DeepMind’s Open Rain Data: 5 Hours vs NASA’s 3.5 Months

Google DeepMind on 13 November released PARDIG, an AI-built global precipitation dataset that it says is updated with an average latency of about five hours, and opened its live feed and a 1999 to 2026 historical archive on Google Cloud Storage under a CC BY 4.0 licence. The benchmark DeepMind compares it with, NASA’s research-quality IMERG Final Run, usually has a latency of about 3.5 months, according to NASA’s GPM data page.

By our arithmetic, five hours against about 3.5 months makes PARDIG’s average roughly 500 times sooner than IMERG Final. DeepMind’s paper also reports higher rain-detection skill than IMERG Final against ground radar in the US, Europe and Japan. NASA’s own near-real-time IMERG Early Run arrives in about 4 hours, so the latency gap is with the Final Run, not with NASA’s fastest product.

Business Pill · Data quality explained

An evergreen Business Engineer explainer on data quality: fewer, better examples. It teaches the general idea only and does not discuss PARDIG or weather data.

The key insight: As we read it, the release moves research-grade rain estimates from a months-long wait to a same-day feed: PARDIG averages about five hours against about 3.5 months for NASA’s IMERG Final, roughly 500 times sooner by our arithmetic, and DeepMind’s paper reports a CSI of 0.37 against 0.26 at 4 mm per hour on US radar.

What DeepMind Released

The 13 November post from DeepMind’s WeatherNext team releases two datasets, PARDIG and SatWeave, and calls both “key to the advances in our recently released WeatherNext 3 model.” DeepMind says it is making historical archives and live feeds of both available.

PARDIG stands for Precipitation AI Reanalysis: Densely Inferred GPM. According to the post, it is trained directly on sparse spaceborne radar observations from the NASA Global Precipitation Measurement (GPM) Core Observatory, using SatWeave imagery and atmospheric fields from ECMWF as inputs, and produces a global precipitation map at 15-minute temporal resolution and roughly 10-kilometer resolution.

The paper, posted to arXiv on 10 November, says the near-real-time version is delivered with an end-to-end latency of 2 to 8 hours from observation time, and that PARDIG never uses ground-based radar or rain gauge observations at any stage. The datasets are openly available under a CC BY 4.0 license, the paper says.

SatWeave is the satellite side. The post says it blends geostationary imagery, naming GOES-East and GOES-West, Meteosat 0 and Meteosat Indian Ocean, and Himawari, into a global mosaic across 17 visible, water vapor, and infrared channels that updates hourly, with a historical archive spanning over a decade.

PARDIG’s historical catalogs are also in Google Earth Engine, where the catalog entry for the long-record version lists data from 1 January 1999 to 31 December 2023 at a 15-minute cadence and describes PARDIG as an experimental dataset.

Typical latency from observation to data, in hours: DeepMind’s PARDIG averages about five hours (2 to 8
Typical latency from observation to data, in hours: DeepMind’s PARDIG averages about five hours (2 to 8 hours in its paper), against about 3.5 months for NASA’s IMERG Final Run; NASA’s IMERG Early Run, at about 4 hours, is the fastest. Latencies from NASA’s GPM IMERG data page, DeepMind’s 13 November 2026 post and the PARDIG paper (ERA5T from the paper); conversion to hours and the ratio of about 511 are our arithmetic.

Five Hours Against 3.5 Months

NASA’s data page gives the reason for IMERG Final’s wait: the Final Run uses a monthly rain-gauge analysis and revised retrievals, and “The pacing items are the GPCC and ERA-5, usually giving a latency of about 3.5 months.” The paper uses the same figure, describing IMERG Final as released with a delay of about 3.5 months.

About 3.5 months is about 2,557 hours, by our arithmetic (3.5 months of 30.44 days). Against PARDIG’s average of about five hours, that is roughly 511 times sooner. NASA’s download list gives a second figure for the Final Run, roughly 2.5 months; on that basis the ratio is about 365 times, by our arithmetic.

NASA’s faster products narrow the gap. The data page says the IMERG Early Run provides the lowest latency available, 4 hours, and that the Late Run is computed about 14 hours after observation time. NASA’s own advice is to use the Early or Late Run for low latency applications, and the Final Run for research purposes where latency is not a concern.

So PARDIG’s pitch, as DeepMind and its paper set it out, is research-grade skill at near-real-time speed: the paper compares it with IMERG Final as the leading retrospective satellite benchmark and with IMERG Early as an operational near-real-time counterpart.

How DeepMind Measured the Accuracy Claim

The paper scores every product with the Critical Success Index (CSI), hits divided by hits plus misses plus false alarms, against three ground radar networks: MRMS over the contiguous US, EURADCLIM over Europe and JMA Radar-AMeDAS over Japan, for 2016 to 2023. DeepMind’s post adds an evaluation against over 36,000 weather stations worldwide.

Against US radar, the paper’s table gives PARDIG’s near-real-time version a CSI of 0.48 at 1 mm per hour, against 0.38 for IMERG Final; 0.37 at 4 mm per hour, against 0.26; and 0.21 at 15 mm per hour, against 0.06. That is about 26% higher, about 42% higher and 3.5 times, by our arithmetic. At 4 mm per hour, IMERG Early scores 0.22.

DeepMind’s post puts it as an improvement in hourly precipitation detection skill over IMERG-Final of “more than 30% on light-to-moderate precipitation.” The paper says the largest improvements occur at extreme intensities.

