def eoldas_inversion(fluxnet_site,
year,
n_years,
green_leaves,
gamma_lai,
n_tries=5):
f = IntProgress(min=0, max=2 * n_tries + 1)
f.value = 1
display(f)
prior_set = []
for train_year in xrange(n_years):
this_year = (year - n_years - 1) + train_year
retval_s, state, obs = tip_inversion(this_year,
fluxnet_site,
green_leaves=green_leaves,
n_tries=2)
prior_set.append(retval_s)
mu_priors = np.array(
[state.__unpack_from_dict(p['real_map']) for p in prior_set])
cov_mtx = np.zeros((322, 322))
for p in prior_set:
cov_mtx = cov_mtx + p['hessian'].todense()
prior = Prior(mu_priors.mean(axis=0), cov_mtx)
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv(
cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd_single = np.where(post_sd > c, c, post_sd)
retval, state, obs = regularised_tip_inversion(
year,
fluxnet_site, [1e-3, 0, 0.1, 1e-3, 0, 0.1, gamma_lai],
x0=retval_s['real_map'],
green_leaves=green_leaves,
n_tries=n_tries,
progressbar=f)
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv(
cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = [
'omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai'
]
post_sd = np.sqrt(np.array(retval['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
#axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
# mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].fill_between(state.state_grid,
retval['real_map'][params[i]] -
post_sd[(i * 46):((i + 1) * 46)],
retval['real_map'][params[i]] +
post_sd[(i * 46):((i + 1) * 46)],
lw=0.8,
color="0.8")
axs[i].vlines(state.state_grid,
retval_s['real_map'][params[i]] -
post_sd_single[(i * 46):((i + 1) * 46)],
retval_s['real_map'][params[i]] +
post_sd_single[(i * 46):((i + 1) * 46)],
lw=0.8,
colors="0.1",
alpha=0.5)
axs[i].plot(state.state_grid,
retval['real_map'][params[i]],
'o-',
mfc="none")
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]],
'--')
if i in [1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel(tip_params[i])
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot([0, 0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0],
obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0],
obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo,
fig.bbox.xmax - 500,
fig.bbox.ymax - 250,
alpha=.4,
zorder=1) #if __name__ == "__main__":
def prior_experiment ( fluxnet_site, year, gamma_lai, green_leaves=False, inflation=2.,
n_tries=5 ):
retval_s, state, obs = tip_inversion( year-1, fluxnet_site, green_leaves=green_leaves,
n_tries=4 )
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
main_diag = np.array(1. / ((inflation * post_sd) ** 2))
prior = Prior ( state.pack_from_dict (retval_s['real_map']),
sp.lil_matrix(np.diag( main_diag )))
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd_single = np.where(post_sd > c, c, post_sd)
retval, state, obs = regularised_tip_inversion( year, fluxnet_site, [1e-3, 0, 0.1, 1e-3, 0, 0.1, gamma_lai ],
x0=retval_s['real_map'], green_leaves=green_leaves,
n_tries=n_tries, prior=prior )
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = ['omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai']
post_sd = np.sqrt(np.array(retval['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
#axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
# mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].fill_between ( state.state_grid, retval['real_map'][params[i]] - post_sd[(i*46):((i+1)*46)],
retval['real_map'][params[i]] + post_sd[(i*46):((i+1)*46)], lw=0.8, color="0.8")
axs[i].vlines ( state.state_grid, retval_s['real_map'][params[i]] - post_sd_single[(i*46):((i+1)*46)],
retval_s['real_map'][params[i]] + post_sd_single[(i*46):((i+1)*46)], lw=0.8,
colors="0.1", alpha=0.5)
axs[i].plot(state.state_grid, retval['real_map'][params[i]], 'o-', mfc="none")
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]], '--')
if i in [ 1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel( tip_params[i] )
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot ( [0,0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0], obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0], obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo, fig.bbox.xmax - 500, fig.bbox.ymax - 250, alpha=.4, zorder=1)#if __name__ == "__main__":
