def gwr_mth(self, pt, mth, nnghs=None, stns_rm=None):
'''
Run geographically weighted regression of daily temperature anomalies
for a specific month and point location. The function interpolates
daily anomalies using GWR and then adds the anomalies to the point's
monthly normal and returns the actual daily values.
Parameters
----------
pt : structured array
A structured array containing the point's latitude, longitude,
elevation, topographic dissection index, and average land skin
temperatures for each month. An empty point can be initialized
with build_empty_pt()
nnghs : int, optional
The number of closest neighboring stations to use for the GWR.
If None, nnghs will be set to the optimized number of GWR neighbors
for the point's region.
stns_rm : ndarray or str, optional
An array of station ids or a single station id for stations that
should not be considered neighbors for the specific point.
Returns
----------
interp_vals : ndarray
A 1-D array containing the GWR interpolated actual daily values
for the specified month.
'''
if nnghs == None:
# Get the nnghs to use from the optimal values
# at surrounding stations
nnghs = self.__get_nnghs(pt, mth, stns_rm)
self.stn_slct.set_ngh_stns(pt[LAT],
pt[LON],
nnghs,
load_obs=True,
stns_rm=stns_rm,
obs_mth=mth)
ngh_obs = self.stn_slct.ngh_obs
ngh_stns = self.stn_slct.ngh_stns
ngh_wgt = self.stn_slct.ngh_wgt
ngh_obs_cntr = ngh_obs - ngh_stns[get_norm_varname(mth)]
# Perform a GWR for each day
X = [ngh_stns[avar] for avar in self.mthly_predictors[mth]]
X = np.column_stack(X)
x = [pt[avar] for avar in self.mthly_predictors[mth]]
x = np.array(x)
interp_anom = _gwr_series(X, x, ngh_obs_cntr, ngh_wgt)
# Add interpolated anomalies to monthly norm to get actual values
interp_vals = interp_anom + pt[get_norm_varname(mth)]
return interp_vals
def run_xval(self, stn_id, abw_nngh):
'''
Run leave-one-out cross validations for a specific station id
and a set of different neighbor bandwidths.
Parameters
----------
stn_id : str
Station id for cross validation
abw_nngh : ndarray
A 1-D array of different neighbor station bandwidths
for which to run cross validation.
Returns
----------
err : ndarray
A 12 * N array of cross validations error (modeled - observed)
where N is the number of bandwidths specified by abw_nngh
'''
xval_stn = self.stn_da.stns[self.stn_da.stn_idxs[stn_id]]
err = np.zeros((12, abw_nngh.size))
xvalNorms = np.array(
[xval_stn[get_norm_varname(mth)] for mth in np.arange(1, 13)])
for bw_nngh, x in zip(abw_nngh, np.arange(abw_nngh.size)):
interp_norms = self.krig.krigall(xval_stn,
bw_nngh,
stns_rm=xval_stn[STN_ID])
err[:, x] = interp_norms - xvalNorms
return err
def proc_write(fpath_stndb, elem, fpath_out, nwrkers):
status = MPI.Status()
nwrkrs_done = 0
stn_da = StationSerialDataDb(fpath_stndb, elem)
stn_ids = stn_da.stn_ids
stns = stn_da.stns
stn_mask = np.logical_and(np.isfinite(stn_da.stns[MASK]),
np.isnan(stn_da.stns[BAD]))
days = stn_da.days
stn_da.ds.close()
stn_da = None
print "WRITER: Creating output station netCDF database..."
create_quick_db(fpath_out, stns, days, DB_VARIABLES[elem])
stnda_out = StationSerialDataDb(fpath_out, elem, mode='r+')
mth_names = []
for mth in np.arange(1, 13):
norm_var_name = get_norm_varname(mth)
stnda_out.add_stn_variable(norm_var_name,
'',
units='C',
dtype='f8',
fill_value=netCDF4.default_fillvals['f8'])
mth_names.append(norm_var_name)
stnda_out.ds.sync()
print "WRITER: Output station netCDF database created."
