This repository contains the code and pretrained models accompanying the paper "Ensembling geophysical models using Bayesian neural networks". This code has the following dependencies: python >=3.6, tensorflow-gpu == 1.15, matplotlib == 3.2.1, numpy == 1.18.5, basemap == 0.1.
The BayNNE for the toy problem can be trained using Toy Problem/Training.py and the plots from the paper may be reproduced by running Toy Problem/Plotting.ipynb. Similarly, the BayNNE for the ozone data can be trained using Ozone/Training.py. Pre-trained neural network model files may be found in Ozone/Pretrained and these can be loaded by Ozone/Loading.ipynb. The plots from the paper and baseline comparisons can be reproduced by running Ozone/Plotting.ipynb. To work with these notebooks, the ozone dataset (https://osf.io/ynax2/download) must be present in the Ozone folder.
An easily runnable (binder) example of the toy example can be found here:
The following table summarizes the prediction RMSEs in Dobson Units on subsets of the validation dataset.
| Method | Temporal extrapolation | Missing North Pole | Missing South Pole | Missing Tropics | Satellite Voids | Small Features |
|---|---|---|---|---|---|---|
| BayNNE | 4.4 | 4.7 | 6.6 | 2.7 | 2.1 | 3.2 |
| Bilinear* | 31.2 | 1.7 | 3.4 | |||
| Spatiotemporal Kriging* | 7.0 | 2.2 | 3.4 | |||
| Multi-model mean | 15.7 | 8.8 | 30.5 | 9.8 | 9.2 | 16.4 |
| Weighted mean | 8.7 | 12.3 | 22.1 | 8.2 | 8.5 | 10.2 |
| Spatially weighted mean | 9.8 | 6.6 | 19.6 | 5.5 | 5.2 | 10.0 |
*Used for interpolation only