ENSO-ASC 1.0.0

This is the code for this paper https://doi.org/10.5194/gmd-2021-213

This project can be built and trained on Ubuntu 18.04.3 LTS, with python3.6/3.7 and CUDA 10.0/cudnn 7.6.5.

0. Environment

conda create -n enso python=3.6
source activate enso

pip install tensorflow-gpu==2.0.0
pip install netCDF4==1.5.3
pip install pandas==0.25.3
pip install scikit-learn==0.21.3
pip install progress==1.5
pip install loguru==0.3.2
pip install absl-py
pip install cmaps
pip install geos
pip install pyproj
pip install h5py==2.10
conda install -c conda-forge basemap-data-hires=1.0.8.dev0
conda install -c conda-forge pygrib

1. Download climate dataset

Scripts in ./data/reanalysis_dataset/meta-data and ./data/remote_sensing_dataset/meta-data are prepared well for download data from NOAA/CIRES Twentieth Century Global Reanalysis Version 2c and Remote Sensing System. Hadley Centre Global Sea Ice and Sea Surface Temperature (HadISST) can be downloaded from website.

python ./data/download_*_*.py

The archieved dataset is also in (not the latest!)

2. For the reanalysis dataset (only need to run once for the later transfer learning)

Firstly, use the following commands to parse and parpare training data.

python -m data.reanalysis_dataset.1_grib2npz
python -m data.reanalysis_dataset.2_interpolation

The output training data files are in ./data/reanalysis_dataset/final

Then, train the model:

python -m data.preprocess_reanalysis_transfer
python -m train.train_multi_gpus [or] python -m train.train_single_gpu

3. For the remote sensing dataset (need to train for every month)

Firstly, use the following commands to parse and parpare training data.

python -m data.remote_sensing_dataset.1_byte2npz
python -m data.remote_sensing_dataset.2_nc2npz
python -m data.remote_sensing_dataset.3_crop_region_and_fill_land

The output training data files are in ./data/remote_sensing_dataset/final

Then, train the model:

python -m data.preprocess_remote_sensing
python -m train.train_multi_gpus [or] python -m train.train_single_gpu

4. Monthly ENSO forecasting

Firstly, download the remote sensing data from the above wetsites and modify the ./data/remote_sensing_dataset/record.txt to supplement the new data for data preprocessing, such as:

2022-1,2,3  # split months by ',' if more than one month

Secondly, prepare the new data:

python -m data.remote_sensing_dataset.1_byte2npz
python -m data.remote_sensing_dataset.2_nc2npz
python -m data.remote_sensing_dataset.3_crop_region_and_fill_land

Thirdly, fine-tune the trained model:

python workflow.py

Finally, make forecasts for the future 18 months:

python forecast.py