Google opens the datasets behind its WeatherNext 3 forecasts

By Sophie Lindqvist

Google’s PARDIG dataset improves hourly detection skill for light-to-moderate precipitation by more than 30% over NASA’s IMERG-Final, according to the company’s evaluations. Google is releasing PARDIG and SatWeave, the two foundational datasets used in its recently released WeatherNext 3 model, as live feeds and historical archives.

In Google’s account, the core constraint is simple: weather models need current, consistent observations, while each satellite and ground network covers a different part of Earth. The company said the open data can give researchers and developers a shared base for weather, hydrology and climate applications, including work in regions with sparse radar coverage.

What does SatWeave add to WeatherNext 3?

SatWeave combines geostationary imagery from five satellites: GOES-East, GOES-West, Meteosat 0, Meteosat Indian Ocean and Himawari. Google described the product as “a live view of developing clouds and storms,” with updates every hour.

Google said the five satellites do not naturally align because they use different map projections, file formats, sensors and scan schedules. Those schedules range from 10 to 30 minutes, and the mismatch makes a coherent global map difficult.

Many existing global composites respond to that mismatch by dropping incompatible inputs, offering only four or fewer channels, or updating every few hours, according to Google. SatWeave instead unifies the imagery and refreshes its global mosaic every hour.

Google said SatWeave allows WeatherNext 3 to produce new global predictions every hour. It contrasted that cadence with models initialized on analysis data refreshed every six hours.

The company is also releasing a historical SatWeave archive spanning more than a decade. It presented the archive as material for testing models and the hourly live feed as current input for operational forecasts.

WeatherNext 3 already uses SatWeave for medium-range weather forecasts. Google said it wants outside researchers and developers to explore other applications, including solar-power forecasting and wildfire monitoring. In practical terms, SatWeave supplies a unified, current view that can support forecasts more often than six-hourly analysis cycles.

How does PARDIG estimate rain?

PARDIG stands for Precipitation AI Reanalysis: Densely Inferred GPM. Google built it on an AI architecture similar to WeatherNext 3 and trained it directly on sparse spaceborne radar observations from the NASA Global Precipitation Measurement (GPM) Core Observatory.

PARDIG uses SatWeave imagery and atmospheric fields from ECMWF as additional inputs. Google said the result is a global precipitation map at 15-minute temporal resolution and roughly 10-kilometer resolution.

Those two resolution figures describe different properties. The 15-minute interval sets the map’s time steps, while roughly 10 kilometers sets its spatial granularity; Google reports an average output latency of about five hours.

Google said ground radars and dense rain-gauge networks provide high-quality precipitation data, but they are not always available or well distributed. That gap matters because accurate rain estimates affect food and water supplies, flood risk and drought.

PARDIG learns from satellite observations rather than relying only on ground measurements. Google said it expects the dataset to have “reliable accuracy everywhere on Earth,” including regions without dense instrument coverage.

The company used PARDIG to train WeatherNext 3’s precipitation forecasts. In practical terms, it gives the model a global rain estimate where ground radar coverage is thin.

How did Google test PARDIG?

Google evaluated PARDIG against three continental radar networks across the US, Europe and Japan, as well as more than 36,000 weather stations worldwide. The company said PARDIG outperformed leading global baselines across all analyzed rain rates.

The comparison included NASA’s IMERG-Final, which Google called a research-quality dataset of precipitation estimates used in many applications. In hourly precipitation detection tests, PARDIG improved light-to-moderate detection skill by more than 30% over IMERG-Final.

Google also reported that PARDIG was particularly effective at capturing extreme precipitation. The company said existing global products can struggle to detect those events.

Those results cover regional radar networks and a much larger set of station checks, according to Google. The clearest published advantage is stronger light-to-moderate rain detection, with the company also reporting particular strength at extreme precipitation.

Where can researchers get the data?

Google made the live and historical files for both datasets available on Google Cloud Storage. PARDIG’s historical catalogs are also available in Google Earth Engine, giving researchers a second place to work with the long precipitation record.

Google called the release the latest addition to its Earth AI portfolio. The company said it hopes researchers can use the materials to connect weather observations with practical applications and build new tools.

The PARDIG archive covers 1999 through 2026, alongside a live feed. SatWeave offers an hourly live feed and a historical archive extending more than a decade.

Google said WeatherNext 3 used PARDIG for medium-range weather forecasting, but the company wants the wider community to explore models for floods, landslides and drought. It identified regions without dense ground radar networks as a particular target.

SatWeave’s wider possibilities include solar power and wildfire monitoring, according to Google. In practical terms, the release combines model-building material with live feeds designed for current applications.

Who built and supplied the data?

The work spans several Google teams. Google said teams across Google DeepMind and Google Research developed the datasets in collaboration with Google Cloud and Google Geo.

The company thanked ECMWF, NASA, JAXA, NOAA, EUMETSAT, KMA, JMA, KNMI and Synoptic Data for their open satellite programs and datasets. Google said those shared programs and data make the work possible.

The release shows how much SatWeave and PARDIG depend on public infrastructure rather than one company’s sensors alone. The next test is whether outside researchers can turn the open feeds into better flood, landslide and drought forecasts in regions without dense radar networks.