def eoldas_inversion(fluxnet_site,
year,
green_leaves,
gamma_lai,
albedo_unc_avg,
albedo_unc_good,
n_tries=10):
f = IntProgress(min=0, max=2 * n_tries + 1)
f.value = 1
display(f)
# convert uncertainties from percent
albedo_unc = [albedo_unc_good * 0.01, albedo_unc_avg * 0.01]
print albedo_unc
retval_s, state, obs = tip_inversion(year,
fluxnet_site,
albedo_unc=albedo_unc,
green_leaves=green_leaves,
n_tries=n_tries,
progressbar=f)
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv(
cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd_single = np.where(post_sd > c, c, post_sd)
retval, state, obs = regularised_tip_inversion(
year,
fluxnet_site, [1e-3, 0, 0.1, 1e-3, 0, 0.1, gamma_lai],
albedo_unc=albedo_unc,
x0=retval_s['real_map'],
green_leaves=green_leaves,
n_tries=n_tries,
progressbar=f)
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv(
cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = [
'omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai'
]
post_sd = np.sqrt(np.array(retval['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
#axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
# mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].fill_between(state.state_grid,
retval['real_map'][params[i]] -
post_sd[(i * 46):((i + 1) * 46)],
retval['real_map'][params[i]] +
post_sd[(i * 46):((i + 1) * 46)],
lw=0.8,
color="0.8")
axs[i].vlines(state.state_grid,
retval_s['real_map'][params[i]] -
post_sd_single[(i * 46):((i + 1) * 46)],
retval_s['real_map'][params[i]] +
post_sd_single[(i * 46):((i + 1) * 46)],
lw=0.8,
colors="0.1",
alpha=0.5)
axs[i].plot(state.state_grid,
retval['real_map'][params[i]],
'o-',
mfc="none")
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]],
'--')
if i in [1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel(tip_params[i])
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot([0, 0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0],
obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0],
obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo,
fig.bbox.xmax - 500,
fig.bbox.ymax - 250,
alpha=.4,
zorder=1)
plt.savefig("regularised_model_%s_%04d_%02dpcnt_%02dpcnt.pdf" %
(fluxnet_site, year, int(
albedo_unc[0] * 100.), int(albedo_unc[1] * 100.)),
dpi=300,
bbox_inches="tight")
def single_observation_inversion( fluxnet_site, year, green_leaves=False, n_tries=5 ):
"""A function to do invert each individual observation in a time series. """
retval_s, state, obs = tip_inversion( year, fluxnet_site,
green_leaves=green_leaves,
n_tries=n_tries )
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = ['omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai']
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].plot(state.state_grid, mu[(i * 46):((i + 1) * 46)], '--')
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]], 'o-', mfc="none")
axs[i].vlines ( state.state_grid, retval_s['real_map'][params[i]] - post_sd[(i*46):((i+1)*46)],
retval_s['real_map'][params[i]] + post_sd[(i*46):((i+1)*46)], lw=0.8, colors="0.1", alpha=0.5)
if i in [ 1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel( tip_params[i] )
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot ( [0,0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0], obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0], obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo, fig.bbox.xmax - 500, fig.bbox.ymax - 250, alpha=.4, zorder=1)
plt.savefig( "single_obs_inversion_%s_%04d.pdf" % ( fluxnet_site, year), dpi=300, boox_inches="tight")
def prior_experiment ( fluxnet_site, year, gamma_lai, green_leaves=False, inflation=2.,
n_tries=5 ):
retval_s, state, obs = tip_inversion( year-1, fluxnet_site, green_leaves=green_leaves,
n_tries=4 )
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
main_diag = np.array(1. / ((inflation * post_sd) ** 2))
prior = Prior ( state.pack_from_dict (retval_s['real_map']),
sp.lil_matrix(np.diag( main_diag )))
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd_single = np.where(post_sd > c, c, post_sd)
