Decadal prediction of climate data subject to different Greenhouse Gas emission scenarios using a variational autoencoder

Overview

Jupyter notebooks for the implementation of a variational autoencoder (VAE) for climate data modelling and prediction.

The Jupyter notebooks demonstrate the process of training and exploring a Variational Autoencoder (VAE) on precipitation and sea-surface temperature data. The VAE is trained on different Shared Socio-economic Pathways (SSPs) data of CMIP6.

Requirements

1. The Jupyter notebooks requires the VAE package, which is available at:

   https://github.com/andr-groth/VAE-project

2. Sample data used in the notebook is included in the data/ folder. The data is in netCDF format and has been prepared with the help of the CDO scripts, which are available at:

   https://andr-groth.github.io/CDO-scripts

   For more information on the data preparation see data/README.md.

Usage

This repository contains two Jupyter notebooks for model training and prediction:

1. To train a VAE on a range of different Greenhouse Gas (GHG) emission scenarios: VAEp_train.ipynb

2. To make predictions with the VAE on new GHG emission scenarios: VAEp_explore.ipynb

The notebooks use data provided in the data/ folder in this repository.

Results

The work is published in

Groth, 2023: Decadal agricultural production risk prediction subject to different Greenhouse Gas emission scenarios in African countries. Technical report.

The report is available at report/VAE-SSP.pdf in this repository.