Analysis-Ready, Cloud Optimized ERA5

Recipes for reproducing Analysis-Ready & Cloud Optimized (ARCO) ERA5 datasets.

Introduction • Roadmap • Data Description • How to reproduce • FAQs • How to cite this work • License

Introduction

Our goal is to make a global history of the climate highly accessible in the cloud. To that end, we present a curated copy of the ERA5 corpus in Google Cloud Public Datasets.

So far, we have ingested meteorologically valuable variables for the land and atmosphere. From this, we have produced a cloud-optimized version of ERA5, in which we have converted grib data to Zarr with no other modifications. Next, we plan on creating an "analysis-ready" version, oriented towards common research workflows, which we will do in the open.

This two-pronged approach for the data serves different user needs. Some researchers need full control over the interpolation of data for their analysis. Most will want a batteries-included dataset, where standard pre-processing and chunk optimization is already applied. In general, we ensure that every step in this pipeline is open and reproducible, to provide transparency in the provenance of all data.

TODO(#1): What have we done to make this dataset possible?

Please view out our walkthrough notebook for a demo of the datasets.

Roadmap

1. Phase 0: Ingest raw ERA5
2. Phase 1: Cloud-Optimize to Zarr, without data modifications
   1. Use Pangeo-Forge to convert the data from grib to Zarr.
   2. Create example notebooks for common workflows, including regridding and variable derivation.
3. Phase 2: Produce an Analysis-Ready corpus
   1. Regrid datasets to lat/long grids.
   2. Convert model levels to pressure levels (at high resolution).
   3. Compute derived variables.
   4. Expand on example notebooks.
4. Phase 3: Create an analysis & machine learning (ML) pipeline toolkit
   1. Dataset generator for ML Models.
   2. Examples of reading data in XArray-Beam pipelines.
   3. Notebooks demoing common data analysis, like Extreme Value Analysis.
5. General future plans...
   1. Include more variables, especially ocean data.
   2. Integrate preliminary ERA5 data (1950 to 1978).
   3. Automatically update with recent data.

Data Description

As of 2022-04-27, all data spans the dates 1979-01-01/to/2021-08-31 (inclusive).

Whenever possible, we have chosen to represent parameters by their native grid resolution. See this ECMWF documentation for more.

Model Level Wind

• Levels: 1/to/137
• Times: 00/to/23
• Grid: Spectral Harmonic Coefficients (docs)
• Size: 305.89 TiB

Warning: We encountered a small error in one of our source data shards (the dve shard for 2008-08-27 actually had data for 2008-02-27). We noticed this only after ECMWF's MARS archive went down due to a data center migration. In order to release the data as soon as possible, we substituted the previous day's worth of data for this 24-hour period. Once the center is back online, we will re-compute this dataset (#8).

import xarray as xr

ml_wind = xr.open_zarr(
    'gs://gcp-public-data-arco-era5/co/model-level-wind.zarr/',
    chunks={'time': 48},
    consolidated=True,
)

Model Level Moisture

• Levels: 1/to/137
• Times: 00/to/23
• Grid: N320, a Reduced Gaussian Grid (docs)
• Size: 707.34 TiB

import xarray as xr

ml_moisture = xr.open_zarr(
    'gs://gcp-public-data-arco-era5/co/model-level-moisture.zarr/',
    chunks={'time': 48},
    consolidated=True,
)

Single Level Surface

• Times: 00/to/23
• Grid: Spectral Harmonic Coefficients (docs)
• Size: 1.11 TiB

import xarray as xr

ml_surface = xr.open_zarr(
    'gs://gcp-public-data-arco-era5/co/single-level-surface.zarr/',
    chunks={'time': 48},
    consolidated=True,
)

Single Level Reanalysis

• Times: 00/to/23
• Grid: N320, a Reduced Gaussian Grid (docs)
• Size: 28.02 TiB

import xarray as xr

sl_reanalysis = xr.open_zarr(
    'gs://gcp-public-data-arco-era5/co/single-level-reanalysis.zarr', 
    chunks={'time': 48},
    consolidated=True,
)

