Generative Adversarial Networks for Improving Earth System Model Precipitation

Description

This repository contains the code for the training a cycle consistent generative adversarial network on Earth system model output data for bias correction.

Requirements

The dependencies are installed in a Singularity container that can be pulled from

singularity pull --arch amd64 library://phess/pytorch-stack/stack.sif:v3

Data

• The W5E5 reanalysis data can be downloaded at Link.
• The CMIP6 data can be downloaded at WCRP Coupled Model Intercomparison Project (Phase 6).

Usage

Training:

1. Define the parameters and file paths in src/configuration.py
2. run:

 singularity run --nv --bind /path/to/current/directory /path/to/container/stack_v3.sif python main.py

Evaluation:

To evaluate the results define parameters and paths in src/configuration.py and use the Jupyther notebooks:

• Evaluation of the GAN model checkpoints: notebooks/summary-statistics.ipynb
• Comparison of the GAN model and baselines: notebooks/analysis-combined-results.ipynb
• Evaluation of spectral densities: notebooks/analysis-spectral-density.ipynb
• Evaluation of fractals: notebooks/analysis-fractal-dimension.ipynb

To start Jupyter Lab run:

 singularity run --nv --bind /path/to/current/directory /path/to/container/stack_v3.sif jupyter-lab