Products
Solutions
Retail & eCommerce
Defense
Logistics
Autonomous Vehicles
Robotics
AR/VR
Content & Language
Smart Port Lab
Federal LLMs
Resources
Company
Customers
See all customers
Mariupol Damage Identification
Public dataset identifying damage sustained in Mariupol as of March 21st, 2022 on a by-structure level. This dataset aims to support the mission-driven AI community across government, humanitarian, and academic sectors.
OVERVIEW
Mariupol Damage Identification Dataset
Developed with commercial satellite imagery, this dataset supports structure and damage identification use cases. This data is composed of:
310
Sq. KM Scene of Mariupol
3
Levels of damage assessment
81,000+
structures in total
DATA COLLECTION
Satellite Constellation Imagery
The image used is an electro-optical multispectral satellite image, with a GSD of approximately 50cm and spherical error of 10 meters.
DATA ANNOTATION
Identifying And Assessing Damage
Scale annotates large areas through best in class human-machine teaming. After imagery ingestion, machine learning data curation separates out areas without structures. The remaining areas are annotated by human labelers using smart annotation tools, programmatic checks and quality sampling to meet statistical confidence goals.
Polygon Labeled
In this Mariupol dataset, each structure is labeled with a polygon. Because of the difficulty of placing precise polygons in dense urban areas, Scale created an auto-segment polygon tool. This tool allows labelers to place a bounding box over an area with a building and leverages machine learning to transform that bounding box into a precise polygon covering the building’s rooftop.
All buildings receive a damage assessment on a three-point scale:
No Damage
No visible holes or cracks in the roof
Damaged
Roofs with holes or partially collapsed
Destroyed
Roof is completely collapsed
