pystuff 0.0.7

Some useful functions for data analysis in python.

pystuff

Some useful functions for data analysis in python.

• Introduction
• Handling time series
• Periodogram
• Principal Component Analysis (PCA)
• Ensemble Subsampling
• Statistics

Introduction

Installation

pip install pystuff

Import pystuff

import pystuff as ps

Import auxiliary packages

import numpy as np
import xarray as xr
import pandas as pd
import matplotlib.pyplot as plt

Read example data

# ERA-Interim SST anomalies, weighted by sqrt(cos(lat))
pathfile='/work/uo1075/u241292/data/ERAI/atmo/monavg_sst_w2_anom.nc'
ds=xr.open_dataset(pathfile)
lat=ds['lat'].values
lon=ds['lon'].values
myvar=ds['sst'].values
time=pd.to_datetime(ds['time'].values)

Handling time series

# Area Average
nasst, sasst_map = ps.getbox([0,60,280,360],lat,lon,myvar,returnmap=True)

# Standardize (center=True is default)
nasstn = ps.standardize(nasst)

# Running Mean
nasstn_rm = ps.runmean(nasstn,window=5*12)
nasstn_rm_fill = ps.runmean(nasstn,window=5*12,fillaround=True)

# Detrend (redundant with ddreg, to some extent)
# If returnTrend=False, only detended series is returned
nasstndt, slope, trend = ps.ddetrend(nasstn, returnTrend=True)

# The trend line can also be taken from:
trend2 = ps.ddreg(range(len(nasst)),nasst)

# Annual Mean
annual=ps.annualmean(nasstndt)

# Low-pass Lanczos Filter
dt=12 # month
cutoff=5 # years
low, low_nonan = ps.lanczos(nasstndt,dt*cutoff,returnNonan=True)

# Example Plot
fig = plt.figure(figsize=(9,4.5))

ax=fig.add_subplot(1,2,1)
ps.nospines(ax)
plt.axhline(0,color='grey',ls='--')
plt.plot(time,nasstn,'b',lw=0.2, label='Monthly NASST')
plt.plot(time,nasstn_rm,'b',lw=2, label='5-year running mean')
plt.plot(time,low+trend,'g',lw=2, label='5-year low-pass Lanczos filter')
plt.plot(time,trend,'-r', lw=2, label='Trend line')
plt.text(0.1,0.8,'Trend=%.2f std year$^{-1}$' %(float(slope)*12), color='r',
         transform=plt.gcf().transFigure)
plt.ylim((-2.5,2.5))
plt.ylabel('std. units')
plt.title('NASST')
plt.title('a', loc='left',fontweight='bold', fontsize=14)
ps.leg(loc='lower right', fontsize=9)

ps.usetex(False)
plt.tight_layout()
plt.show()

fig.savefig('/home/zmaw/u241292/scripts/python/pystuff/figs/timeseries.png')		

Periodogram

fig = plt.figure(figsize=(5,4.5))

ax=fig.add_subplot(111)
ps.nospines(ax)

# Calculate yearly periodogram on monthly data (Monte Carlo with 10k iterations)
f, psd, pctl, max5, meanRed = ps.periods(nasstndt, dt=12, nsim=10000)

plt.plot(f, psd,'b', lw=2, label='NASST Power Spectrum')
plt.plot(f,meanRed, 'r', lw=2, label='Mean of 10k Red-Noise Spectra')
plt.plot(f,pctl[:,0], '--k', label='80% Percentile')
plt.plot(f,pctl[:,1], '-k', label='90% Percentile')
plt.plot(f,pctl[:,2], '--g', label='95% Percentile')
plt.plot(f,pctl[:,3], '-g', label='99% Percentile')
plt.xlabel('Frequency [year$^{-1}$]')
plt.ylabel('PSD [Units$^{2}$ year]')
plt.xlim((0,2.5))
plt.title('Periodogram of NASST')
plt.title('b', loc='left',fontweight='bold', fontsize=14)
ps.leg(fontsize=10, frameon=True)

# Print maximum periods on graph
for i in range(5):
    t=plt.text(max5[i,0],max5[i,1],'%.0f yr' %round(max5[i,2]), color='blue')
    t.set_bbox(dict(facecolor='white', alpha=0.5, edgecolor='w'))


plt.tight_layout()
plt.show()

fig.savefig('/home/zmaw/u241292/scripts/python/pystuff/figs/periodogram.png')

Principal Component Analysis (PCA)

# Get data from any 3 grid cells
x1=myvar[:,100,100]
x2=myvar[:,100,200]
x3=myvar[:,100,300]
X=np.transpose([x1,x3,x2]) # np.shape(X) = (468, 3)

