danijonesocean/ozone-clustering

Unsupervised classification of ozone profiles

Ozone clustering in UKESM1

Here we use Gaussian mixture modelling as implimented in the scikit-learn python package to classify ozone profiles in the UKESM1 model. We train the model using annual mean profiles from:

1. The historical simulation from 2004-2014

2. SSP126 from 2090-2100

3. SSP585 from 2090-2100

By training the GMM using a combined dataset with profiles from (1)-(3) we ensure that the GMM captures a wide variety of profile types across the forcing scenarios. We then label the profiles separately and examine how the class properties change over time and with forcing scenario.

This repository contains:

• bic : a plot of the BIC score with number of classes to guide selection of N
• classes-n : collections of plots and statistical models using 'n' classes
• data_in : both the NetCDF data extracted from UKESM1 and the same data formatted as CSV
• get_UKESM_ozone_profiles : the script used to extract ozone profiles from UKESM1 on Pangeo
• Ozone profile clustering : notebook containing the fitting and labeling of classes

ozone-clustering's Issues

Decide how to treat interannual variability, change code accordingly

At present, only one year is selected for plotting in the notebook. We could take the mean over years for each decade instead. To implement this, use the Pandas groupby function with the reset_index function to convert the result back to single indices. For example,

df.groupby(['x','y']).mean().reset_index()

That way we can be more general than just plotting individual years. We could also consider plotting the interannual variability in some way as well.

Add background information about UKESM1 biases

Luke Abraham showed at the ACSIS meeting (July 2021) that UKESM1 has total column ozone values that are too high relative to the rest of the CMIP6 ensemble. In addition, it has a greater relative drop than other CMIP6 members.