# Google opens precipitation datasets

> Rainfall estimates have measurement and timing limits.

By BIGCHANGE.AI Editorial

Published: 2026-11-13T15:43:39.152Z
Updated: 2026-11-13T15:43:39.152Z
Canonical: https://bigchange.ai/blog/google-pardig-satweave-rainfall-datasets

![Illustration of a precipitation satellite above Earth.](https://bigchange.ai/api/media/file/gpm-v1.png)
AI-generated illustration for BIG CHANGE.

Google's November release opens PARDIG precipitation estimates and SatWeave satellite mosaics.

## The big change

- PARDIG-NRT estimates past precipitation at 15-minute intervals, not future rainfall. Publication occurs every six hours with roughly two to eight hours' latency, rather than instantly.
- Satellite measurements need interpretation. NASA explains that radar reflections and microwave radiation are converted into precipitation information by algorithms; the sensors do not directly count rainfall reaching every location.
- Data frequency and delivery delay answer different questions. NASA's existing IMERG products illustrate why a researcher studying a completed storm may choose a different version from someone needing recent observations.

## Learning from incomplete observations

The [November 10 preprint](https://arxiv.org/html/2611.13941v1) describes AI learning from sparse GPM radar/radiometer precipitation targets, then combining atmospheric fields and SatWeave imagery to estimate precipitation globally.

[NASA's instrument account](https://gpm.nasa.gov/missions/GPM/core-observatory) explains why those satellite observations are intermittent. The GPM Core Observatory circles Earth roughly sixteen times a day, sampling precipitation at different times. Its radar observes a strip about 245 kilometres wide; its microwave imager covers a wider swath. A passing strip is not continuous observation of every location.

The instruments also measure different physical signals. [NASA describes](https://gpm.nasa.gov/science/precipitation-algorithms) radar as sending a signal and examining its reflection from precipitation particles. The microwave imager passively measures natural thermal radiation from the scene, including clouds and the surface. Algorithms transform those signals into precipitation information, and ground observations provide a way to check the retrievals. This explains why a rainfall estimate is a processed interpretation of measurements, rather than a rain gauge everywhere.

## What the reported comparison establishes

Authors report better rain detection than competing products using US/European/Japanese radar as reference. Their 2016–2023 evaluation overlaps training years, but ground observations were not training targets; later US/Japan radar tests cover 2024–2025. The preprint does not establish universal local accuracy or improved flood response.

[Google's release notes](https://developers.google.com/weathernext/release-notes) list CC BY 4.0 data on Cloud Storage. The paper flags lower early-period means before 2008 in the 1999–2023 archive.

## Choosing data for a real question

NASA's separate [IMERG guidance](https://gpm.nasa.gov/data/imerg) gives a concrete example of the tradeoff between delivery speed and processing. Its Early version propagates observations forward; the Late version can also use later observations to interpolate backward. Waiting allows additional satellite transmissions to arrive. The Final version incorporates monthly rain-gauge information and is intended for research where the delay is acceptable.

For a completed storm analysis, later corrections may be worth waiting for. An operational user cannot use information that has not yet arrived. NASA also cautions that some products express a rainfall rate in millimetres per hour, while others give accumulated millimetres. Comparing them requires matching units and time windows before judging whether two maps agree.

These established product distinctions explain the questions a data user should ask. They are not an independent validation of Google's new estimates. A fine time grid alone establishes neither immediate availability nor reliable performance at a particular site.

## Sources

- [WeatherNext release notes](https://developers.google.com/weathernext/release-notes) — Dataset release and access.
- [PARDIG research preprint](https://arxiv.org/html/2611.13941v1) — Methods and evaluation.
- [NASA GPM Core Observatory](https://gpm.nasa.gov/missions/GPM/core-observatory) — Satellite instruments.
- [NASA precipitation algorithms](https://gpm.nasa.gov/science/precipitation-algorithms) — How observations become estimates.
- [NASA IMERG](https://gpm.nasa.gov/data/imerg) — Existing products and delivery.