apendergrass / cmip6hack-precip Goto Github PK

sd_hist = (xr.apply_ufunc(signal.detrend, data_tmp['hist'].fillna(0),
                                kwargs={'axis': 0}).where(~data_tmp['hist'].isnull())
           .sel(time=data_tmp['hist'].time.dt.season.isin(seas))
           .std("time"))
sd_futr = (xr.apply_ufunc(signal.detrend, data_tmp['futr'].fillna(0),
                                kwargs={'axis': 0}).where(~data_tmp['futr'].isnull())
           .sel(time=data_tmp['futr'].time.dt.season.isin(seas))
           .std("time"))

i think detrending is applied to the whole timeseries for each month, but the seasonal mean should be calculated first, then the timeseries of seasonsal means should be detrended, then the std should come last

Calculate precip mean and std change, and huss change, for three regions (all land):

• tropics (< |30° lat|)
• NH XT (>30° lat)
• SH XT (<-30° lat)
  And for two seasons:
• DJF
• JJA

For the plots, average the NH XT, DJF with SH XT, JJA and vice versa to get XT land in winter and XT land in summer

the current CSVs lack the first two time periods, which are commented out in the list

Make the plot look nice.

• Make markers for pr std change bigger than the original
• Consider using markers for mean pr change and huss change.
• Consider making more use of color (no need to continue single-color scheme from paper for this context)

Models have constant field of "areacella" to correctly area average quantities.

• check that we have them for all models in use for cloud
• check that we have them for all models in use for glade
• modify weights from np.cos(latr) to areacella
• leave some if checks for those models without areacella and leave np.cos(latr) as alternative

Add comments throughout

Isolate to land, hopefully using each model's land mask (sftlf variable)

https://github.com/hdrake/cmip6hack-multigen/blob/master/notebooks/0_calc_tas_skill.ipynb

seems to have done the first step of our project nicely.

Currently the code has:

dsdG_djf=results.copy()
dsdG_jja=results.copy()
dmuG_djf=results.copy()
dmuG_jja=results.copy()
for model in results:
for timeperiod in results[model]:
dsdG_djf[model][timeperiod]=results[model][timeperiod]['DJF'].dsdG.values.item(0)
dsdG_jja[model][timeperiod]=results[model][timeperiod]['JJA'].dsdG.values.item(0)
dmuG_djf[model][timeperiod]=results[model][timeperiod]['DJF'].dmuG.values.item(0)
dmuG_jja[model][timeperiod]=results[model][timeperiod]['JJA'].dmuG.values.item(0)

But what actually made the one csv that worked was

dsdG_djf=results
for model in results:
for timeperiod in results[model]:
dsdG_djf[model][timeperiod]=results[model][timeperiod]['DJF'].dsdG.values.item(0)

The problem is that it alters "results" and overwrites the other data. The challenge is to figure out how to get the data out without overwriting.

Fix duplicative calls etc

Enforce that any model with a NaN in any variable has a NaN in all variables before plotting. Right now one model has a lot of NaNs

save data from the loading/calculate notebook into a csv

load and plot it in another notebook

Proposal: envelope of 95% confidence interval of average pr std change based on t-test for spread across models in pr std change at each time.

Could also add envelopes for mean precip and huss (might require more graphic design, but they should be smaller envelopes)

Merge subset_models_with_intake.ipynb, which implements intake-esm but doesn't actually load data, into mean_var_calcs.ipynb, which opens data and calculates and plots things, but only works on the cloud.

Ssps start in 2006 but the first period to evaluate change over starts in 1991

Calculate the change in tas, pr mean, pr std, and huss mean for each ensemble member / variant of a model, and average them together

to get 30 years of DJF seasons where each season is a contiguous 30 years, there needs to be 31 years of data to start, and the partial seasons should be thrown out

for xarray.dataset (eg temp_dapr) - temp_dapr.to_netcdf()

Read in huss, take its mean over the regions/seasons, and incorporate into plots (analogous to mean pr)