def subsample(training, target, sample):
base_dir = get_main_dir()
# randomly subsample the data and store the subsampling distribution
fname = os.path.join(
os.path.join(
os.path.join(os.path.join(base_dir, 'input'), 'calibrations'),
'idealised'), 'subsample_%d.npy' % (sample))
if not os.path.isfile(fname):
size = 7
doys = np.unique(training['doy'])
sub = np.unique(np.random.randint(doys[0], doys[-1] + 1, size=size))
while len(sub) < size:
sub = np.unique(
np.append(
sub,
np.unique(
np.random.randint(doys[0],
doys[-1] + 1,
size=size - len(sub)))))
ssub = np.asarray(training.loc[training['doy'].isin(sub)].index)
"""
# there are small differences in day time hours... same sizes?
if sample > 1: # the first distribution is the reference size
ref = np.load(fname.replace('%d.npy' % (sample), '1.npy'))
diff = len(ref) - len(ssub)
while diff > 0: # random extra data points from any one day
sub = np.unique(np.append(sub,
np.unique(np.random.randint(doys[0],
doys[-1] + 1, size=1))))
ssub = np.asarray(training.loc[training['doy'].isin(sub)].index)
diff = len(ref) - len(ssub)
if diff < 0: # randomly remove the excess
rm = np.random.randint(0, len(ssub), size=abs(diff))
ssub = np.delete(ssub, rm)
"""
np.save(fname, ssub)
else:
ssub = np.load(fname)
# now accordingly subsample input and output
Y = target[ssub]
X = training.iloc[ssub]
X.reset_index(inplace=True, drop=True)
return X, Y
def check_X_Y(swaters):
base_dir = get_main_dir()
# check that the 4 week forcing file exists
fname1 = os.path.join(
os.path.join(
os.path.join(os.path.join(base_dir, 'input'), 'calibrations'),
'idealised'), 'training_x.csv')
if not os.path.isfile(fname1): # create file if it doesn't exist
params = InForcings().defparams
params.doy = random.randrange(92, 275) # random day within GS
InForcings().run(fname1, params, Ndays=7 * 4)
for profile in swaters:
# check that the output file from the reference model exists
fname2 = os.path.join(
os.path.join(
os.path.join(os.path.join(base_dir, 'input'), 'calibrations'),
'idealised'), 'training_%s_y.csv' % (profile))
if not os.path.isfile(fname2):
df1, __ = read_csv(fname1)
# add the soil moisture profile to the input data
df1['sw'], df1['Ps'] = soil_water(df1, profile)
df1['Ps_pd'] = df1['Ps'].copy() # pre-dawn soil water potentials
df1['Ps_pd'].where(df1['PPFD'] <= 50., np.nan, inplace=True)
# fixed value for the wind speed
df1['u'] = df1['u'].iloc[0]
# non time-sensitive: last valid value propagated until next valid
df1.fillna(method='ffill', inplace=True)
__ = hrun(fname2,
df1,
len(df1.index),
'Farquhar',
models=['Medlyn'],
inf_gb=True)
return
def prep_training_N_target(profile, sub=None):
base_dir = get_main_dir()
# path to input data
fname = os.path.join(
os.path.join(
os.path.join(os.path.join(base_dir, 'input'), 'calibrations'),
'idealised'), 'training_x.csv')
df1, __ = read_csv(fname)
# path to output data from the reference model
fname = os.path.join(
os.path.join(
os.path.join(os.path.join(base_dir, 'input'), 'calibrations'),
'idealised'), 'training_%s_y.csv' % (profile))
df2, __ = read_csv(fname)
# add the soil moisture profile to the input data
df1['sw'], df1['Ps'] = soil_water(df1, profile)
df1['Ps_pd'] = df1['Ps'].copy() # daily pre-dawn soil water potentials
df1['Ps_pd'].where(df1['PPFD'] <= 50., np.nan, inplace=True)
# fix the wind speed
df1['u'] = df1['u'].iloc[0]
# non time-sensitive: last valid value propagated until next valid
df1.fillna(method='ffill', inplace=True)
# drop everything below min threshold for photosynthesis and reindex
Y = np.asarray(df2['gs(std)'][df1['PPFD'] > 50.]) * 1000. # mmol m-2 s-1
X = df1[df1['PPFD'] > 50.]