Single Level Forecast

• Times: 06:00/18:00
• Steps: 0/1/2/3/4/5/6/7/8/9/10/11/12/13/14/15/16/17/18
• Grid: N320, a Reduced Gaussian Grid (docs)
• Size: 24.52 TiB

import xarray as xr

sl_forecasts = xr.open_zarr(
    'gs://gcp-public-data-arco-era5/co/single-level-forecast.zarr/', 
    chunks={'time': 48},
    consolidated=True,
)

How to reproduce

All phases of this dataset can be reproduced with scripts found here. To run them, please clone the repo and install the project.

git clone https://github.com/google-research/arco-era5.git

Or, via SSH:

git clone [email protected]:google-research/arco-era5.git

Then, install with pip:

cd arco-era5
pip install -e .

Acquire & preprocess raw data

Please consult the instructions described in raw/.

Cloud-Optimization

All our tools make use of Apache Beam, and thus are portable to any cloud (or Runner). We use GCP's Dataflow to produce this dataset. If you would like to reproduce the project this way, we recommend the following:

1. Ensure you have access to a GCP project with GCS read & write access, as well as full Dataflow permissions (see these "Before you begin" instructions).
2. Export the following variables:

   export PROJECT=<your-gcp-project>
   export REGION=us-central1
   export BUCKET=<your-beam-runner-bucket>

From here, we provide examples of how to run the recipes at the top of each script.

pydoc src/single-levels-to-zarr.py

You can also discover available command line options by invoking the script with -h/--help:

python src/model-levels-to-zarr.py --help

Making the dataset "Analysis Ready" & beyond...

This phase of the project is under active development! If you would like to lend a hand in any way, please check out our contributing guide.

FAQs

How did you pick these variables?

This dataset originated in Loon, Alphabet’s project to deliver internet service using stratospheric balloons, and is now curated by Google Research & Google Cloud Platform. Loon’s Planning, Simulation and Control team needed accurate data on how the stratospheric winds have behaved in the past to evaluate the effectiveness of different balloon steering algorithms over a range of weather. This led us to download the model-level data. But Loon also needed information about the atmospheric radiation to model balloon gas temperatures, so we downloaded that. And then we downloaded the most commonly used meteorological variables to support different product planning needs (RF propagation models, etc)...

Eventually, we found ourselves with a comprehensive history of weather for the world.

Where are the U/V components of wind? Where is geopotential height? Why isn’t X variable in this dataset?

We intentionally did not include many variables that can be derived from other variables. For example, U/V components of wind can be computed from divergence and vorticity; geopotential is a vertical integral of temperature.

In the second phase of our roadmap (towards "Analysis Ready" data), we aim to compute all of these variables ourselves. If you’d like to make use of these parameters sooner, please check out our example notebooks where we demo common calculations. If you notice non-derived missing data, such as surface variables, please let us know of your needs by filing an issue, and we will be happy to incorporate them into our roadmap.

Do you have plans to get all of ERA5?

We aim to support hosting data that serves general meteorological use cases, rather than aim for total completeness. Wave variables are missing from this corpus, and are a priority on our roadmap. If there is a variable or dataset that you think should be included here, please file a Github issue.

For a complete ERA5 mirror, we recommend consulting with the Pangeo Forge project (especially staged-recipes#92).

Why are there two model-level datasets and not one?

It definitely is possible for all model level data to be represented in one grid, and thus one dataset. However, we opted to preserve the native representation for variables in ECMWF's models. A handful of core model variables (wind, temperature and surface pressure) are represented as spectral harmonic coefficients , while everything else is stored on a Gaussian grid. This avoids introducing numerical error by interpolating these variables to physical space. For a more in depth review of this topic, please consult these references:

• Fundamentals of Numerical Weather Prediction by Jean Coiffier
• Atmospheric modeling, data assimilation, and predictability by Eugenia Kalnay

Please note: in a future releases, we intend to create a dataset version where all model levels are in one grid and Zarr.

Why doesn’t this project make use of Pangeo Forge Cloud?

We are big fans of the Pangeo Forge project, and of Pangeo in general. While this project does make use of their Recipes, we have a few reasons to not use their cloud. First, we would prefer to use internal rather than community resources for computations of this scale. In addition, there are several technical reasons why Pangeo Forge as it is today would not be able to handle this case (0, 1, 2, 3). To work around this, we opted to combine familiar-to-us infrastructure with Pangeo-Forge's core and to use the right tool for the right job.

