Dataset Card for WxC-Bench

WxC-Bench primary goal is to provide a standardized benchmark for evaluating the performance of AI models in Atmospheric and Earth Sciences across various tasks.

Dataset Details

WxC-Bench contains datasets for six key tasks:

1. Nonlocal Parameterization of Gravity Wave Momentum Flux
2. Prediction of Aviation Turbulence
3. Identifying Weather Analogs
4. Generation of Natural Language Weather Forecasts
5. Long-Term Precipitation Forecasting
6. Hurricane Track and Intensity Prediction

Dataset Description

1. Nonlocal Parameterization of Gravity Wave Momentum Flux

The input variables consist of three dynamic atmospheric variables (zonal and meridional winds and potential temperature), concatenated along the vertical dimension. The output variables are the zonal and meridional components of vertical momentum flux due to gravity waves.

• Curated by: Aman Gupta

2. Generation of Natural Language Weather Forecasts

The dataset includes the HRRR re-analysis data paired with NOAA Storm Prediction Center daily reports for January 2017. This task aims to generate human-readable weather forecasts.

• Curated by: NASA IMPACT

3. Long-Term Precipitation Forecasting

This dataset contains daily global rainfall accumulation records and corresponding satellite observations. The goal is to predict rainfall up to 28 days in advance.

• Curated by: Simon Pfreundschuh (Colorado State University)

4. Aviation Turbulence Prediction

Aimed at detecting turbulence conditions that impact aviation safety.

• Curated by: NASA IMPACT

5. Hurricane Track and Intensity Prediction

Provides HURDAT2 data for predicting hurricane paths and intensity changes.

• Curated by: NASA IMPACT

6. Weather Analog Search

Data to identify analog weather patterns for improved forecasting.

• Curated by: NASA IMPACT

Dataset Sources

Nonlocal Parameterization of Gravity Wave Momentum Flux

Developed using ERA5 reanalysis data (top 15 pressure levels above 1 hPa are excluded). Inputs were coarsely grained from winds and temperatures on a 0.3° grid.

Long-Term Precipitation Forecasting

Precipitation data sources include the PERSIANN CDR dataset (until June 2020) and IMERG final daily product. Satellite observations are sourced from PATMOS-x, GridSat-B1, and SSMI(S) brightness temperatures CDRs, with baseline forecasts from ECMWF and the UK Met Office S2S database.

Dataset Structure

WxC-Bench datasets are organized by task directories:

WxC-Bench
aviation_turbulence
nonlocal_parameterization
weather_analogs
hurricane
weather_forecast_discussion
long_term_precipitation_forecast

Each directory contains datasets specific to the respective downstream tasks.

Dataset Creation

Curation Rationale

The WxC-Bench dataset aims to create a unified standard for assessing AI models applied to complex meteorological and atmospheric science tasks.

Source Data

The datasets were created using multiple authoritative data sources, such as ERA5 reanalysis data, NOAA Storm Prediction Center reports, PERSIANN CDR, and IMERG products. Data processing involved spatial and temporal alignment, quality control, and variable normalization.

Citation

BibTeX:

@misc{shinde2024wxcbenchnoveldatasetweather,
      title={WxC-Bench: A Novel Dataset for Weather and Climate Downstream Tasks}, 
      author={Rajat Shinde and Christopher E. Phillips and Kumar Ankur and Aman Gupta and Simon Pfreundschuh and Sujit Roy and Sheyenne Kirkland and Vishal Gaur and Amy Lin and Aditi Sheshadri and Udaysankar Nair and Manil Maskey and Rahul Ramachandran},
      year={2024},
      eprint={2412.02780},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2412.02780}, 
}

Dataset Card Authors

• Rajat Shinde
• Christopher E. Phillips
• Sujit Roy
• Ankur Kumar
• Aman Gupta
• Simon Pfreundschuh
• Sheyenne Kirkland
• Vishal Gaur
• Amy Lin
• Aditi Sheshadri
• Manil Maskey
• Rahul Ramachandran

Dataset Card Contact

For each task, please contact:

• Nonlocal Parameterization of Gravity Wave Momentum Flux: Aman Gupta
• Aviation Turbulence Prediction: Christopher E. Phillips
• Identifying Weather Analogs: Christopher E. Phillips, Rajat Shinde
• Natural Language Weather Forecasts: Rajat Shinde, Sujit Roy
• Long-Term Precipitation Forecasting: Simon Pfreundschuh
• Hurricane Track and Intensity Prediction: Ankur Kumar