**PARDIG** (*Precipitation AI Reanalysis - Densely Inferred GPM* ) is an
experimental dataset of global, AI-driven precipitation estimates developed by
[Google DeepMind](https://deepmind.google/) and
[Google Research](https://research.google/). PARDIG is trained directly on
precipitation retrievals from the Combined Radar-Radiometer Algorithm (CORRA)
aboard the NASA/JAXA Global Precipitation Measurement (GPM) Core Observatory.
Relying on satellite observations and ECMWF atmospheric fields as inputs,
PARDIG produces global precipitation rate estimates **every 15 minutes** at
**0.1° (\~10 km) spatial resolution** from **1999 to the present**.

PARDIG is also used as one of the precipitation training targets for
[WeatherNext 3](https://developers.google.com/weathernext/guides/models.md.txt) (`experimental_tp_1hr`). For
technical details and validation benchmarks, see the
[research paper](https://arxiv.org/abs/2611.13941).


[Access on Cloud Storage](https://developers.google.com/weathernext/guides/pardig.md.txt#access-on-google-cloud-storage)
[Explore in Earth Engine](https://developers.google.com/weathernext/guides/pardig.md.txt#access-on-earth-engine)
[Explore Variables](https://developers.google.com/weathernext/guides/pardig.md.txt#supported-variables)

## At a glance

PARDIG is released in two variants: **PARDIG-NRT** (Near-Real-Time) from 2016
to the present, and **PARDIG-LR** (Long-Record) from 1999 to 2023.

| Attribute | PARDIG-NRT (Near-Real-Time) | PARDIG-LR (Long-Record) |
| Temporal Coverage | **2016 to present** (`2016-01-01T07:45:00` UTC onward; continuously updated on Cloud Storage) | **1999-2023** (`1999-01-01T00:00:00` to `2023-12-31T23:45:00` UTC; static archive) |
| Spatial Coverage | Global (90°S to 90°N, 0° to 360°E) | Global (90°S to 90°N, 0° to 360°E) |
| Spatial Resolution | 0.1° (\~10 km) (1801 × 3600 grid) | 0.1° (\~10 km) (1801 × 3600 grid) |
| Temporal Resolution | 15 minutes (`:00`, `:15`, `:30`, `:45` UTC) | 15 minutes (`:00`, `:15`, `:30`, `:45` UTC) |
| Update Frequency \& Latency | Updated every 6 hours on Cloud Storage (\~5 hours average, \~2-8 hours effective latency); static archive on Earth Engine | Static historical archive (no real-time updates) |
| Training Target | NASA/JAXA GPM Core Observatory Combined Radar-Radiometer Algorithm (CORRA) precipitation swaths ||
| Model Inputs | ECMWF HRES (0.1°) + [SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt) geostationary satellite mosaic (0.1°, 11 channels) | ECMWF ERA5 (0.25°) + NOAA GridSat-B1 + polar-orbiting (LEO) passive microwave sounders (AMSU-A, MHS, ATMS) |
| Available Surfaces | - **[Google Cloud Storage (Zarr)](https://developers.google.com/weathernext/guides/pardig.md.txt#access-on-google-cloud-storage):** Full 2016 to present archive + live 6-hourly operational stream - **[Earth Engine](https://developers.google.com/weathernext/guides/pardig.md.txt#access-on-earth-engine):** Static archive (January 2016 through October 2026 at launch) | - **[Google Cloud Storage (Zarr)](https://developers.google.com/weathernext/guides/pardig.md.txt#access-on-google-cloud-storage):** Full 1999-2023 archive - **[Earth Engine](https://developers.google.com/weathernext/guides/pardig.md.txt#access-on-earth-engine):** Full 1999-2023 archive |
| License \& Access | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · Publicly accessible with any Google Cloud account (no access request form required) ||
|---|---|---|

### Choosing between PARDIG-NRT and PARDIG-LR

- **Prefer PARDIG-NRT for operational use and any analysis from 2016 onward:** If you need low-latency operational data, or if your study only requires data from 2016 onward (including the 2016-2023 overlap), use PARDIG-NRT, which ingests higher-resolution `0.1°` ECMWF HRES and [SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt) inputs.
- **Multi-decadal historical records (1999 onward):** If you need a longer record extending back to 1999, you can explore PARDIG-LR (`1999-2023`) on its own or a combination of PARDIG-LR and PARDIG-NRT, keeping in mind the input-resolution differences and historical observing-system transitions described in [Limitations and caveats](https://developers.google.com/weathernext/guides/pardig.md.txt#limitations-and-caveats).