# Calculate PCA
scores, eigenvals, eigenvecs, expl, expl_acc, means, stds, north, loadings = ps.ddpca(X)

# Combo-plot
from matplotlib.gridspec import GridSpec
gs = GridSpec(nrows=2, ncols=4)
f=plt.figure(figsize=(6,6))

ax=f.add_subplot(gs[0, 0:2])
ps.usetex()
ps.nospines(ax)
plt.errorbar(np.arange(1,len(expl)+1),expl,yerr=[north,north], fmt='o',color='b',markeredgecolor='b')
for i in range(3):
    plt.text(i+1.2,expl[i],'$%.1f \pm %.1f$' %(expl[i],north[i]))
plt.xticks(np.arange(1,len(expl)+1,1))
plt.xlim((0.5,4))
plt.xlabel('PC')
plt.ylabel('Explained Variance [\%]')
plt.text(0.6,46.5,'a',fontsize=14,fontweight='heavy')

ax=f.add_subplot(gs[0, 2:])
ps.nospines(ax)
wdt=0.15
plt.axhline(0,color='k')
plt.bar(np.arange(3)-wdt,eigenvecs[0,:],facecolor='r',edgecolor='r',width=wdt)
plt.bar(np.arange(3)    ,eigenvecs[1,:],facecolor='g',edgecolor='g',width=wdt)
plt.bar(np.arange(3)+wdt,eigenvecs[2,:],facecolor='b',edgecolor='b',width=wdt)
plt.ylabel('Eigenvector values')
plt.xticks(np.arange(3)+0.125,['$1$','$2$','$3$'],usetex=True)
plt.xlabel('PC')
plt.axvline(0.625,color='lightgrey',ls='--')
plt.axvline(1.625,color='lightgrey',ls='--')
plt.text(0.8,0.5,'Series 1',color='r')
plt.text(0.8,0.4,'Series 2',color='g')
plt.text(0.8,0.3,'Series 3',color='b')
plt.text(-0.4,0.9,'b',fontsize=14,fontweight='heavy')

ax1=f.add_subplot(gs[1, 1:-1])
ax1.set_xlim((-3,3)); plt.ylim((-3,3))
ax1.set(xlabel='PC1', ylabel='PC2')
ax1.axvline(0,color='k')
ax1.axhline(0,color='k')
ax1.plot(scores[:,0],scores[:,1],'o',markersize=3,
         markeredgecolor='lightgrey',markerfacecolor='lightgrey')
ax1.text(-2.8,2.5,'c',fontsize=15,fontweight='heavy')
ax2 = ps.twinboth(ax1)
ax2.set_xlim((-1.5,1.5)); plt.ylim((-1.5,1.5))
ax2.set_xlabel('PC1 Loadings', labelpad=3)
ax2.set_ylabel('PC2 Loadings', labelpad=14)
ax2.arrow(0,0,loadings[0,0],loadings[0,1],width=0.005,color='r',lw=2)
ax2.arrow(0,0,loadings[1,0],loadings[1,1],width=0.005,color='g',lw=2)
ax2.arrow(0,0,loadings[2,0],loadings[2,1],width=0.005,color='b',lw=2)

plt.tight_layout()
plt.show()

f.savefig('figs/pca.png')

Ensemble Subsampling

# Create a fake ensemble of nsim members of red noise (preserving the lag-dt) correlation
nsim=1000
dt=1
ens=ps.rednoise(len(nasstndt),ps.rhoAlt(nasstndt, dt=dt), nsim=nsim)

# Get Percentiles of Ensemble Spread
spread=ensPctl(ens,pctl=[0.025,0.975])

# Get the best 5% ensemble members, based on correlation
bestens, ensmean, nmembers = bestEns(ens,nasstndt,pctl=0.95)


# Plot
f=plt.figure(figsize=(8,6))

ax=f.add_subplot(2,1,1)
ps.nospines(ax)
plt.plot(time, ens,'lightgrey', lw=0.5)
# plt.fill_between(time, spread[:,0],spread[:,1],facecolor='k',alpha=0.1,label='2 std')
plt.plot(time, ens[:,1],'lightgrey', lw=1, label='All 10k Ensemble Members')
plt.plot(time, ens[:,1],'r', lw=0.5, label='Member 1')
plt.plot(time, np.nanmean(ens,axis=1),'k', lw=2, label='Full Ensemble Mean')
plt.plot(time, nasstndt,'b',lw=1,label='NASST')
plt.ylim((-6,6))
plt.ylabel('s.d.')
plt.title('Full Ensemble (10k Members)')
ps.leg(loc='upper right', frameon=True)

