. DPMs are designed for analysis support in rapid response and disaster mapping. ARIA offers Standard Displacement Products as well as DPMs and Flood Proxy Maps (FPMs).
Link provided for the 2023 Turkey and Syria earthquake.
ARIA Product Search
InSAR standard data products are currently searchable and downloadable through the ARIA Product Search website. This site requires a NASA Earthdata URS user login and requires users to add “ARIA Product Search” to their URS approved applications.
Event Response
Data products that have been generated for event response are made available on the ARIA Share site, which does not require a login. The products are generated for both emergency response and scientific rapid response. These include Damage Proxy Maps, Flood Proxy Maps, InSAR and GPS results for earthquake and volcanic eruptions.
This resource is a Github repository developed by Microsoft AI for Good that provides a workflow to develop and fine tune a building damage assessment model using satellite imagery. The model is tiered, providing an annotation tool, a more advanced segmentation mask, and finally scripts for fine-tuning the model and finalizing inference.
The repository also provides a workflow for analyzing damage proxy maps (DPMs), which are pre-made maps using SAR imagery that provide measure of coherence pre- and post-disaster. The repository also provides a script to measure the damage levels recorded by changes in coherence across the DPMs.
The .py files included in the Github repository contain workflows specific to each desired task; i.e., creating segmentation masks, downloading building footprints, etc.
Event-based emergency mapping products for disasters worldwide. Rapid Mapping activations often include damage assessment layers (such as building damage grading), reference maps, and delineation layers (such as affected area and infrastructure impacts), published as downloadable GIS files and cartographic PDFs.
Open the Copernicus EMS Mapping portal and go to the Rapid Mapping section.
Find the specific activation for the disaster of interest (by country, date, or activation ID).
Download the GIS deliverables (commonly SHP/GeoPackage, GeoTIFFs, and PDFs) for the relevant Area of Interest.
In GIS software, load the damage grade layer (if provided), then join or summarize by admin boundaries to produce counts by district, neighborhood, or other reporting units.
Satellite-derived crisis mapping and damage assessment products created for specific disasters. Many products include building damage assessments and supporting maps; availability of downloadable GIS layers varies by event, but PDFs are commonly published quickly for operational use
Search the UNOSAT products catalog for the disaster (country + event name + year).
Download the latest damage assessment map PDFs and, when provided, the accompanying GIS files (often SHP/KML/GeoPackage).
In GIS, load the damage layer and use simple summaries (counts by admin unit, distance to key infrastructure, density hot spots) to support briefs and response planning.
The BRIGHT dataset is a VHR imagery dataset designed to support the development of deep learning models for disaster response. It covers five types of natural disasters and two types of man-made disasters across 14 disaster events in 23 regions worldwide, with a particular focus on developing countries.
BRIGHT can be used to train AI models and perform semi-automated or fully automated building damage classification. It is open source through a Github repository and supported by Linux OS.
The Github repository provides a tutorial for preparing an environment for and downloading the BRIGHT dataset. The dataset can be accessed and utilized anywhere you use Python.
The repo also provides additional supplemental resources, including publications that have successfully utilized BRIGHT.
Overview:
Use Whatsapp and mobile applications such as ArcGIS Field Maps and UN-ASIGN to coordinate field teams and collect records of damaged buildings immediately following a disaster event.
Crowd-sourced photographs can be georeferenced and added to mapping software and interactive dashboards delineating the extent of the damaged area, as well as assist in aid and rescue efforts.
Ideal for early-phase response and environments where technical capacity may be limited.
Audience: Local NGOs, civil protection agencies, volunteer groups
Prerequisites: Smartphones with GPS; basic data connectivity; a shared communication channel (WhatsApp/Signal/SMS); basic GIS knowledge.
Data & Tools:
Agree on a damage scale that field teams can apply consistently.
Use 4 levels with plain language: No visible damage, Minor, Major, Destroyed. Add one “Unknown” option for unsafe access or uncertainty.
Prepare a collection form that forces the essentials.
Minimum fields: building status (the 4-level scale), photo required, notes, collector name, timestamp, and a confidence flag (High, Medium, Low).
Load building footprints so reports snap to real buildings.
Download Google Open Buildings for the affected area and publish it as a reference layer in the map used by field teams. The goal is to tap the correct building polygon (or nearest point) instead of dropping random points.
Create a simple assignment grid for coverage.
Split the area into small zones (neighborhood blocks, a fishnet grid, or admin sub-areas). Assign each team a zone. This prevents ten teams surveying the same street while another area is untouched.
Collect reports with photos.
If possible, each report should include: one wide photo (shows the building and context) and one close photo (shows damage). Try to stand still for 2 seconds before capturing so GPS stabilizes.