Why use this dataset? What uses are there for the data?

ERA5 can be used in many applications. It can be used to train ML models that predict the impact of weather on different phenomena. ERA5 data could also be used to train and evaluate ML models that forecast the weather. The data could be used to compute climatologies, or the average weather for a region over a given period of time. ERA5 data can be used to visualize and study historical weather events, such as Hurricane Sandy.

Where should I be cautious? What are the limitations of the dataset?

Mumbai, India	San Francisco, USA
Tokyo, Japan	Singapore

ERA5 Topography	GMTED2010 Topography

It is important to remember that a reanalysis is an estimate of what the weather was, it is not guaranteed to be an error-free estimate. There are several areas where the novice reanalysis user should be careful.

First, the user should be careful using reanalysis data at locations near coastlines. The first figure shows the fraction of land (1 for land, 0 for ocean) of ERA5 grid points at different coastal locations. This is important because the land-surface model used in ERA5 tries to blend in the influence of water with the influence of land based on this fraction. The most visible effect of this blending is that as the fraction of land decreases, the daily variation in temperature will also decrease. Looking at the first figure, there are sharp changes in the fraction of land between neighboring grid cells so there could be differences in daily temperature range that might not be reflected in actual weather observations.

The user should also be careful when using reanalysis data in areas with large variations in topography. The second figure is a plot of ERA5 topography around Mount Everest compared with GMTED2010 topography. The ERA5 topography is completely missing the high peaks of the Everest region and missing most of the structure of the mountain valleys. Topography strongly influences temperature and precipitation rate, so it is possible that ERA5’s temperature is too warm and ERA5’s precipitation patterns could be wrong as well.

ERA5’s precipitation variables aren’t directly constrained by any observations, so we strongly encourage the user to check ERA5 against observed precipitation (for example, Wu et al., 2022). A study comparing reanalyses (not including ERA5) against gridded precipitation observations showed striking differences between reanalyses and observation Lisa V Alexander et al 2020 Environ. Res. Lett. 15 055002.

Can I use the data for {research,commercial} purposes?

Yes, you can use our ERA5 data according to the terms of the Copernicus license.

Researchers, see the next section for how to cite this work.

Commercial users, please be sure to provide acknowledgement to the Copernicus Climate Change Service according to the Copernicus Licence terms.

How to cite this work

Please cite our presentation at the 22nd Conference on Artificial Intelligence for Environmental Science describing ARCO-ERA5.

Carver, Robert W, and Merose, Alex. (2023): ARCO-ERA5: An Analysis-Ready Cloud-Optimized Reanalysis Dataset. 22nd Conf. on AI for Env. Science, Denver, CO, Amer. Meteo. Soc, 4A.1, https://ams.confex.com/ams/103ANNUAL/meetingapp.cgi/Paper/415842

In addition, please cite the ERA5 dataset accordingly:

Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., 
Muñoz‐Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., 
Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G., 
Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., De Chiara, G., 
Dahlgren, P., Dee, D., Diamantakis, M., Dragani, R., Flemming, J., 
Forbes, R., Fuentes, M., Geer, A., Haimberger, L., Healy, S., 
Hogan, R.J., Hólm, E., Janisková, M., Keeley, S., Laloyaux, P., 
Lopez, P., Lupu, C., Radnoti, G., de Rosnay, P., Rozum, I., Vamborg, F.,
Villaume, S., Thépaut, J-N. (2017): Complete ERA5: Fifth generation of 
ECMWF atmospheric reanalyses of the global climate. Copernicus Climate 
Change Service (C3S) Data Store (CDS). (Accessed on DD-MM-YYYY)

Hersbach et al, (2017) was downloaded from the Copernicus Climate Change 
Service (C3S) Climate Data Store. We thank C3S for allowing us to 
redistribute the data.

The results contain modified Copernicus Climate Change Service 
information 2022. Neither the European Commission nor ECMWF is 
responsible for any use that may be made of the Copernicus information 
or data it contains.

License

This is not an official Google product.

Copyright 2022 Google LLC

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    https://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.