On November 13, 2026, Google DeepMind released historical records and live feeds for two datasets that support its weather AI, WeatherNext 3: "SatWeave" and "PARDIG." SatWeave integrates observations from multiple weather satellites into a common format, while PARDIG provides near-surface rainfall estimates derived from satellite observations and other inputs. This makes it easier for outside researchers and developers to work on weather AI without building satellite image preprocessing and rainfall training data from scratch. Google DeepMind's WeatherNext team says that, in addition to historical archives of both datasets, continuously updated live data is available.
SatWeave supplies satellite observations such as clouds and water vapor as input to AI models, and PARDIG provides rainfall estimates that can be used to train precipitation forecasting models. Google DeepMind also describes for developers how WeatherNext ingests SatWeave directly and initializes forecasts every hour. However, recording data at fine time intervals is not the same as being able to obtain it in near real time. When using the public data for actual forecasting or disaster monitoring, users need to consider delivery frequency and latency, not just the temporal resolution of the records.
Merging multiple satellites' imagery, and estimating global rainfall from limited precipitation observations
SatWeave merges imagery from multiple geostationary weather satellites, including Himawari, GOES, and Meteosat. The original satellite data is provided by national space and meteorological agencies, but projections, calibration methods, and file formats differ from satellite to satellite, and observation intervals are not uniform, ranging from 10 to 30 minutes. SatWeave reprojects these images onto a common latitude-longitude grid, calibrates them, and stitches them together. The grid spacing is 0.1 degrees, equivalent to roughly 10 km. Detailed specifications are published in the SatWeave developer documentation.
The data comprises 7 spectral bands showing reflected sunlight and 10 bands showing brightness temperature, such as water vapor and thermal infrared. This allows not only cloud distribution but also information related to atmospheric water vapor and temperature to be handled on a common grid. However, the coverage of geostationary weather satellites generally extends only to about 60–65 degrees north and south latitude, and all 17 bands are not available for every region and time. Bands that a satellite does not carry, and data such as reflected sunlight that cannot be captured at night, are missing. In addition, the generational transition of the GOES satellites in 2016–2018 changed which bands are available.
PARDIG, meanwhile, is an AI model that estimates near-surface rainfall from satellite imagery and data describing atmospheric conditions. It was trained on "CORRA," a precipitation estimate product from the primary satellite of the Global Precipitation Measurement (GPM) mission, a joint effort of NASA and JAXA. CORRA combines radar and radiometer observations, but the area the GPM primary satellite can observe at one time is limited to a relatively narrow swath along its orbital track. PARDIG learns the relationship between these rainfall estimates and the continuously available satellite imagery and atmospheric data, and uses it to estimate global rainfall distribution. The model uses a Transformer built on a spherical graph grid. Technical details are explained in the research paper.
PARDIG comes in two versions: "PARDIG-NRT," which supports continuous data delivery, and "PARDIG-LR," which provides a long historical record. NRT takes the high-resolution atmospheric field data "HRES" and SatWeave as inputs, using 11 of the 17 published SatWeave bands. LR, on the other hand, uses the "ERA5" meteorological reanalysis, the "GridSat-B1" satellite imagery dataset, and passive microwave observations from polar-orbiting satellites. Both provide output at 0.1-degree grid spacing and 15-minute intervals. The PARDIG developer documentation summarizes the data structure and usage of each.
Precipitation detection evaluated against ground radar and rain gauges
To examine how well PARDIG's rainfall estimates reproduce actual observations, the research team evaluated performance using ground radar and rain gauge data. These ground observations were not used for model training or calibration and were held out as independent evaluation data. The main radar comparisons covered the contiguous United States, Europe, and Japan for 2016–2023. The rain gauge evaluation used quality-controlled data from more than 36,000 observation sites for 2018–2023. Although the comparison used observations not seen in training, the evaluation itself was carried out by the research team and should be distinguished from independent third-party verification.
Google DeepMind says that, for light to moderate rainfall, hourly precipitation detection skill improved by more than 30% compared with the existing precipitation product "IMERG Final." The metric used was the CSI (Critical Success Index), calculated as the number of correctly detected rainfall events divided by the sum of hits, misses, and false alarms. A CSI closer to 1 indicates higher performance, but results change depending on the rainfall threshold used to define precipitation. It is therefore important to check what rainfall intensity a comparison targets, not just the rate of improvement.