retval, state, obs = regularised_tip_inversion( year, fluxnet_site, [1e-3, 0, 0.1, 1e-3, 0, 0.1, gamma_lai ],
x0=retval_s['real_map'], green_leaves=green_leaves,
n_tries=n_tries, prior=prior )
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = ['omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai']
post_sd = np.sqrt(np.array(retval['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
#axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
# mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].fill_between ( state.state_grid, retval['real_map'][params[i]] - post_sd[(i*46):((i+1)*46)],
retval['real_map'][params[i]] + post_sd[(i*46):((i+1)*46)], lw=0.8, color="0.8")
axs[i].vlines ( state.state_grid, retval_s['real_map'][params[i]] - post_sd_single[(i*46):((i+1)*46)],
retval_s['real_map'][params[i]] + post_sd_single[(i*46):((i+1)*46)], lw=0.8,
colors="0.1", alpha=0.5)
axs[i].plot(state.state_grid, retval['real_map'][params[i]], 'o-', mfc="none")
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]], '--')
if i in [ 1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel( tip_params[i] )
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot ( [0,0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0], obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0], obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo, fig.bbox.xmax - 500, fig.bbox.ymax - 250, alpha=.4, zorder=1)#if __name__ == "__main__":
def single_observation_inversion( fluxnet_site, year, green_leaves=False, n_tries=5 ):
"""A function to do invert each individual observation in a time series. """
retval_s, state, obs = tip_inversion( year, fluxnet_site,
green_leaves=green_leaves,
n_tries=n_tries )
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = ['omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai']
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].plot(state.state_grid, mu[(i * 46):((i + 1) * 46)], '--')
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]], 'o-', mfc="none")
axs[i].vlines ( state.state_grid, retval_s['real_map'][params[i]] - post_sd[(i*46):((i+1)*46)],
retval_s['real_map'][params[i]] + post_sd[(i*46):((i+1)*46)], lw=0.8, colors="0.1", alpha=0.5)
if i in [ 1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel( tip_params[i] )
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot ( [0,0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0], obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0], obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo, fig.bbox.xmax - 500, fig.bbox.ymax - 250, alpha=.4, zorder=1)
plt.savefig( "single_obs_inversion_%s_%04d.pdf" % ( fluxnet_site, year), dpi=300, boox_inches="tight")
def eoldas_inversion ( fluxnet_site, year, n_years, green_leaves, gamma_lai,
n_tries=5 ):
f = IntProgress(min=0, max=2*n_tries + 1)
f.value = 1
display(f)
prior_set = []
for train_year in xrange ( n_years ):
this_year = (year - n_years - 1) + train_year
retval_s, state, obs = tip_inversion( this_year, fluxnet_site, green_leaves=green_leaves,
n_tries=2 )
prior_set.append ( retval_s )
mu_priors = np.array( [ state.__unpack_from_dict (p['real_map']) for p in prior_set ] )
cov_mtx = np.zeros (( 322, 322))
for p in prior_set:
cov_mtx = cov_mtx + p['hessian'].todense()
prior = Prior ( mu_priors.mean(axis=0), cov_mtx )
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd_single = np.where(post_sd > c, c, post_sd)
retval, state, obs = regularised_tip_inversion( year, fluxnet_site, [1e-3, 0, 0.1, 1e-3, 0, 0.1, gamma_lai ],
x0=retval_s['real_map'], green_leaves=green_leaves,
n_tries=n_tries, progressbar=f)
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = ['omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai']
post_sd = np.sqrt(np.array(retval['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
#axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
# mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].fill_between ( state.state_grid, retval['real_map'][params[i]] - post_sd[(i*46):((i+1)*46)],
retval['real_map'][params[i]] + post_sd[(i*46):((i+1)*46)], lw=0.8, color="0.8")
axs[i].vlines ( state.state_grid, retval_s['real_map'][params[i]] - post_sd_single[(i*46):((i+1)*46)],