Use satellite imagery as a context layer, not as proof, in this workflow.
If Maxar Open Data or Planet Disaster Data exists for the event, use it to help teams navigate, understand blocked roads, and prioritize clusters, but keep the source of truth as field verification at this stage.
Publish a live view for decision makers with three filters only.
Filters: damage level, date/time window, and confidence. Avoid adding lots of layers; try to streamline presentation so viewers can understand (1) where the worst damage is, and (2) where more coverage is needed.
Output: georeferenced dataset of damaged buildings; verified crowd-sourced photo library; interactive dashboard of reported damages & their severities
Validation: cross-verify with open building footprints; integrate population counts; verify coordinates & time-stamps
Risks: Inconsistent labeling, duplicated records, device battery and connectivity failures, and safety constraints that prevent verification in the hardest hit areas
Overview:
Use satellite imagery (SAR, optical, or LIDAR) to annotate buildings based on visual inspection of damage, comparing pre- and post-disaster imagery.
Ideal for team-based analysis environments, where image analysts can work in shifts following a disaster to identify damage and coordinate response.
Audience: GIS units; university research teams; data science NGOS; technical response teams
Prereqs: QGIS, a NASA Earthdata login, basic imagery handling (load, reproject, clip), and a shared labeling guide.
Data & Tools:
Clip building footprint layer to disaster area.
Download Google Open Buildings for the region and clip to the affected AOI in QGIS.
Pull “before” and “after” imagery with a date rule.
Pick one pre-event image close to the event date, and one post-event image as soon as possible after the event. If you have VHR imagery (Maxar or Planet), use it as the main interpretation layer. If not, Sentinel-2 can still support neighborhood-scale screening, but it will miss single-building detail.
If clouds or smoke block optical imagery, switch to Sentinel-1.
Sentinel-1 SAR sees through clouds and works at night. It does not present as optical imagery, so treat it as a change detector: you are looking for areas/ pixel values that changed a lot between dates.
Arrange layers in QGIS.
Load layers in this order: basemap, pre-event imagery, post-event imagery, footprints, then any SAR-derived change layer if available. Use swipe tools or two map panels to compare.
Define damage guidelines and visual cues with annotators.
Damage in optical VHR can look like: collapsed roofs, missing roof sections, debris piles, shadows that indicate partial collapse, and irregular roof texture changes.
Damage in Sentinel-2 can look like: large clusters of destruction, roof color shifts across blocks, widespread debris signatures, and burned areas, but not reliable per-building detail.
SAR coherence drops can indicate structural change, but also vegetation change, soil moisture changes, or rubble clearing.
Annotate by selecting a building footprint and assigning a class.
Select damaged building footprints and assign a standardized damage class. You can also add notes on evidence from VHR imagery or coherence scores for context.
Outputs: binary or multi-class damage assessment maps; quantitative statistics
Validation: compare classification outputs with ground-truth samples; conduct accuracy assessments; VHR imagery validation where available
Risks & Considerations: Manual labeling can drift across annotators. Standardized guidelines and a reviewer are important.
Overview:
Use satellite imagery (SAR, optical, or LIDAR) and machine learning methods to perform semi-automated building damage assessment.
Ideal for teams with higher technical capacity and for classifying large amounts of buildings to a high degree of accuracy.
Audience: GIS and remote sensing analysts; data scientists; humanitarian agencies or government units with geospatial capacity
Prereqs: NASA Earthdata login; basic Python ability; cloud notebook environment; QGIS capability (including working with raster data)
Data & Tools:
Investigate ARIA availability.
For many disasters, ARIA publishes event response layers (including DPMs) that are meant to speed up early analysis. If a DPM exists, it becomes the fastest way to screen where damage is likely concentrated.
Download the DPM and load it in QGIS.
Add the DPM raster, confirm its projection, and clip it to the AOI. Set a color ramp that makes “high change” stand out.
Interpret the DPM correctly before using it operationally.
The DPM can be helpful in identifying areas with high potential damage. High change can mean building damage, but it can also reflect vegetation loss, soil moisture shifts, rubble clearing, or water. The DPM is most effective when (1) combined with building footprints and (2) at least one visual confirmation source.
Attach a DPM score to each building footprint.
Use zonal statistics (mean or max DPM value within each footprint). Output a building layer with a new field like “DPM_score.” This converts a raster into a list that operations teams can sort and filter.
Create a triage threshold and a review queue.
Split buildings into three bins based on DPM_score: High priority review, Medium, Low. Start by reviewing only the High bin so the workflow stays fast.
Confirm high-score buildings with imagery, using the best available source.