X.reset_index(inplace=True, drop=True)
# add Rnet to the input (no ET or soil albedo feedbacks, this can be done)
X['Rnet'] = net_radiation(X)
X['scale2can'] = 1.
if sub is not None: # randomly subsample one week out of the data
X, Y = subsample(X, Y, sub)
return X, Y
def __init__(self, method='powell', store=None, inf_gb=True):
# fitting method
self.method = method # which solver is used
# MCMC-specific
self.steps = 15000
self.nchains = 4
self.burn = 1000
self.thin = 2
if store is None: # default storing path for the outputs
self.base_dir = get_main_dir() # working paths
self.opath = os.path.join(os.path.join(self.base_dir, 'output'),
'calibrations')
else: # user defined storing path for the outputs
self.opath = store
self.inf_gb = inf_gb # whether to calculate gb or not
def main(fname1,
fname2,
fname3,
calibs='both',
orientation='both',
colours=None):
base_dir = get_main_dir()
dirname = os.path.join(
os.path.join(os.path.join(base_dir, 'output'), 'calibrations'),
'idealised')
# load in the data
df1 = (pd.read_csv(os.path.join(dirname, fname1), header=[0]).dropna(
axis=0, how='all').dropna(axis=1, how='all').squeeze())
df2 = (pd.read_csv(os.path.join(dirname, fname2), header=[0]).dropna(
axis=0, how='all').dropna(axis=1, how='all').squeeze())
df3 = (pd.read_csv(os.path.join(dirname, fname3), header=[0]).dropna(
axis=0, how='all').dropna(axis=1, how='all').squeeze())
if orientation == 'both':
for orientation in ['landscape', 'portrait']:
plt_setup(calibs, orientation, colours=colours) # rendering
calib_info_plot(df1.copy(),
df2.copy(),
df3.copy(),
calibs=calibs,
orientation=orientation)
solver_info_plot(df1)
else:
plt_setup(calibs, orientation, colours=colours) # rendering
solver_info_plot(df1.copy())
calib_info_plot(df1, df2, df3, calibs=calibs, orientation=orientation)
return
def build_calibrated_forcing(training):
base_dir = get_main_dir() # working paths
# forcing file used to calibrate the models
fname = os.path.join(os.path.join(os.path.join(os.path.join(base_dir,
'input'), 'calibrations'), 'obs_driven'),
'%s_x.csv' % (training))
df1, columns = read_csv(fname)
# file containing the best calibrated params
fname = os.path.join(os.path.join(os.path.join(os.path.join(base_dir,
'output'), 'calibrations'), 'obs_driven'),
'best_fit.csv')
df2 = (pd.read_csv(fname, header=[0]).dropna(axis=0, how='all')
.dropna(axis=1, how='all').squeeze())
df2 = df2[df2['training'] == training]
# attribute the first (and second and third) parameter(s)
for i in df2.index:
df1.loc[0, df2.loc[i, 'p1']] = df2.loc[i, 'v1']
if not pd.isnull(df2.loc[i, 'v2']):
df1.loc[0, df2.loc[i, 'p2']] = df2.loc[i, 'v2']
if not pd.isnull(df2.loc[i, 'v3']):
df1.loc[0, df2.loc[i, 'p3']] = df2.loc[i, 'v3']
# save the forcing file containing the calibrated params
df1.columns = columns # original columns
df1.drop([('Tleaf', '[deg C]')], axis=1, inplace=True) # drop Tleaf
df1.to_csv(os.path.join(os.path.join(os.path.join(os.path.join(base_dir,
'input'), 'simulations'), 'obs_driven'),
'%s_calibrated.csv' % (training)), index=False, na_rep='',
encoding='utf-8')
return
ax.set_rgrids([0., 0.25, 0.5, 0.75], ['0', '', '', '0.75'])
ax.set_rmax(0.75)
# draw lines angles
ax.set_thetagrids(np.degrees([0, -45]), [])
ax.set_title('Reading Key:', fontsize=7., x=-1.4, y=0.05)
fig.savefig(figname)
plt.close()
return
###############################################################################
base_dir = get_main_dir()
figname = os.path.join(os.path.join(os.path.join(base_dir, 'output'), 'plots'),
'model_sensitivities_ST_1.5.jpg')
#if not os.path.isfile(figname):
fname = os.path.join(
os.path.join(
os.path.join(
os.path.join(os.path.join(base_dir, 'output'), 'simulations'),
'idealised'), 'sensitivities'),
'overview_of_sensitivities_1.5MPa.csv')
df = (pd.read_csv(fname,
header=[0]).dropna(axis=0,
how='all').dropna(axis=1,
how='all').squeeze())
plot_sensitivities(df, figname)
def calib_info_plot(df1, df2, df3, calibs='wet', orientation='landscape'):
# user-defined plot attributes
vscale = 0.45 # scaling factor around the median
pscale = 1.1 # positions for each model's parameters
pspace = 0.025 # second parameter positions
wbox = 0.9 # width of the violin plots
vert = True
s1 = 'left'
s2 = 'right'
if orientation == 'portrait':
vert = False
s1 = 'top'
s2 = 'bottom'
if calibs == 'both': # colours for the violin plot edge lines
c = plt.rcParams['axes.prop_cycle'].by_key()['color'][-1]
pspace /= 2.5
# landscape characteristics
fs = (6., 3.)