> [!NOTE]
> **Update schedule and effective latency (\~2-8 hours):** Live operational PARDIG-NRT frames on Google Cloud Storage (`gs://pardig_1/NRT/zarr/2026_to_present/`) are published in batches every **6 hours** following each ECMWF HRES analysis cycle (`00`, `06`, `12`, and `18 UTC`). Once a 6-hourly HRES cycle arrives (\~6 hours after initialization) alongside the surrounding [SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt) mosaics, PARDIG-NRT generates **6 consecutive hours (24 frames)** of 15-minute estimates at once. Because all 24 frames in a 6-hour batch are published around the same time, the newest 15-minute frame has an **effective latency of \~2 hours** at publication while the earliest frame in that batch is **\~7.5-8 hours** old, averaging **\~5 hours** overall.

## Key features

- **High accuracy at low latency:** Bridges the gap between low-latency operational precipitation products and gauge-calibrated research archives that lag real time by several months, delivering reanalysis-grade global estimates on Cloud Storage with \~2-8 hours effective latency.
- **Learned from spaceborne radar and radiometer observations:** Trains directly on sparse NASA/JAXA GPM CORRA swaths paired with [SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt) (or historical satellite archives) and ECMWF atmospheric fields to infer dense global precipitation at `0.1°` (\~10 km) and 15-minute resolution.
- **Consistent coverage over oceans and data-sparse regions:** Because PARDIG learns from spaceborne sensors rather than ground-based rain gauges, it maintains consistent skill over oceans, the tropics, and land regions without dense ground radar or gauge networks.
- **Validated against continental radar and global station networks:** Evaluated against three independent regional ground-based radar networks across the United States (MRMS), Europe (EURADCLIM), and Japan (JMA Radar-AMeDAS), as well as over 36,000 weather stations worldwide.
- **Forecast model training target:** Used to train [WeatherNext 3](https://developers.google.com/weathernext/guides/models.md.txt)'s experimental precipitation head (`experimental_tp_1hr`), achieving \~15% lower probabilistic precipitation forecast error (CRPS) than the IMERG-trained head and \~30% lower error than the reanalysis-trained head and WeatherNext 2 at early lead times when evaluated against independent ground radar.

*** ** * ** ***

## Supported variables and grid schema

Both PARDIG-NRT and PARDIG-LR provide a single data variable on a uniform global
`0.1°` (\~10 km) latitude-longitude grid at 15-minute intervals:

| Variable Name | Description | Data Type | Units |
|---|---|---|---|
| `precipitation` | Instantaneous precipitation rate at the valid timestamp (`step_type: instant`) | `float32` | m/h |

> [!NOTE]
> **Units (`m/h`):** `precipitation` in PARDIG represents the **instantaneous precipitation rate** at each 15-minute valid timestamp in **meters per hour (`m/h`, stored in Zarr metadata as
> `m h**-1`)** . Multiply values by **1,000** to convert to millimeters per hour (`mm/h`). Note that in [WeatherNext 3](https://developers.google.com/weathernext/guides/models.md.txt#supported-variables) forecasts, `experimental_tp_1hr` instead reports 1-hour accumulated depth in meters (`m`).

### Coordinate dimensions

| Coordinate | Shape \& Type | Grid Convention |
|---|---|---|
| `time` | `datetime64[ns]` | Valid timestamp in UTC at 15-minute intervals (`:00`, `:15`, `:30`, `:45`). Stored as a 1D dimension in multi-year Zarr archives and as a scalar (0D) coordinate in individual `2026_to_present/` frames. |
| `latitude` | `1801` (`float64`) | Global 0.1° grid ordered south-to-north from `-90.0°` to `90.0°` (`degrees_north`) in Zarr stores, and north-to-south (`90.0°` to `-90.0°`) in Earth Engine. |
| `longitude` | `3600` (`float64`) | Global 0.1° grid in **0° to 360°** convention (`0.0°` to `359.9°`, `degrees_east`) in Zarr stores (convert `-180°..180°` longitudes using `lon_360 = lon % 360`), and `-180.0°` to `179.9°` (`EPSG:4326`) in Earth Engine. |

*** ** * ** ***

## Accessing PARDIG data

> [!NOTE]
> **Open access:** Unlike real-time operational forecast feeds, PARDIG does **not** require submitting a data access request form. Both the Google Cloud Storage Zarr stores and the Earth Engine Data Catalog collections are publicly accessible to any authenticated Google Cloud account under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

### Access on Google Cloud Storage (Zarr)

Google Cloud Storage hosts the complete PARDIG-LR and PARDIG-NRT archives as
well as the live operational stream in [Zarr v3](https://zarr.dev/) format under
the `gs://pardig_1/` bucket.

