Google maps its AI stack for disaster warnings

Google Research has published a white paper that assembles the company’s flood, weather, wildfire, earthquake and damage-assessment systems into one operational picture. Co-authored by Yossi Matias and Kate Brandt, the crisis resilience paper explains how Google wants to detect hazards earlier, distribute warnings and support recovery.

Many of the underlying models have been published or made openly available during the past two years. The paper adds deployment details, reported warning times and a clearer account of how forecasts reach public agencies, emergency services and affected communities.

Sparse sensors leave blind spots

Google cites more than 350 major weather-related disasters in 2025 that affected over 110 million people. It estimates annual costs above $2 trillion when indirect losses are included.

Traditional numerical weather models solve physical equations on supercomputers, which makes high-resolution forecasting computationally expensive. Their local accuracy also depends on observations from gauges, radar and weather stations. Many rivers in the Global South lack gauges, while cities often have little historical data about highly localized flash floods.

Machine-learning models can produce forecasts faster and infer patterns across areas with limited observations. Their reliability still depends on training data, regional validation and a distribution system capable of turning model output into timely alerts.

News reports become flood labels

Google developed Groundsource to address missing historical records for urban flash floods. The company used Gemini to process more than five million flood-related news reports covering 20 years, producing 2.6 million event records across more than 150 countries. The resulting dataset is available through the Groundsource announcement.

News archives expand coverage beyond instrumented regions, though they remain an imperfect proxy for sensors. Reporting density, language coverage and publication access can shape which floods appear in the dataset. An absent record does not establish that no flood occurred, so regional bias and label quality require attention when evaluating models trained on the corpus.

Google’s downstream model estimates the probability of a flash flood within the next 24 hours for each 20 km by 20 km urban grid cell. Inputs include a seven-day hindcast, meaning reconstructed meteorological and geophysical conditions leading up to the prediction, plus weather forecasts for the next 24 hours. The grid supports regional warning, while street-level decisions require finer local data. Forecasts are available through Flood Hub.

Five hazards run on different clocks

The systems use different operational measures. Cyclone and flood products forecast future conditions, wildfire tools refresh detected boundaries, and earthquake alerts begin only after sensors detect a rupture. Their reported windows therefore are not directly comparable.

WeatherNext generates ensembles of plausible weather outcomes instead of relying on one deterministic forecast. Google reports that its three-day cyclone forecasts achieved accuracy comparable to two-day forecasts from earlier models, roughly matching a decade of prior operational improvement. The U.S. National Hurricane Center validated the model during the 2025 season.

Combined with regional models, WeatherNext helped the center predict that Hurricane Melissa would intensify to Category 5 while the storm remained at Category 1. According to Google, this was the first accurate prediction of that scale of intensification.

Phones race seismic waves

The Android Earthquake Alerts System turns participating phones into a distributed seismic network. When a phone’s accelerometer detects motion resembling a primary wave, or P wave, it sends location data to Google’s servers. Signals from many devices allow the system to estimate the earthquake’s epicenter and magnitude.

P waves travel faster and usually cause less damage than the later secondary waves associated with stronger shaking. That speed difference creates a brief warning window for people farther from the epicenter. The system provides earthquake early warning rather than earthquake prediction because detection begins after the rupture starts.

During a magnitude-7.2 and magnitude-7.5 doublet in Venezuela in June 2026, Google sent the first alert three seconds after the initial rupture. More than 11 million alerts followed, with warning times ranging from seconds to two minutes depending on location. Performance depends on phone density, network connectivity, processing latency and distance from the epicenter.

FireSat targets a 20-minute view

Google’s satellite-driven wildfire boundary tracking currently covers 34 countries and publishes updates through Search and Maps. The service refreshes fire perimeters every 15 to 20 minutes, helping users distinguish an active boundary from a static incident marker.

The planned FireSat constellation, developed with the nonprofit Earth Fire Alliance and Muon Space, aims to detect fires as small as 5 meters by 5 meters anywhere on Earth every 20 minutes. Three additional satellites launched in July 2026 as part of a planned fleet exceeding 50 spacecraft. Full global detection depends on completing and operating that constellation.

The pieces developers can use

Several datasets, models and APIs described in the paper are available for research or operational integration. Access terms vary, and an openly documented model should not be confused with an unrestricted production API.

The Czech Hydrometeorological Institute has integrated Google’s hydrology framework into Delft-FEWS, a forecasting platform used by water agencies. That deployment offers one example of incorporating the model into an existing operational system instead of replacing an agency’s full forecasting stack.

SKAI remains limited to partners. After Hurricane Melissa, the damage-assessment model scored more than 385,000 buildings. GiveDirectly used those scores to direct cash assistance toward the worst-affected parishes in Jamaica.

Models converge in the alert pipeline

Google presents its previously published models as a connected pipeline: sensors, satellites and reports supply observations; models convert those inputs into forecasts or damage estimates; distribution partners deliver warnings and guide relief decisions. Operational value depends on forecast quality, coverage, delivery latency and the recipient’s ability to act.

Those partners include national meteorological and hydrological services, the World Meteorological Organization, the United Nations Office for Disaster Risk Reduction, GiveDirectly and the Watch Duty wildfire service. Their role explains the paper’s emphasis on deployment and communication alongside model benchmarks.

The paper reports lead times and deployments more consistently than false-alarm rates, missed-event rates, calibration, uptime, regional performance and API limits. Developers and emergency agencies need those measures, along with licensing and support terms, to compare the systems with existing operational tools.

Google identifies generative AI as its next synthesis layer for command centers receiving large volumes of incomplete or conflicting reports. The proposed systems would organize evidence and surface relevant updates, with people retaining responsibility for verification and operational decisions.