if orientation == 'portrait':
fs = (3.25, 6.)
# declare the figure and the axes
fig, ax = plt.subplots(nrows=1, figsize=fs)
# model order?
models = automate_model_order(df1.copy())
# modify and order model names across all dfs
df1 = update_model_names(df1, models)
df2 = update_model_names(df2, models)
df3 = update_model_names(df3, models)
# organise, normalise, and scale the data consistently
df1, w, i = sorted_data(df1)
wet1 = scaled_data(w, sc=vscale)
inter1 = scaled_data(i, sc=vscale)
# subset of the top 3 solvers
df2, w, i = sorted_data(df2, norm_wet=w, norm_inter=i)
wet2 = scaled_data(w, sc=vscale)
inter2 = scaled_data(i, sc=vscale)
# best params
df3, w, i = sorted_data(df3, norm_wet=w, norm_inter=i)
best_w = scaled_data(w, sc=vscale)
best_i = scaled_data(i, sc=vscale)
# param names in model order?
params = parameter_names(w)
# where are there 2nd params?
all = np.array([i for i in range(len(wet1)) if np.nansum(wet1[i]) != 0.])
isec = np.array([i for i in range(len(all)) if (all[i] % 2) != 0])
pos = np.arange(float(len(all))) * pscale
pos[isec] -= 8. * pspace # second parameter position
pos2 = pos + pspace
if orientation == 'portrait':
pos2 = pos - pspace
# now that we've reworked the positions, only keep those
wet1 = [wet1[i] for i in all]
wet2 = [wet2[i] for i in all]
inter1 = [inter1[i] for i in all]
inter2 = [inter2[i] for i in all]
best_w = [best_w[i] for i in all]
best_i = [best_i[i] for i in all]
# all solver data
Npoints = 1000 # smooth violins
bw = 0.35
if calibs != 'inter':
vp1 = ax.violinplot(wet1,
showextrema=False,
points=Npoints,
positions=pos,
vert=vert,
widths=wbox,
bw_method=bw)
for vp in vp1['bodies']:
vp.set_alpha(0.7)
if calibs != 'wet':
vp2 = ax.violinplot(inter1,
showextrema=False,
points=Npoints,
positions=pos2,
vert=vert,
widths=wbox,
bw_method=bw)
for vp in vp2['bodies']:
vp.set_alpha(0.7)
if calibs == 'both':
slice_vplot(vp1, s1, ec='gray')
slice_vplot(vp2, s2, ec='gray')
# top 3 solver data, if no substantial improvement, no plot
bw *= 2.
plt_wet = np.array([(np.amax(wet2[i]) - np.amin(wet2[i])) < 0.75 *
(np.amax(wet1[i]) - np.amin(wet1[i]))
for i in range(len(wet1))])
plt_inter = np.array([(np.amax(inter2[i]) - np.amin(inter2[i])) < 0.75 *
(np.amax(inter1[i]) - np.amin(inter1[i]))
for i in range(len(inter1))])
wet2 = [wet2[i] for i in range(len(wet2)) if plt_wet[i]]
inter2 = [inter2[i] for i in range(len(inter2)) if plt_inter[i]]
if calibs != 'inter':
vp3 = ax.violinplot(wet2,
showextrema=False,
points=Npoints,
positions=pos[plt_wet],
vert=vert,
widths=wbox,
bw_method=bw)
for vp in vp3['bodies']:
vp.set_alpha(0.7)
vp.set_hatch('/' * 6)
if calibs != 'wet':
vp4 = ax.violinplot(inter2,
showextrema=False,
points=Npoints,
positions=pos2[plt_inter],
vert=vert,
widths=wbox,
bw_method=bw)
for vp in vp4['bodies']:
vp.set_alpha(0.7)
vp.set_hatch('/' * 6)
if calibs == 'both':
slice_vplot(vp3, s1, ec=c)
slice_vplot(vp4, s2, ec=c)
# best params
if calibs == 'both':
x = np.append(pos - wbox / 8., pos + wbox / 4. + pspace / 2.)