#### Directory structure and chunk layouts

    gs://pardig_1/
    ├── LR/
    │   ├── zarr/
    │   │   └── 1999_to_2023/
    │   │       └── predictions.zarr           # Spatial access: chunked (4, 1801, 3600)
    │   └── zarr_rechunked_as_timeseries/
    │       └── 1999_to_2023/
    │           └── predictions.zarr           # Time-series access: chunked (time: -1, 5, 5)
    └── NRT/
        ├── zarr/
        │   ├── 2016_to_2025/
        │   │   └── predictions.zarr           # Spatial access: chunked (4, 1801, 3600); starts 2016-01-01T07:45:00
        │   └── 2026_to_present/               # Live operational 15-minute frames (updated every 6 hours)
        │       ├── <YYYYMMDDTHHMMSS>/
        │       │   └── predictions.zarr       # Single 15-minute frame · 2D (1801, 3600)
        │       └── ...
        └── zarr_rechunked_as_timeseries/
            └── 2016_to_2025/
                └── predictions.zarr           # Time-series access: chunked (time: -1, 5, 5)

Choose the Zarr directory layout that matches your access pattern:

| Directory | Chunk Shape `(time, lat, lon)` | Recommended Access Pattern |
|---|---|---|
| `zarr/` | `(4, 1801, 3600)` for archives; `(1801, 3600)` for `2026_to_present/` | Reading global or regional spatial maps at specific timestamps or short time windows, and querying live `2026_to_present/` frames. |
| `zarr_rechunked_as_timeseries/` | `(time: -1, 5, 5)` (full time axis per `0.5° × 0.5°` tile) | Extracting multi-year 15-minute time series at individual coordinates, weather stations, or small bounding boxes across the historical archives (`1999_to_2023` or `2016_to_2025`). |

#### Common query recipes

Install the required Python libraries:

```bash
pip install xarray zarr obstore
```

##### 1. Open the multi-year archive and inspect the schema

Open the historical PARDIG-NRT (`2016_to_2025`, starting at
`2016-01-01T07:45:00` UTC) or PARDIG-LR (`1999_to_2023`) Zarr v3 store lazily
using `obstore` and `xarray`:

<br />

**Python**

<br />

```python
import obstore
import xarray as xr
import zarr

# Open the PARDIG-NRT multi-year Zarr archive on Google Cloud Storage
gcs_store = obstore.store.GCSStore(
    bucket="pardig_1",
    prefix="NRT/zarr/2016_to_2025/predictions.zarr",
)
ds = xr.open_zarr(zarr.storage.ObjectStore(gcs_store), chunks={})
print(ds)
```

Expected dataset structure:

    <xarray.Dataset> Size: 9TB
    Dimensions:        (time: 350657, latitude: 1801, longitude: 3600)
    Coordinates:
      * time           (time) datetime64[ns] 3MB 2016-01-01T07:45:00 ... 2025-12-31T23:45:00
      * latitude       (latitude) float64 14kB -90.0 -89.9 -89.8 ... 89.9 90.0
      * longitude      (longitude) float64 29kB 0.0 0.1 0.2 ... 359.7 359.8 359.9
    Data variables:
        precipitation  (time, latitude, longitude) float32 9TB dask.array<chunksize=(4, 1801, 3600), meta=np.ndarray>

<br />

<br />

##### 2. Slice a regional bounding box from a near-real-time frame

Each timestamp directory under `NRT/zarr/2026_to_present/` stores a single
15-minute frame (`predictions.zarr`) with 2D spatial dimensions
`(latitude: 1801, longitude: 3600)` and a scalar `time` coordinate:

<br />

**Python**

<br />

```python
import obstore
import xarray as xr
import zarr

# Open a single near-real-time 15-minute frame from 2026_to_present
gcs_store = obstore.store.GCSStore(
    bucket="pardig_1",
    prefix="NRT/zarr/2026_to_present/20260901T000000/predictions.zarr",
)
ds_nrt = xr.open_zarr(zarr.storage.ObjectStore(gcs_store), chunks={})

# Crop to a regional bounding box (longitudes in 0-360° convention)
# Example: Central United States (lat: 30°N to 45°N, lon: 255°E to 275°E)
region_rate_mm_hr = (
    ds_nrt["precipitation"].sel(
        latitude=slice(30.0, 45.0),
        longitude=slice(255.0, 275.0),
    )
    * 1000.0  # Convert m/h to mm/h
)
print(region_rate_mm_hr)
```