retval_s['real_map'][params[i]] + post_sd_single[(i*46):((i+1)*46)], lw=0.8,
colors="0.1", alpha=0.5)
axs[i].plot(state.state_grid, retval['real_map'][params[i]], 'o-', mfc="none")
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]], '--')
if i in [ 1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel( tip_params[i] )
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot ( [0,0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0], obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0], obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo, fig.bbox.xmax - 500, fig.bbox.ymax - 250, alpha=.4, zorder=1)#if __name__ == "__main__":
def eoldas_inversion ( fluxnet_site, year, green_leaves, gamma_lai,
albedo_unc_avg, albedo_unc_good,
n_tries=10 ):
f = IntProgress(min=0, max=2*n_tries + 1)
f.value = 1
display(f)
# convert uncertainties from percent
albedo_unc = [ albedo_unc_good*0.01, albedo_unc_avg*0.01 ]
retval_s, state, obs = tip_inversion( year, fluxnet_site, albedo_unc=albedo_unc,
green_leaves=green_leaves,
n_tries=n_tries, progressbar=f )
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd_single = np.where(post_sd > c, c, post_sd)
retval, state, obs = regularised_tip_inversion( year, fluxnet_site, [1e-3, 0, 0.1, 1e-3, 0, 0.1, gamma_lai ],
albedo_unc=albedo_unc,
x0=retval_s['real_map'], green_leaves=green_leaves,
n_tries=n_tries, progressbar=f)
mu = state.operators['Prior'].mu
cinv = state.operators['Prior'].inv_cov
c = np.array(np.sqrt(np.linalg.inv (cinv.todense()).diagonal())).squeeze()
post_sd = np.sqrt(np.array(retval_s['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
fig, axs = plt.subplots(nrows=5, ncols=2, figsize=(14, 12))
axs = axs.flatten()
fig.suptitle("%s (%d)" % (fluxnet_site, year), fontsize=18)
params = ['omega_vis', 'd_vis', 'a_vis', 'omega_nir', 'd_nir', 'a_nir', 'lai']
post_sd = np.sqrt(np.array(retval['post_cov'].todense()).squeeze())
post_sd = np.where(post_sd > c, c, post_sd)
for i, p in enumerate(tip_params):
#axs[i].axhspan(mu[(i*46):((i+1)*46)][0]+c[(i*46):((i+1)*46)][0],
# mu[(i*46):((i+1)*46)][0] - c[(i * 46):((i + 1) * 46)][0], color="0.9" )
axs[i].fill_between ( state.state_grid, retval['real_map'][params[i]] - post_sd[(i*46):((i+1)*46)],
retval['real_map'][params[i]] + post_sd[(i*46):((i+1)*46)], lw=0.8, color="0.8")
axs[i].vlines ( state.state_grid, retval_s['real_map'][params[i]] - post_sd_single[(i*46):((i+1)*46)],
retval_s['real_map'][params[i]] + post_sd_single[(i*46):((i+1)*46)], lw=0.8,
colors="0.1", alpha=0.5)
axs[i].plot(state.state_grid, retval['real_map'][params[i]], 'o-', mfc="none")
axs[i].plot(state.state_grid, retval_s['real_map'][params[i]], '--')
if i in [ 1, 4, 6]:
axs[i].set_ylim(0, 6)
else:
axs[i].set_ylim(0, 1)
axs[i].set_ylabel( tip_params[i] )
fwd = np.array(obs.fwd_modelled_obs)
axs[7].plot(obs.observations[:, 0], fwd[:, 0], 'k+', label="VIS")
axs[7].plot(obs.observations[:, 1], fwd[:, 1], 'rx', label="NIR")
axs[7].set_xlabel("Measured BHR [-]")
axs[7].set_ylabel("Predicted BHR [-]")
axs[7].plot ( [0,0.9], [0, 0.9], 'k--', lw=0.5)
axs[7].legend(loc='best')
axs[8].vlines(obs.mask[:, 0], obs.observations[:, 0] - 1.96 * obs.bu[:, 0],
obs.observations[:, 0] + 1.96 * obs.bu[:, 0])
axs[8].plot(obs.mask[:, 0], obs.observations[:, 0], 'o')
axs[9].vlines(obs.mask[:, 0], obs.observations[:, 1] - 1.96 * obs.bu[:, 1],
obs.observations[:, 1] + 1.96 * obs.bu[:, 1])
axs[9].plot(obs.mask[:, 0], obs.observations[:, 1], 'o')
axs[8].set_ylabel("BHR VIS [-]")
axs[9].set_ylabel("BHR NIR [-]")
axs[8].set_xlabel("DoY [d]")
axs[9].set_xlabel("DoY [d]")
for i in xrange(10):
if i != 7:
axs[i].set_xlim(1, 370)
# Hide the right and top spines
axs[i].spines['right'].set_visible(False)
axs[i].spines['top'].set_visible(False)
# Only show ticks on the left and bottom spines
axs[i].yaxis.set_ticks_position('left')
axs[i].xaxis.set_ticks_position('bottom')
fig.figimage(logo, fig.bbox.xmax - 500, fig.bbox.ymax - 250, alpha=.4, zorder=1)
plt.savefig("regularised_model_%s_%04d_%02dpcnt_%02dpcnt.pdf" % (fluxnet_site, year,
int(albedo_unc[0]*100.), int(albedo_unc[1]*100.)),
dpi=300, bbox_inches="tight")