For example, in the paper's comparison for the contiguous United States in 2016–2023, the CSI for rainfall of 1 mm per hour or more was 0.48 for PARDIG-NRT and 0.38 for IMERG Final. Raising the threshold to 4 mm per hour, PARDIG-NRT scored 0.37 and IMERG Final 0.26. PARDIG-NRT came out ahead under both conditions, but the size of the improvement varies with the rainfall intensity being evaluated.
PARDIG-NRT was trained on 2016–2023 data, and the research team also evaluated 2024–2025, which falls outside the training period, using ground radar in the United States and Japan. The authors report finding no performance degradation. However, the research paper by Guy Shalev and colleagues is a preprint posted to arXiv on November 10, 2026, and no third-party replication has been confirmed at this point. The reported performance should be read as the research team's own evaluation results.
Even with 15-minute rainfall data, delivery lags by about 5 hours on average
When using the public data, it is important not to confuse the data's time interval with its actual delivery frequency and latency. PARDIG-NRT provides rainfall estimates at 15-minute intervals, but live data is updated every 6 hours, and the delay from the target time to delivery averages about 5 hours. In other words, even data that represents rainfall conditions every 15 minutes cannot be obtained 15 minutes later. SatWeave's live data, by contrast, is recorded at 1-hour intervals, updated hourly, and delivered with a delay of about 1 hour.
Based on the official developer documentation, the record period, time interval, update frequency, and delivery delay of each dataset compare as follows.
| Data and usage mode | Record period | Data time interval | Update frequency | Delivery delay from target time |
|---|---|---|---|---|
| SatWeave historical record | 2016–2025 | 30 minutes | Historical record | — |
| SatWeave live feed | 2026 onward | 1 hour | Hourly | About 1 hour |
| PARDIG-NRT historical archive | 2016–2025 | 15 minutes | Historical record | — |
| PARDIG-NRT live / operational record | 2026 onward | 15 minutes | Every 6 hours | About 2–8 hours, average about 5 hours |
| PARDIG-LR historical record | 1999–2023 | 15 minutes | Historical record | — |
Note: This table is organized based on the specifications of data published on Google Cloud Storage and is not the result of measuring actual retrieval speeds. "—" indicates that live-delivery delay does not apply because the data is a historical record. On Google Earth Engine, static archives were available at the time of release, with PARDIG-NRT covering January 2016 to October 2026 and PARDIG-LR covering 1999–2023.
Both datasets are released under the Creative Commons "CC BY 4.0" license and can be retrieved in Zarr format from Google Cloud Storage. SatWeave can be used with a Google Cloud account without submitting an access request form. PARDIG's historical records can also be retrieved from Google Earth Engine. Researchers and developers can choose a retrieval method according to the period they need and their environment, but note that update status differs between live feeds and static archives.
When analyzing long-term changes in rainfall, what matters is not only the length of the record but also how observing conditions changed over that time. PARDIG-LR and PARDIG-NRT use different inputs, and the satellite constellations also changed over the years. Connecting the two series or comparing data from different periods could therefore introduce discontinuities unrelated to actual climate change. The paper evaluates the consistency of data from 2008 onward, but cautions about changes in rainfall distribution before then. To interpret long-term trends as the effect of climate change, the influence of changes in the observing system must be separated out.
There are also limits on estimation accuracy by region and precipitation type. Regions poleward of about 67 degrees north and south fall outside the training range based on GPM observations. Accuracy is also limited for mountain precipitation, winter snowfall, and drizzle. In addition, the deterministic rainfall intensity provided at 15-minute intervals has an upper cap of 75 mm per hour. PARDIG is experimental AI-estimated data at the research stage, and for decisions involving life or property, official information issued by national meteorological agencies and local authorities should take priority.
Google DeepMind hopes the released datasets will be used in research and countermeasures for floods, landslides, droughts, and similar hazards. With standardized satellite observation data and rainfall estimates available, researchers can spend less effort preparing basic data and focus on model development and performance validation suited to their own uses. Whether the data is suitable for actual operations, however, depends on the observing conditions of the target region, the required temporal resolution, and delivery latency. Going forward, independent performance evaluations that account for these constraints, and validation of effectiveness in specific applications, will be important.