<br />

<br />

##### 3. Extract a multi-year 15-minute point time series

Use `zarr_rechunked_as_timeseries/` to read the full multi-year 15-minute record
at a single location in a single chunk read:

<br />

**Python**

<br />

```python
import obstore
import xarray as xr
import zarr

# Open the time-series-chunked PARDIG-NRT archive (chunked time: -1, 5, 5)
gcs_store = obstore.store.GCSStore(
    bucket="pardig_1",
    prefix="NRT/zarr_rechunked_as_timeseries/2016_to_2025/predictions.zarr",
)
ds_ts = xr.open_zarr(zarr.storage.ObjectStore(gcs_store), chunks={})

# Example: New York City (40.7°N, 74.0°W -> convert -180..180 lon via % 360)
lat, lon_360 = 40.7, (-74.0) % 360
point_rate_mm_hr = (
    ds_ts["precipitation"]
    .sel(latitude=lat, longitude=lon_360, method="nearest")
    .compute()
    * 1000.0  # Convert m/h to mm/h
)
print(point_rate_mm_hr)
```

<br />

<br />

*** ** * ** ***

### Access on Earth Engine

Static archives for PARDIG-NRT and PARDIG-LR are available in the public
[Google Earth Engine](https://earthengine.google.com/) Data Catalog as two
`ImageCollection` assets:

| Collection | Earth Engine Asset ID | Resolution | Availability |
| [PARDIG-NRT](https://developers.google.com/earth-engine/datasets/catalog/projects_gcp-public-data-weathernext_assets_PARDIG_1_NRT.md.txt) | `projects/gcp-public-data-weathernext/assets/PARDIG_1_NRT` | 0.1° · 15 min | `2016-01-01T07:45:00Z` through October 2026 (static archive at launch) |
| [PARDIG-LR](https://developers.google.com/earth-engine/datasets/catalog/projects_gcp-public-data-weathernext_assets_PARDIG_1_LR.md.txt) | `projects/gcp-public-data-weathernext/assets/PARDIG_1_LR` | 0.1° · 15 min | `1999-01-01T00:00:00Z` to `2023-12-31T23:45:00Z` (static archive) |
|---|---|---|---|

> [!NOTE]
> **Static archives on Earth Engine versus live updates on Cloud Storage:** Both Earth Engine `ImageCollection` assets are static archives at launch (with `PARDIG_1_NRT` covering January 2016 through October 2026). For continuous 6-hourly near-real-time updates, query the Zarr store on [Google Cloud Storage
> (`gs://pardig_1/NRT/zarr/2026_to_present/`)](https://developers.google.com/weathernext/guides/pardig#access-on-google-cloud-storage).

#### Common query recipes

Use the Earth Engine Python API to filter PARDIG collections by timestamp and
load instantaneous precipitation rates:

##### 1. Load a 15-minute precipitation rate frame in millimeters per hour

<br />

**Python**

<br />

```python
import ee

ee.Initialize(project="YOUR_PROJECT_ID")

# Load the PARDIG-NRT ImageCollection and select a single 15-minute frame
pardig_nrt = ee.ImageCollection(
    "projects/gcp-public-data-weathernext/assets/PARDIG_1_NRT"
)
frame = (
    pardig_nrt.filterDate("2024-09-26T12:00:00Z", "2024-09-26T12:15:00Z")
    .first()
    .select("precipitation")
)

# Convert instantaneous rate from m/h to mm/h
rate_mm_hr = frame.multiply(1000.0)
print("Band names:", rate_mm_hr.bandNames().getInfo())
```

<br />

<br />

*** ** * ** ***

## Starter guides

For interactive Colab notebooks with end-to-end sample code and visualization
workflows:

- [PARDIG Starter Guide - Zarr on Google Cloud Storage](https://storage.googleapis.com/weathernext-public/colabs/Pardig_Starter_Guide_Zarr_on_Google_Cloud_Storage.ipynb)
- [PARDIG Starter Guide - Earth Engine](https://storage.googleapis.com/weathernext-public/colabs/Pardig_Starter_Guide_Earth_Engine.ipynb)

