Skip to content

A spherical CNN for weather forecasting

License

Notifications You must be signed in to change notification settings

deepsphere/deepsphere-weather

Repository files navigation

DeepSphere-Weather - Deep Learning on the sphere for weather/climate applications.

weather forecast

t850_state_obs_vs_pred.mp4

The code in this repository provides a scalable and flexible framework to apply convolutions on spherical unstructured grids for weather/climate applications.

ATTENTION: The code is subject to changes in the coming weeks / months.

The folder experiments (will) provide examples for:

  • Weather forecasting using autoregressive CNN models
  • Weather field downscaling (aka superesolution) [in preparation].
  • Classication of atmospheric features (i.e. tropical cyclone and atmospheric rivers) [in preparation].

The folder tutorials (will) provide jupyter notebooks describing various features of DeepSphere-Weather.

The folder docs (will) contains slides and notebooks explaining the DeepSphere-Weather concept.

Installation

For a local installation, follow the below instructions.

  1. Clone this repository.

    git clone https://github.com/deepsphere/deepsphere-weather.git
    cd deepSphere-weather
  2. Install manually the following dependencies:

    • Install first pytorch and its extensions on GPU:
      conda install -c conda-forge pytorch-gpu  
    • If you don't have GPU available install it on CPU:
      conda install -c conda-forge pytorch-cpu  
    • Install the required packages:
    conda create --name weather python=3.8
    conda install xarray dask cdo h5py h5netcdf netcdf4 zarr numcodecs rechunker xskillscore
    conda install notebook jupyterlab
    conda install matplotlib-base cartopy pycairo seaborn cycler
    conda install numpy pandas numba scipy bottleneck
    conda install yaml tabulate tqdm deepdiff
    conda install healpy igl shapely      
    pip install git+https://github.com/epfl-lts2/pygsp@sphere-graphs
    pip install torchinfo
    • Clone the required repository:
    git clone [email protected]:ghiggi/xverif.git
    git clone [email protected]:ghiggi/xscaler.git
    git clone [email protected]:ghiggi/xsphere.git
    git clone [email protected]:ghiggi/xforecasting.git
  3. Alternatively install the dependencies using one of the appropriate below environment.yml files:

    conda env create -f environment_python3.8.5.yml
    conda env create -f environment_python3.9.yml

    and clone the required repository:

    git clone [email protected]:ghiggi/xverif.git
    git clone [email protected]:ghiggi/xscaler.git
    git clone [email protected]:ghiggi/xsphere.git
    git clone [email protected]:ghiggi/xforecasting.git
  4. Add the PYTHONPATH (i.e. in the .bashrc) for the following packages. Here is an example:

     export PYTHONPATH="${PYTHONPATH}:/home/ghiggi/Python_Packages/xscaler"
     export PYTHONPATH="${PYTHONPATH}:/home/ghiggi/Python_Packages/xverif"
     export PYTHONPATH="${PYTHONPATH}:/home/ghiggi/Python_Packages/xsphere"
     export PYTHONPATH="${PYTHONPATH}:/home/ghiggi/Python_Packages/xforecasting"  
     export PYTHONPATH="${PYTHONPATH}:/home/ghiggi/Projects/deepsphere-weather"

Tutorials

Data structure

Here we are going to document the data structure

The data required for model training and evaluation are stored with the following folder hierarchy.

preprocessed
└───ERA5_HRES
|   └─── <sampling>
|         └─── Data
|         └─── Scalers
|         └─── Benchmarks 
|         └─── Climatology 
| 
└───IF5_HRES
|    └─── <sampling>
|         └─── Data
|    
└───IF5_ENS
|   └─── <sampling>
         └─── Data

On the LTE servers, data are available in the directory /ltenas3/data/DeepSphere

The data are available upon request to the authors.

Reproducing our results

Contributors

License

The content of this repository is released under the terms of the MIT license.