y = np.append(best_w, best_i)
elif calibs == 'wet':
x = pos
y = best_w
else:
x = pos
y = best_i
if orientation == 'portrait':
y = x
if calibs == 'both':
x = np.append(best_i, best_w)
elif calibs == 'wet':
x = best_w
else:
x = best_i
ax.plot(x, y, lw=0, marker='*', mec='k', zorder=9)
# add custom legend
if calibs == 'both': # add custom legend
handles = custom_legend(calibs, orientation, ec='gray')
else:
handles = custom_legend(calibs, orientation)
if orientation == 'landscape':
ax.legend(handles=handles, loc=2, bbox_to_anchor=[-0.025, 1.015])
else:
ax.legend(handles=handles,
ncol=3,
columnspacing=1.,
handlelength=1.,
handletextpad=0.4,
frameon=False,
loc=2,
bbox_to_anchor=[0., 1.03])
# add grid and format the axes
lpos = np.asarray([0.05, 0.5, 0.9, 1., 1.1, 2., 20.])
mpos = np.copy(pos)
mpos[isec - 1] += (mpos[isec] - mpos[isec - 1]) / 2.
mpos = np.delete(mpos, isec)
mlines = np.copy(pos) + pscale / 2.
mlines[isec] += pspace * 2. / 3.
mlines = np.delete(mlines, isec - 1)
custom_grid(mlines, -(vscale**lpos), ax, orientation)
if orientation == 'landscape':
ax.set_yticks(-(vscale**lpos))
if calibs == 'both':
ax.set_xlim([np.amin(pos) - 0.5, np.amax(pos) + 0.5])
else:
ax.set_xlim([np.amin(pos) - 0.55, np.amax(pos) + 0.6])
ax.set_xticks(mpos)
else:
ax.set_xticks(-(vscale**lpos))
if calibs == 'both':
ax.set_ylim([np.amin(pos) - 0.5, np.amax(pos) + 0.5])
else:
ax.set_ylim([np.amin(pos) - 0.6, np.amax(pos) + 0.55])
ax.set_yticks(mpos + 0.15)
# nicer display of the model names and normalised param values
pvals = ['0.05', '0.5', '0.9', '', '1.1', '2', '20']
models[models.index('WUE-LWP')] = r'WUE-$f_{\varPsi_l}$'
models[models.index('SOX-OPT')] = r'SOX$_\mathrm{\mathsf{opt}}$'
if orientation == 'landscape':
ax.set_yticklabels(pvals)
ax.set_xticklabels(models, va='top', rotation=25., size=7.)
for i in range(len(params)): # add param names
t = ax.text(pos[i],
-(vscale**0.2),
params[i],
va='top',
ha='center')
t.set_bbox(
dict(boxstyle='round,pad=0.1', fc='w', ec='none', alpha=0.8))
# move the y labels to the right side
ax.yaxis.set_label_position('right')
ax.yaxis.tick_right()
ax.set_ylabel('Normalised parameter values')
else:
ax.set_xticklabels(pvals)
ax.set_yticklabels(models, ha='left', va='top', size=7.)
for i in range(len(params)): # add param names
yy = pos[i] + 0.125
if i in [9, 14, 15, 16]:
yy = pos[i] + 0.375
t = ax.text(-(vscale**4.4), yy, params[i], ha='right', va='top')
if i == 12:
t.set_path_effects([
path_effects.Stroke(linewidth=0.75, foreground='w'),
path_effects.Normal()
])
ax.tick_params(axis='y', direction='in', pad=-5.)
plt.setp(ax.get_yticklabels(),
bbox=dict(boxstyle='round', fc='w', ec='none'))
ax.set_xlabel('Normalised parameter values')
# remove the ticks themselves
ax.xaxis.set_tick_params(length=0.)
ax.yaxis.set_tick_params(length=0.)
base_dir = get_main_dir()
opath = os.path.join(os.path.join(base_dir, 'output'), 'plots')
fig.tight_layout()
plt.savefig(
os.path.join(opath, 'model_calibs_%s_%s.png' % (calibs, orientation)))
plt.savefig(
os.path.join(opath, 'model_calibs_%s_%s.jpg' % (calibs, orientation)))
def solver_info_plot(df):
# declare the figure and the axes
fig, ax = plt.subplots(figsize=(2.75, 3.))
if 'training' not in df.columns:
size = 30.
else:
size = 20.