If Maxar Open Data or Planet is available, confirm visually for the High bin.
If not, use Sentinel-2 for neighborhood-scale confirmation and flag buildings for field verification.
If optical is blocked, use Sentinel-1 change cues as supporting evidence.
Assign a final damage class and store evidence.
For each reviewed building, store: damage class, evidence type (VHR, field photo, SAR-only), and a short explanatory note.
(Optional) Run a semi-automated model workflow for scale.
If the AOI is large, use the Microsoft toolkit workflow to structure training labels and run inference, especially when you have VHR imagery and enough labeled examples from step 7. Keep the first run simple: binary damage versus no-damage often performs better early than four-class labeling.
(Optional) Automate repeat runs for new imagery releases.
Put the steps into a Colab notebook with three configuration inputs: AOI polygon, pre and post date windows, and the chosen imagery source. Re-run when new post-event imagery arrives, then compare outputs day to day.
Publish a map and a table that match how response teams work.
Map for spatial clusters, table for action lists. Provide a ranked list of buildings with high DPM_score and confirmed damage class so teams can assign inspections, debris planning, or shelter assessments.
Outputs: A ranked building list with damage classes and evidence, plus map layers suitable for dashboards and briefings.
Validation:
Maintenance: Save the AOI, thresholds, and labeling guide as a “deployment kit” so the next event starts from a known baseline. Re-run the pipeline when new imagery arrives.
Risks & Considerations: Misinterpretation of proxy layers; over-confidence when optical confirmation is missing; inconsistent labeling across analysts.
Overview:
Use the BRIGHT multimodal building damage assessment dataset to train a machine learning workflow for semi-automated building damage assessment using optical and SAR imagery.
This workflow is ideal for organizations or researchers with moderate-to-high technical capacities that seek to scale damage assessments across large AOIs and rapidly classify damaged buildings after disasters.
The workflow introduces the use of pretrained models and cloud notebooks for damage assessment, as well as outlines a process for triage setup in semi-automated workflows.
BRIGHT supports workflows using optical and SAR imagery, which means that BRIGHT-based methods can continue operating during cloud cover, smoke, or low-visibility conditions.
To learn more about BRIGHT, explore these resources:
Audience: GIS and remote sensing analysts; humanitarian mapping teams; government/NGO disaster response units with ML capability
Prereqs: Basic Python ability; cloud notebook environment (Google Colab recommended); experience with raster imagery processing
Data & Tools:
Software & Libraries:
*not comprehensive
Download the BRIGHT dataset and organize the imagery and labels into training, validation, and testing folders.
The dataset includes pre-event optical imagery, post-event SAR imagery, and building-level damage labels.
Maintain a consistent folder structure so you can keep things straight during preprocessing and training.
2. Review the labeling schema before training.
Inspect damage classes included in the dataset and review how labels are distributed across categories. This can look like running simple descriptive statistics on the training data.
3. Preprocess optical and SAR imagery.
Despeckle; normalize; reduce noise and/or cloud cover.
4. Build a baseline model workflow.
Build a model workflow using a relatively simple and well-documented model architecture such as UNet (Ronneberger et al., 2015), DeepLabV3+ (Chen et al., 2018), or Prithvi (Jakubik et al., 2023).
Refer to package documentation to load and build model workflows.
5. Train and fine-tune the model.
Train the model using the imagery provided by the BRIGHT dataset and using the building damage labels. Some metrics to look out for during training are the F1 score and IoU (Intersection over Union). Adjusting brightness, rotation, or even flipping the image can improve model generalization in some cases.
6. Once the model has been trained, run inference on your AOI.
Apply the trained model to new disaster imagery that coverings your AOI. The model will generate predictions for individual buildings, and should include both a damage class and a confidence score. Using these outputs, one can produce GIS-compatible layers for further analysis.
7. But first, triage.
Sort model outputs in Colab into high-, medium-, and low-confidence groups to prioritize analyst review. Within these confidence groups, analysts should first inspect buildings with high predicted damage scores (or classes) to identify the most impacted areas within the AOI.
Model predictions can be validated against VHR imagery, field photographs, or external damage assessments if available.
8. Publish operational outputs.
Export results as GIS layers, building tables, or dashboard-ready products response teams can use operationally. Useful outputs typically include:
Maps of your AOI will help identify spatial clusters or damage, and tables help support field inspections, debris planning, and broader coordination.
Outputs: Building-level damage classifications; confidence ranked building lists; GIS-compatible vector layers; operational layers
Validation: compare outputs to VHR imagery, if available; compare outputs to external assessments (i.e., UNOSAT)
Risks & Considerations: GPU requirements may limit deployment in low-resourced environments