The paper also reports a weakness. Over the European Alps it says PARDIG exhibits an exaggerated seasonal amplitude, underestimating wintertime precipitation relative to summertime precipitation, and it offers as one hypothesis the GPM radar’s reduced sensitivity to snowfall.

Grouped bar chart of Critical Success Index against US MRMS radar, 2016 to 2023: at 1 mm per hour PARDIG 0.48, IMERG Final 0.38, IMERG Early 0.34; at 4 mm per hour 0.37, 0.26 and 0.22; at 15 mm per hour 0.21, 0.06 and 0.05
Hourly rain-detection skill (Critical Success Index, higher is better) against MRMS ground radar over the contiguous US, 2016 to 2023, for PARDIG’s near-real-time version and NASA’s IMERG Final and Early. Values from Table A.10.1 of DeepMind’s PARDIG paper, the developer’s own evaluation.

Why Rain Data Is AI Infrastructure

The paper calls precipitation “a primary driver for hydrological modeling, agricultural planning, and hazard management.” DeepMind’s post says it wants to work with the community on models for floods, landslides, and drought, and on SatWeave uses such as solar power and wildfire monitoring.

The datasets also feed Google’s own forecasts. The paper says WeatherNext 3 was trained with two precipitation output heads, one against IMERG Final and one against PARDIG, and that at 6-hour lead times the PARDIG head’s CRPS error against US radar is approximately 15% lower than the IMERG head’s.

Google’s WeatherNext developer page describes the layer above the data: it offers forecast data across Google Cloud Storage, Earth Engine and BigQuery, and custom inference that generates tailored forecasts on demand using dedicated accelerators on Google Cloud. We covered an earlier piece of that stack in Google’s Earth AI agent, and DeepMind’s Kohli on a phase transition in AI science.

The Structural Read

As we read it, the comparison DeepMind chose matters: against IMERG Final the latency gap is roughly 500 times, by our arithmetic, but NASA’s IMERG Early Run arrives in about 4 hours, so against Early the paper’s case rests on skill, a CSI of 0.37 against 0.22 at 4 mm per hour on US radar.

As we read it, the open data also feeds Google’s own forecasting stack: the paper says the PARDIG-trained WeatherNext 3 head had approximately 15% lower CRPS than the IMERG-trained head at 6-hour lead times, and Google’s developer page offers custom inference on dedicated accelerators on Google Cloud.

Google DeepMind, 13 November 2026

“Because PARDIG learns directly from satellites rather than relying solely on ground-based measurements, we expect it to have reliable accuracy everywhere on Earth, even in areas where these measurements are not available.”

Three Implications

FOR FLOOD, DROUGHT AND FARM MODELS The paper calls precipitation a primary driver for hydrological modeling, agricultural planning, and hazard management; PARDIG’s live feed and 1999 to 2026 archive are under CC BY 4.0.

FOR USERS OF NASA’S IMERG NASA advises the Early or Late Run for low latency and the Final Run for research; the PARDIG paper reports higher CSI than both IMERG Early and IMERG Final against radar in the US, Europe and Japan.

RELATED Google’s Earth AI agent covers an earlier piece of Google’s Earth AI stack.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework says: “Every technology stack has layers. Some are scarce, some are abundant.” Its rule: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, Google opened the data layer under CC BY 4.0 while its developer page offers custom inference on Google Cloud accelerators, a split that matches the framework’s pattern. The framework does not discuss weather data, PARDIG or WeatherNext.

What Is Not Established

The accuracy figures are from DeepMind’s own paper, a first version on arXiv, and its blog post. We did not download the datasets or run an evaluation of our own, and we did not contact Google DeepMind or NASA.

The latency comparison uses averages and typical figures as each source states them: about five hours for PARDIG in DeepMind’s post (2 to 8 hours in the paper), and about 3.5 months for IMERG Final on NASA’s page, which also lists roughly 2.5 months. The paper calls PARDIG’s estimates experimental and says they are not official severe weather warnings.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Open vs Closed Meta-Framework

The Business Engineer framework for deciding which layer of a technology stack to open and which to keep closed.

Read the Map of AI →

The Bottom Line

DeepMind’s PARDIG arrives with an average latency of about five hours, against about 3.5 months for NASA’s IMERG Final, roughly 500 times sooner by our arithmetic, and its paper reports a CSI of 0.37 against IMERG Final’s 0.26 at 4 mm per hour on US radar.

The live feed and the 1999 to 2026 archive are openly available under CC BY 4.0, and NASA’s IMERG Early Run, at about 4 hours, remains the faster NASA product.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. We read Google DeepMind’s 13 November 2026 blog post in full, and the PARDIG paper on arXiv (first version, submitted 10 November 2026): its abstract, dataset, baseline, metric and results sections and its appendix table of Critical Success Index values. NASA’s latency figures for IMERG come from NASA’s GPM IMERG data page; the Earth Engine dates and licence come from Google’s catalog entry for PARDIG LR, and the forecast-access details from Google’s WeatherNext developer page.

A Tech Times article on 16 November 2026 pointed us to the release. The conversion of months to hours, the latency ratios and the percentage gaps between CSI values are our arithmetic. The accuracy results are DeepMind’s own evaluation; we did not download or test the datasets, and we did not contact Google DeepMind or NASA.

Sources: Google DeepMind blog, 13 Nov 2026 · PARDIG paper, arXiv:2611.13941 · NASA GPM IMERG data page · Google Earth Engine catalog, PARDIG LR · Google WeatherNext developer page

Scroll to Top
FourWeekMBA