Using machine learning to analyze physical causes of climate change: a case study of U.S. Midwest extreme precipitation
Supporting code for Davenport and Diffenbaugh (2021) 'Using machine learning to analyze physical causes of climate change: a case study of U.S. Midwest extreme precipitation'
Davenport, F. V. & Diffenbaugh, N. S. (2021). Using machine learning to analyze physical causes of climate change: a case study of U.S. Midwest extreme precipitation. Geophysical Research Letters, 48, e2021GL093787, https://doi.org/10.1029/2021GL093787
If you find meaningful errors in the code of have questions, please contact Frances Davenport at fvdav@stanford.edu
- input_data: input data used for analysis (not all raw data is included due to size; see details below)
- notebooks: jupyter notebooks and python scripts to read and analyze data, and create figures
- processed_data: processed data from analysis
- project_utils: python utilities used in analysis (see instructions below to install)
- figures (placeholder directory for figure pdfs created by running figure notebooks in notebooks directory)
- environment.yml (specifies python packages needed to run notebooks)
- environment_tf1.yml (specifies python packages needed to run the LRP notebook)
Data analyzed in the paper is publicly available from the following sources:
-
PRISM daily precipitation: available from the PRISM Climate Group, Oregon State University
-
NCEP/NCAR-R1 reanalysis: daily reanalysis is available through ftp access from NOAA. The following variables are used in the analysis: geopotential height (hgt), sea-level pressure (slp), zonal wind (uwind), meridional wind (vwind), specific humidity (shum).
-
GHCN precipitation station data: daily precipitation observations from the Global Historical Climatology Network are available from NOAA.
Due to the large size of the raw data, pre-processed data has been uploaded to figshare and is available to download here. See notebooks directory for more detail on recreating the analysis starting from the pre-processed data.
- download this repository
- download pre-processed data here into the processed_data folder.
- install the required python modules using conda. The environment.yml and environment_tf1.yml provide information on the required modules. (The environment_tf1.yml files specifies the tensorflow 1 compatible environment needed to calculate the layerwise relevance propagation - see notebooks directory for more details)
- install project_utils in conda environment with the following command (should be run from within the main project directory):
pip install -e . --user- run and/or edit the notebooks.