*** ** * ** ***

## Limitations and caveats

Keep the following data characteristics and physical retrieval constraints in
mind when analyzing PARDIG:

- **Temporal homogeneity in PARDIG-LR and across variants:** PARDIG-LR (`1999-2023`) uses coarser `0.25°` ERA5 atmospheric fields, historical GridSat-B1 imagery, and an evolving constellation of polar-orbiting passive microwave sounders (AMSU-A, MHS, ATMS), whereas PARDIG-NRT (`2016 to present`) uses `0.1°` HRES and [SatWeave](https://developers.google.com/weathernext/guides/satweave.md.txt) inputs. Changes in satellite sensor coverage over time within PARDIG-LR, as well as the input transition if concatenating PARDIG-LR with PARDIG-NRT, can introduce non-climatic shifts; exercise caution when interpreting multi-decadal trends as purely climatic signals.
- **Polar latitudes (`> ~67°N/S`):** Although PARDIG outputs a complete `90°S` to `90°N` grid, the GPM Core Observatory orbit only samples within approximately `±67°` latitude, and geostationary satellite coverage degrades beyond `~60°-65°` latitude. Estimates poleward of `~67°N/S` rely on unconstrained spatial extrapolation outside the spaceborne radar training domain and should be treated with caution.
- **Complex terrain, snowfall, and light drizzle:** Skill margins over gauge-adjusted baselines narrow over high-elevation mountain ranges (such as the Alps and the Rockies). Wintertime underestimation in snowy and orographic regimes reflects both fine-scale terrain variability and the sensitivity limits of spaceborne radar and passive microwave sensors to light snowfall and shallow drizzle.
- **Instantaneous rate cap and upstream input artifacts:** Deterministic 15-minute precipitation rates cap at `75 mm/h` (`0.075 m/h`, the upper calibration threshold of the model), which can clip localized peak convective intensities. In addition, occasional artifacts or missing swaths in upstream satellite or NWP inputs (such as geostationary limb or scan seams) can propagate into individual precipitation frames.

*** ** * ** ***

## Terms of service and disclaimers

PARDIG is an experimental AI dataset under active research. Estimates and model
outputs are provided "as-is" for informational and research purposes and don't
constitute official weather observations, forecasts, watches, or warnings. Users
are responsible for their use of the data. For the protection of life and
property, never rely on PARDIG as a sole source of information and always defer
to official alerts and advisories from your national meteorological service and
local emergency authorities.

### Data licensing and terms of use

PARDIG is licensed under the
[Creative Commons Attribution International License, Version 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).
For complete terms, liability provisions, and acceptable use restrictions, see
the [Terms of Service and Disclaimers](https://developers.google.com/weathernext/guides/disclaimers.md.txt) page.

### Citation

If you use PARDIG in your research or applications, cite the accompanying paper:

```tex
@article{shalev2026pardig,
  title={PARDIG: Global Precipitation Estimates Learned from Sparse Satellite Observations},
  author={Shalev, Guy and Rasp, Stephan and Babenko, Boris and El-Kadi, Andrew and Hassen, Mohammed Alewi and Merchant, Samier and Zyda, Fred and Masters, Dominic and Alet, Ferran and R{\"u}hling Cachay, Salva and Price, Ilan and Uddin, Akib and Peled Levi, Nofar and Gaiarin, Ben and Bell, Aaron and Battaglia, Peter and Sanchez-Gonzalez, Alvaro},
  journal={arXiv preprint arXiv:2611.13941},
  year={2026}
}
```

### Acknowledgements and third-party attributions

This experimental dataset was generated by a model using data and products of
the European Centre for Medium-Range Weather Forecasts (ECMWF), NASA, JAXA,
NOAA, EUMETSAT, and JMA, as well as additional third-party providers. For a
complete list of attributions and acknowledgements, see the
[PARDIG Dataset Attributions and Acknowledgements PDF](https://storage.googleapis.com/weathernext-public/pardig-attributions-acknowledgements.pdf).

Use of the third-party materials referred to in the acknowledgements may be
governed by separate terms and conditions or license provisions. Your use of
third-party materials is subject to any such terms, and you should verify that
you can comply with any applicable restrictions before use.

*** ** * ** ***

See the [Glossary](https://developers.google.com/weathernext/guides/glossary.md.txt) for definitions of
meteorological terms, data formats, and Google Cloud platforms.