Sub-Seasonal Climate Forecasting (SSF) Dataset

University of Illinois at Urbana-Champaign (UIUC) University of Minnesota, Twin Cities (UMN)

George Mason University (GMU) Carnegie Mellon University (CMU)

Overview

The SSF climate dataset is a benchmark dataset for training and evaluating machine learning models for sub-seasonal climate forecasting, including a variety of high-resolution climate variables over the atmosphere, ocean, and land from 1980 to recent.

The codebase developed for data extraction, preprocessing, and SSF model training and evaluation using the SSF dataset are publicly available at https://github.com/SSF-climate/SSF and https://github.com/Sijie-umn/SSF-MIP/.

Details

The SSF Dataset is constructed by climate variables representing the condition of atmosphere, land, and ocean from 1980 until the recent, which are collected from a diverse collection of data sources. In addition, we include some climate indexes monitoring El Niño-Southern Oscillation (Niño indexes) and North Atlantic Oscillation (NAO), etc. All the climate variables are interpolated to 0.5 degree latitude by 0.5 degree longitude grid and daily temporal resolution from 1980 to the recent. The climate variables of each year are saved as multi-index pandas DataFrames and can be read using python 3.

Spatiotemporal Climate Variables

tmp2m: 2 meter temperature
precip: precipitation
sst: sea surface temperature
slp: sea level pressure
icec: sea ice concentration
hgt10, hgt200, hgt500, hgt700: geopotential height at 10mb, 200mb, 500mb, and 700mb
rhum.sig995: relative humidity at level sig 995
sm: soil moisture

Temporal Climate Variables

mei: Multivariate ENSO Index Version 2
mjo_phase: Madden/Julian Oscillation phase
mjo_amplitude: Madden/Julian Oscillation amplitude
nao: North Atlantic Oscillation index
nino1+2, nino3, nino4, nino3.4: nino sst indicies
ssw: sudden stratospheric warming index

Spatial Climate Variables

elevation
masks: latitude and longitude coverage of the United States, North Atlantic, and North Pacific Ocean, etc.

SubX (NCEP-CFSv2 and GMAO-GEOS)

the climatology of tmp2m of weeks 3 and 4 computed from the corresponding hindcast periods
the SubX hindcasts of tmp2m anomalies of weeks 3 and 4 over the western U.S. (1999 - 2015)
the SubX forecasts of tmp2m anomalies of weeks 3 and 4 over the western U.S. (2017 - 2020)
the initialization dates in the forecast periods of NCEP-CFSv2 and GMAO-GEOS

Groundtruth (computed from the NOAA’s Climate Prediction Center (CPC) Global Gridded Temperature dataset)

the week 3 & 4 temperature anomalies over the western U.S. from 1990 to 2020
the climatology computed from 1990 to 2017 for each grid point over the western U.S. and each month-day combination

Preprocessed data

pca-zscore: The folder contains covariates after preprocessing (pca first then z-scored) from 1986 to 2018.
train-validation: The folder contains the training and validation data used for hyper parameter tuning.
train-test: The folder contain the training and test data used for tmp2m forecasting over 2017-2018