# solver performance info
df = solver_performance(df)
c = plt.rcParams['axes.prop_cycle'].by_key()['color'][-1]
# counts plot where the points are bigger as more points overlap
ax.scatter(df['sv'],
df['Rank'],
marker='o',
c='grey',
alpha=0.7,
s=df['counts'] * size)
# plot the N best data
ax.scatter(df[df['sv'] < 3]['sv'],
df[df['sv'] < 3]['Rank'],
marker='o',
c=c,
s=df[df['sv'] < 3]['counts'] * size / 2.)
# plot the average ranks
ax.plot(pd.unique(df['sv']),
df.groupby('sv')['waRank'].mean(),
c='k',
lw=1.5)
# format the axes
ax.set_xticks(np.arange(len(df['solver'].unique())) + 0.5)
ax.set_xticklabels(df['solver'].unique(),
rotation=55.,
ha='right',
va='top')
ax.xaxis.set_tick_params(length=0.)
# replace y axis with skill arrow
ax.get_yaxis().set_visible(False)
ax.text(-0.125,
0.95,
'Low\nskill',
va='center',
ha='center',
transform=ax.transAxes)
ax.annotate('High\nskill',
xy=(-0.125, 0.9),
xytext=(-0.125, 0.05),
xycoords='axes fraction',
va='center',
ha='center',
arrowprops=dict(arrowstyle='<-', lw=0.75))
for spine in ax.spines.values(): # thinner spines
spine.set_visible(True)
spine.set_linewidth(0.25)
base_dir = get_main_dir()
opath = os.path.join(os.path.join(base_dir, 'output'), 'plots')
fig.tight_layout()
plt.savefig(os.path.join(opath, 'solver_performance.png'))
plt.savefig(os.path.join(opath, 'solver_performance.jpg'))
return
handles, labels = ax.get_legend_handles_labels()
ax.legend(handles[:len(mods) + 1],
labels[:len(mods) + 1],
bbox_to_anchor=(1.025, 1. / 3.),
loc=3)
fig.savefig(fname)
plt.close()
###############################################################################
# first, activate user defined rendering options
plt_setup()
base_dir = get_main_dir() # dir paths
ifdir = os.path.join(
os.path.join(os.path.join(base_dir, 'input'), 'simulations'), 'idealised')
ofdir = os.path.join(
os.path.join(os.path.join(base_dir, 'output'), 'simulations'), 'idealised')
# path to input data
fname1 = os.path.join(ifdir, 'wet_calibration.csv')
df1, __ = read_csv(fname1)
# initialise soil moisture forcings
df1['sw'] = df1['theta_sat']
df1.fillna(method='ffill', inplace=True)
# plot the atmospheric forcings
figdir = os.path.join(os.path.join(base_dir, 'output'), 'plots')
def obs_calibs(df1, df2, figname):
fig = plt.figure(figsize=(6.5, 8.))
gs = fig.add_gridspec(nrows=96, ncols=16, hspace=0.3, wspace=0.2)
ax2 = fig.add_subplot(gs[52:, 6:]) # conductance data
ipath = os.path.join(
os.path.join(os.path.join(get_main_dir(), 'input'), 'simulations'),
'obs_driven')
labels = []
for i, what in enumerate(df1['site_spp'].unique().dropna()):
if i < 13:
nrow = int(i / 4) * 16
ncol = (i % 4) * 4
ax1 = fig.add_subplot(gs[nrow:nrow + 16, ncol:ncol + 4])
else:
nrow += 16
ax1 = fig.add_subplot(gs[nrow:nrow + 16, :4])
sub = df1.copy()[df1['site_spp'] == what]
sub = sub.select_dtypes(exclude=['object', 'category'])
sub = sub[sub['Pleaf'] > -9999.]
