Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseResearch Paper · For R&DDeepMind Blog→ original

DeepMind Releases Weather AI Training Data: Rain Detection 30% Better Than NASA Benchmark

DeepMind has opened up the data it used to train WeatherNext 3, its weather AI model: the hourly satellite mosaic SatWeave and the precipitation map PARDIG. For light and moderate rain, PARDIG detects more than 30% better than IMERG-Final, NASA's reference precipitation dataset. The map is built at 15-minute steps and roughly 10 km resolution, even where there are no radars. For developers, it is a ready foundation for flood and drought models.

AI-processed from DeepMind Blog; edited by Hamidun News
DeepMind Releases Weather AI Training Data: Rain Detection 30% Better Than NASA Benchmark
Source: DeepMind Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceDeepMind Blog ↗
Editor-in-Chief
Key metric #1SatWeave updates once an hour, and its archive spans more than ten years
PARDIGarchive from 1999 to 2026, with fresh data arriving about 5 hours later
Key metric #3The data sits in Google Cloud Storage, and PARDIG is also available in Earth Engine

Predicting the weather takes more than a good neural network. You also need to know what is happening over the planet right now and how much rain has actually fallen. That is hard: radars and rain gauges are not everywhere, and satellites each see things in their own way.

Google DeepMind has decided to open up exactly this part of the kitchen. The WeatherNext team has published two datasets that underpin its new weather AI model, WeatherNext 3: SatWeave (an hourly satellite picture of the whole world) and PARDIG (a precipitation map covering the entire planet). The headline figure of the release concerns PARDIG.

It detects light and moderate rain more than 30% better than IMERG-Final, NASA's reference set of precipitation estimates, which many applications rely on. Consider an agricultural insurance app: if the old dataset misses every third rainfall or paints downpours where it was dry, payouts are calculated incorrectly. The new map reduces such misses.

Both collections are available as a historical archive and as a live feed.

SatWeave: a single picture of Earth from many satellites

The problem will be familiar to anyone who has stitched together data from several sources. The planet is photographed by different geostationary satellites: the American GOES-East and GOES-West, the European Meteosat, and Japan's Himawari. Each has its own map projection, file formats, sensors, and imaging schedule: a frame arrives every 10 to 30 minutes. As a result, ready-made global mosaics are scarce, and the ones that exist are usually limited: four channels or fewer, or updates only every few hours.

SatWeave combines everything into a single world map with 17 channels: visible light, water vapor, and infrared. It updates once an hour. For WeatherNext 3 this matters: the model gets a fresh view of forming clouds and storms, rather than relying on analysis data that many models refresh only once every six hours. The hourly feed and an archive more than ten years deep are open. DeepMind also names specific applications: forecasting output from solar power plants and monitoring forest fires.

PARDIG: a precipitation map where there are no radars

The second dataset tackles a different task. Precipitation can be measured accurately with ground radars and a dense network of rain gauges, but these are far from ubiquitous. Over oceans, in mountains, and in many developing countries, coverage is sparse.

PARDIG (Precipitation AI Reanalysis: Densely Inferred GPM) works around this limitation. The model, architecturally similar to WeatherNext 3, learns directly from the sparse measurements of NASA's GPM Core Observatory space radar. Its inputs are SatWeave images and atmospheric fields from the ECMWF centre. The output is a precipitation map of the whole planet at 15-minute steps and roughly 10 km resolution. Fresh data arrives with an average delay of about five hours.

Validation was carried out against three continental radar networks (the US, Europe, and Japan) and more than 36,000 weather stations. According to DeepMind, PARDIG outperforms leading global analogues at every rain intensity level, including IMERG-Final. The 30% figure is measured with the Critical Success Index, which shows how often the hourly forecast matched the radar and penalizes both missed rain and false alarms. The company also notes that the model handles extreme downpours better, which global products often miss. A telling example is the July 2025 flooding in Texas, where PARDIG was compared with the NOAA ground radar and with IMERG-Final.

The PARDIG historical archive covers the period from 1999 to 2026.

What to do with this, and the caveats

For developers and researchers, this is a rare case of a top lab's training data being openly available. The data is hosted in Google Cloud Storage, and the PARDIG historical catalogue is also available in Google Earth Engine. It can be used to train your own flood, landslide, and drought models, especially for regions without dense radar networks. It also suits applied services, from energy to agricultural insurance.

There are caveats too. The resolution of about 10 km is noticeably coarser than ground radar: in the Texas flooding example, the NOAA radar provides roughly 1 km. For local urban tasks, PARDIG is not a replacement for radar but a way to fill the gaps. The claim of reliability "at any point on Earth" remains the authors' expectation: in regions without ground measurements, there is no direct way to verify it. All the comparison figures so far come from DeepMind itself, and the article contains no independent tests. Finally, these are data for geospatial specialists: working with them requires skills in processing satellite rasters.

Still, the step is significant. An hourly-updated satellite mosaic and a global precipitation map that has existed for years are becoming publicly available infrastructure rather than an internal Google resource.

Story

Key Questions Answered

What are the key details on SatWeave: a single picture of Earth from many satellites?

The problem will be familiar to anyone who has stitched together data from several sources. The planet is photographed by different geostationary satellites: the American GOES-East and GOES-West, the European Meteosat, and Japan's Himawari. Each has its own map projection, file formats, sensors, and imaging schedule: a frame arrives every 10 to 30 minutes. As a result, ready-made global mosaics are scarce, and the ones that exist are usually limited: four channels or fewer, or updates only every few hours.

What are the key details on PARDIG: a precipitation map where there are no radars?

The second dataset tackles a different task. Precipitation can be measured accurately with ground radars and a dense network of rain gauges, but these are far from ubiquitous. Over oceans, in mountains, and in many developing countries, coverage is sparse. PARDIG (Precipitation AI Reanalysis: Densely Inferred GPM) works around this limitation. The model, architecturally similar to WeatherNext 3, learns directly from the sparse measurements of NASA's GPM Core Observatory space radar.

What to do with this, and the caveats?

For developers and researchers, this is a rare case of a top lab's training data being openly available. The data is hosted in Google Cloud Storage, and the PARDIG historical catalogue is also available in Google Earth Engine. It can be used to train your own flood, landslide, and drought models, especially for regions without dense radar networks. It also suits applied services, from energy to agricultural insurance. There are caveats too.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…