ax=f.add_subplot(2,1,2)
ps.nospines(ax)
plt.plot(time,bestens[:,0],'lightgrey', label='Best 5\% of Ensemble (50 members)')
plt.plot(time,bestens,'lightgrey')
plt.plot(time,ensmean,'k', lw=2, label='Best Ensemble Mean')
plt.plot(time,nasstndt,'b',lw=1,label='NASST')
plt.plot(time,low,'g',lw=2,label='NASST low-pass filt.')
plt.ylim((-6,6))
plt.ylabel('s.d.')
plt.title('Best Ensemble (50 Members)')
ps.leg(loc='upper right', frameon=True)

plt.tight_layout()
plt.show()
f.savefig('figs/ensemeble.png')

Statistics

# Annual means of monthly values
y=ps.annualmean(x)

# Autocorrelation of lag 1
r=ps.rho(x)

# Autocorrelation of lag n
r=ps.rhoAlt(x,dt=n)

# Rednoise time series
red=ps.rednoise(length, lag-1 correlation, nsim)

# Value of percentile
x95=getPercentile(x,pctl=0.95)

# Multiple Linear Regression (Uncertainties based on t-test)
X = [x1, x2, ... xn]
fitted, lo, up, r2, r2adj, [coefs] = ps.mlr(X, y, stds=2, returnCoefs=False, printSummary=False)

# Multiple Linear Regression (Uncertainties baed on Bootstrap)
fitted_0, up, lo, r2adj_0, r2adj_un, [r2_0, r2_unc, coefs_0, coefs] = ps.bootsmlr(X, y, n=1000, conflev=0.95, positions='new', uncertain='Betas', details=False, printSummary=False)

# Predict MLR, using coefs from the two above
pred, lo, up = ps.mlr_predict(X,coefs,stds=2)

# Residuals frmo MLR
res = ps.mlr_res(X, y)

# Histogram
binlims, x_count = ps.ddhist(x,vmin,vmax,binsize, zeronan=True, density=False)

# Order (sort ascending) x and y based on values of x
xo, yo = ps.order(x,y)

# Horizontal divergence
div = ps.hdivg(u,v,lat,lon,regulargrid=True)

# Vorticity (vertical component of relative vorticity)
vor = ps.hcurl(u,v,lat,lon,regulargrid=True)

Miscellaneous

# Saving CDO- and NCView-compatible NetCDF

# import pandas as pd
# mytime = pd.date_range('1979-01-01', periods=numYears, freq='1A')

newds = xr.Dataset(data_vars={'lat': (["lat"], lat),
                              'lon': (["lon"], lon),
                              'time': (["time"], np.arange(firstYear,lastYear+1)),
                              'clim' : (["lat","lon"], clim),
                              'trend': (["lat","lon"], trend),
                              'std'  : (["lat","lon"], std),
                              'stdDt': (["lat","lon"], stdDt),
                              'prop' : (["lat","lon"], prop),
                              'hs'   : (["time","lat","lon"], tempArray)})

newds['lat'].attrs  = {'standard_name':'lat','units':'degrees_north', '_CoordinateAxisType':'Lat'}
newds['lon'].attrs  = {'standard_name':'lon','units':'degrees_east',  '_CoordinateAxisType':'Lon'}
newds['time'].attrs = {'standard_name':'time', 'units':'years since 1979-01-01 00:00:00',
                       'calendar': 'proleptic_gregorian', '_CoordinateAxisType':'Time'}

newds['hs'].attrs    = {'standard_name':'hs',   'units':'m'}
newds['clim'].attrs  = {'standard_name':'clim', 'units':'m'}
newds['trend'].attrs = {'standard_name':'trend','units':'m/year'}
newds['std'].attrs   = {'standard_name':'std',  'units':'m'}
newds['stdDt'].attrs = {'standard_name':'stdDt','units':'m'}
newds['prop'].attrs  = {'standard_name':'prop', 'units':'%'}

newds.attrs = {'script' :'/home/u241292/scripts/python/acdtools/acdtools/Hs_Clim.ipynb',
               'content':'Statistics and Time Series of Significant Wave Height (Hs) from ERA5 remapcon2 WAM'}
newds.to_netcdf('/work/uo1075/u241292/outputs/CEModel/era5-rampcon2-wam_StatsTs_1979-2000_Hs.nc')
#                 unlimited_dims = 'time',
#                 encoding = {'time': {'dtype': 'datetime64[ns]'}})