sub['gs'] /= sub['gs'].max()
for day in sub['doy'].unique():
mask = sub['doy'] == day
plot_obs(ax1, sub['Pleaf'][mask], sub['gs'][mask])
x0, x1, obs_popt = fit_Tuzet(sub)
x = np.linspace(sub['Pleaf'].max(), sub['Pleaf'].min(), 500)
ax1.plot(x, fsig_tuzet(x, obs_popt[0], obs_popt[1]), 'k', zorder=30)
ax1.vlines(x0, 0., 1., linestyle=':')
ax1.vlines(x1, 0., 1., linestyle=':')
# get the integrated VC given by the obs and site params
ref, __ = read_csv(os.path.join(ipath, '%s_calibrated.csv' % (what)))
b, c = Weibull_params(ref.iloc[0])
int_VC = np.zeros(len(sub))
for j in range(len(sub)):
int_VC[j], __ = quad(f,
sub['Pleaf'].iloc[j],
sub['Ps'].iloc[j],
args=(b, c))
plot_obs(ax2, i, np.log(sub['E'] / int_VC), which='kmax')
# subplot titles (including labelling)
what = what.split('_')
species = r'\textit{%s %s}' % (what[-2], what[-1])
labels += [r'\textit{%s. %s}' % (what[-2][0], what[-1])]
if 'Quercus' in what:
species += ' (%s)' % (what[0][0])
labels[-1] += ' (%s)' % (what[0][0])
txt = ax1.annotate(r'\textbf{(%s)} %s' %
(string.ascii_lowercase[i], species),
xy=(0.025, 0.98),
xycoords='axes fraction',
ha='left',
va='top')
txt.set_bbox(
dict(boxstyle='round,pad=0.1', fc='w', ec='none', alpha=0.8))
# format axes ticks
ax1.xaxis.set_major_locator(mpl.ticker.NullLocator())
if (i == 13) or ((ncol > 0) and (nrow == 32)):
render_xlabels(ax1, r'$\Psi_{l}$', 'MPa')
if ncol == 0:
ax1.yaxis.set_major_locator(mpl.ticker.MaxNLocator(3))
ax1.yaxis.set_major_formatter(
mpl.ticker.FormatStrFormatter('%.1f'))
ax1.set_ylabel(r'$g_{s, norm}$')
else:
ax1.yaxis.set_major_locator(mpl.ticker.MaxNLocator(3))
ax1.set_yticklabels([])
ax2.annotate(r'\textbf{(%s)}' % (string.ascii_lowercase[i + 1]),
xy=(0.05, 0.98),
xycoords='axes fraction',
ha='right',
va='top')
# add max conductance parameter values
params, models = get_calib_kmax(df2)
params = np.asarray(params)
locs = np.arange(len(df1['site_spp'].unique()))
# update colour list
colours = ([
'#6023b7', '#af97c5', '#009231', '#6b3b07', '#ff8e12', '#ffe020',
'#f10c80', '#ffc2cd'
]) * len(params)
for i in range(params.shape[1]):
if i < 8:
ax2.scatter(locs,
params[:, i],
s=50,
linewidths=0.25,
c=colours[i],
alpha=0.9,
label=models[0][i],
zorder=4)
else:
ax2.scatter(locs,
params[:, i],
s=50,
linewidths=0.25,
c=colours[i],
alpha=0.9,
zorder=4)
# tighten the subplot
ax2.set_xlim(locs[0] - 0.8, locs[-1] + 0.8)
ax2.set_ylim(np.log(0.025) - 0.1, np.log(80.))
# ticks
ax2.set_xticks(locs + 0.5)
ax2.set_xticklabels(labels, ha='right', rotation=40)
ax2.xaxis.set_tick_params(length=0.)
yticks = [0.025, 0.25, 1, 5, 25, 75]
ax2.set_yticks([np.log(e) for e in yticks])
ax2.set_yticklabels(yticks)
render_ylabels(ax2, r'k$_{max}$', 'mmol m$^{-2}$ s$^{-1}$ MPa$^{-1}$')
handles, labels = ax2.get_legend_handles_labels()
labels[3] = 'SOX$_\mathrm{\mathsf{opt}}$'
ax2.legend(handles,
labels,
ncol=3,
labelspacing=1. / 3.,
columnspacing=0.5,
loc=3)
# save
fig.savefig(figname)
#==============================================================================
to_fit = True
sample = None # None, 1, 2, or 3
swaters = ['wet', 'inter'] # two different soil moisture profiles
# declare empty dataframe which will be used to analyse the calibrations
odf = pd.DataFrame(columns=[
'Model', 'training', 'solver', 'BIC', 'Rank', 'p1', 'v1', 'p2', 'v2'
])
# where should the fitting solvers' outputs be stored?
base_dir = get_main_dir() # working paths
check_X_Y(swaters) # check that the training calibration files exist
if to_fit:
for swater in swaters: # loop over the training soil moisture profiles
X, Y = prep_training_N_target(swater, sub=sample)
# where should the calibration output be stored?
opath = os.path.join(
os.path.join(os.path.join(base_dir, 'output'), 'calibrations'),
'idealised')
if sample is not None: # move files to the relevant sub-dir
