Back to Bigger Picture

BIGGER PICTURE Disaster preparedness

Finding damaged buildings before every roof is checked

After a disaster, satellite images arrive faster than experts can inspect every building. SKAI helps turn those images into an initial map of where to look.

AI-generated illustration: Conceptual layered aerial view comparing intact and storm-damaged roofs.
Layered aerial view comparing intact and storm-damaged roofs.AI-generated illustration of the idea, not a photograph of the actual people, equipment or event.
Evaluation
Nine emergencies
Initial assessment
Sixfold shorter analysis time, on average
Area covered
Sevenfold expansion, on average

How the idea works

  1. Locate structures

    Image models outline buildings in the affected area.

  2. Teach event-specific damage

    Analysts label examples to adapt the damage model.

  3. Map priority areas

    The model extends assessment across the wider scene.

Every roof is a question

Following an earthquake or storm, humanitarian analysts compare before-and-after satellite images to judge building damage. A city can contain tens of thousands of structures. Those individual decisions add up precisely when responders need an overview quickly: which neighbourhoods may need shelter, where access may be difficult, and where detailed inspection should start.

Sources: [1][3]

Find the buildings, then learn this disaster

The UN workflow combines three tools. Open Buildings identifies structures; a general SKAI model highlights likely damage; analysts label a sample of buildings to adapt a second model to the particular event. That adapted model examines the wider area. The experts’ local judgments therefore guide the larger assessment rather than disappearing from it.

Sources: [1][2][3]

The useful result was an earlier overview

In an evaluation reported by UN Global Pulse in 2024, nine emergencies showed an average sevenfold increase in assessed area and a sixfold reduction in time to directional findings, bringing that analysis below a day. These are workflow outcomes. They mean responders can obtain an initial geographic picture sooner and repeat assessments as a crisis evolves.

Sources: [1]

The clock also depends on the satellite

The timing starts with usable imagery. Obtaining suitable before-and-after views can itself take days, and cloud cover may prevent optical analysis. Misaligned images, different lighting and unfamiliar building types also complicate the task. SKAI is an aid to expert damage mapping, not a direct detector of trapped people or a replacement for ground verification.

Sources: [1][3][2]

The original sources

  1. UN Global PulseAI from Google Research and UN boosts humanitarian disaster response
  2. Google ResearchSKAI open-source project
  3. Google ResearchMachine Learning-based Damage Assessment for Disaster Relief

What to explore next

Forecasting a downpour without blurring it awayHelping a new support agent find the expert’s next move
Explore all 99 applications