HASTE delivered an overnight earthquake damage map for Venezuela relief teams

Following the June 2026 earthquakes in Venezuela, Catholic Relief Services received an AI-supported satellite damage map by the next morning. Its digital-programming lead estimated that a comparable manual assessment would have taken about two weeks. Microsoft's September 22 account describes how the map helped teams locate damage and plan assistance. The comparison is the relief organisation's operational estimate, not a controlled evaluation of lives saved.[1]

Microsoft's High-speed Assessment and Satellite Tracking for Emergencies, or HASTE, compares imagery from before and after a disaster. An analyst identifies samples of damaged and intact structures, and the system applies those patterns across a much larger area. This is a form of event-specific machine learning rather than an all-purpose model that can infer damage perfectly from any photo. Satellite coverage, image resolution, clouds, local architecture and the time the imagery was captured all affect what its outputs can mean.[1][3]

The map became a decision input

Microsoft reported that HASTE had supported assessments for 23 major disasters across 16 countries since 2023. In Venezuela, Catholic Relief Services worked with local partner Caritas and also adapted the analysis to flag potentially damaged routes. The United Nations Office for the Coordination of Humanitarian Affairs received the results and circulated them through its response network. The useful result was a faster shared view of where visible building damage clustered, enabling teams to discuss priorities before visiting every affected community.[1]

OCHA also brought HASTE assessments together with eight other disaster-mapping products in a shared viewer. Microsoft's data included metadata and confidence rankings, helping responders compare different assessments. That provides a concrete account of coordination: organisations could examine where several products agreed, where they differed and which areas still needed attention.[1]

In the same reporting, emergency managers described using early damage estimates after wildfires in Spokane, Washington, to anticipate how many households might need assistance. That is another concrete administrative outcome: faster resource estimates while formal inspections were still pending. The examples do not imply satellite imagery can replace structural assessments or find everyone who needs help. A building may appear intact from above despite an unsafe interior, and social vulnerability may be invisible in the pixels.[1]

A human still chooses what a model should look for

The workflow is revealing because its human input is not merely signing off at the end. An analyst must identify the particular kind of damage expected after that event and select examples that teach the model where to look. An earthquake, cyclone and wildfire leave different signatures. If the model learns the wrong examples, it may produce a persuasive map of the wrong phenomenon. Humanitarians then combine the model's estimates with field reports, road conditions, local knowledge and other satellite products before committing resources.[1][3]

Microsoft's public repository makes the implementation more inspectable than a proprietary dashboard. It describes an application with an interface, queue workers, geospatial processing and cloud-optimized imagery. The published quick start supports a local Docker Compose deployment for development and evaluation. The project specifically warns that its convenient development defaults—including disabled authentication—are not appropriate for a production disaster-response system. That is not a trivial technical footnote when location assessments may be shared across many organisations.[2][4]

The broader opportunity is for smaller organisations to build local response capacity rather than wait for an outside team to hand over a static image. That requires satellite access, trained users, software maintenance and a process for checking damage calls against reality. HASTE offers an implemented tool and documented use; it does not remove the local decisions on which effective disaster assistance depends.[1][2][3]

Report an error or suggest a correction