mths = np.arange(12)
stat_chk = StatusCheck(np.sum(stn_mask), 50)
while 1:
stn_id, tair_daily, tair_norms = MPI.COMM_WORLD.recv(
source=MPI.ANY_SOURCE, tag=MPI.ANY_TAG, status=status)
if status.tag == TAG_STOPWORK:
nwrkrs_done += 1
if nwrkrs_done == nwrkers:
print "WRITER: Finished"
return 0
else:
x = np.nonzero(stn_ids == stn_id)[0][0]
stnda_out.ds.variables[elem][:, x] = tair_daily
for i in mths:
stnda_out.ds.variables[mth_names[i]][x] = tair_norms[i]
stnda_out.ds.sync()
stat_chk.increment()
def run_xval(self, stn_id, a_nnghs):
xval_stn = self.stn_da.stns[self.stn_da.stn_idxs[stn_id]]
xval_obs = self.stn_da.load_obs(xval_stn[STN_ID])
biasAll = np.zeros((a_nnghs.size, 12))
maeAll = np.zeros((a_nnghs.size, 12))
r2All = np.zeros((a_nnghs.size, 12))
for x in np.arange(a_nnghs.size):
nnghs = a_nnghs[x]
biasMths = np.zeros(12)
maeMths = np.zeros(12)
r2Mths = np.zeros(12)
for mth in np.arange(1, 13):
xval_anom = xval_obs[self.stn_da.mth_idx[mth]] - xval_stn[
get_norm_varname(mth)]
interp_tair = self.gwr.gwr_mth(xval_stn,
mth,
nnghs,
stns_rm=xval_stn[STN_ID])
interp_anom = interp_tair - xval_stn[get_norm_varname(mth)]
difs = interp_anom - xval_anom
bias = np.mean(difs)
mae = np.mean(np.abs(difs))
r_value = stats.linregress(interp_anom, xval_anom)[2]
r2 = r_value**2 # r-squared value; variance explained
biasMths[mth - 1] = bias
maeMths[mth - 1] = mae
r2Mths[mth - 1] = r2
biasAll[x, :] = biasMths
maeAll[x, :] = maeMths
r2All[x, :] = r2Mths
return biasAll, maeAll, r2All
def krigall(self, pt, nnghs, stns_rm=None):
'''
Run moving window variogram fitting and regression kriging
to interpolate monthly temperature normals to a single point location.
Parameters
----------
pt : structured array
A structured array containing the point's latitude, longitude,
elevation, topographic dissection index, and average land skin
temperatures for each month. An empty point can be initialized
with build_empty_pt()
nnghs : int
The number of neighboring stations to use.
stns_rm : ndarray or str, optional
An array of station ids or a single station id for stations that
should not be considered neighbors for the specific point.
Returns
----------
interp_norms : ndarray
A 1-D array of size 12 containing
the interpolated monthly normals
'''
self.stn_slct.set_ngh_stns(pt[LAT],
pt[LON],
nnghs,
load_obs=False,
stns_rm=stns_rm)
nghs = self.stn_slct.ngh_stns
ngh_lon = ri.FloatSexpVector(nghs[LON])
ngh_lat = ri.FloatSexpVector(nghs[LAT])
ngh_elev = ri.FloatSexpVector(nghs[ELEV])
ngh_tdi = ri.FloatSexpVector(nghs[TDI])
ngh_wgt = ri.FloatSexpVector(self.stn_slct.ngh_wgt)
ngh_dists = ri.FloatSexpVector(self.stn_slct.ngh_dists)
interp_norms = np.zeros(12)
for mth in np.arange(1, 13):
ngh_lst = ri.FloatSexpVector(nghs[get_lst_varname(mth)])
ngh_tair = ri.FloatSexpVector(nghs[get_norm_varname(mth)])
pt_svp = ri.FloatSexpVector((pt[LON], pt[LAT], pt[ELEV], pt[TDI],
pt[get_lst_varname(mth)]))
rslt = self.r_func(pt_svp, ngh_lon, ngh_lat, ngh_elev, ngh_tdi,
ngh_lst, ngh_tair, ngh_wgt, ngh_dists)
interp_norms[mth - 1] = rslt[0]
return interp_norms
def gwr_mth(self, pt, mth, nnghs=None, stns_rm=None):
if nnghs == None:
# Get the nnghs to use from the optimal values at surrounding stations
nnghs = self._GwrTairAnom__get_nnghs(pt, mth, stns_rm)
self.stn_slct.set_ngh_stns(pt[LAT],
pt[LON],
nnghs,
load_obs=True,
stns_rm=stns_rm,
obs_mth=mth)
ngh_obs = self.stn_slct.ngh_obs
ngh_stns = self.stn_slct.ngh_stns
ngh_wgt = self.stn_slct.ngh_wgt
ngh_obs_cntr = ngh_obs - ngh_stns[get_norm_varname(mth)]
a_pt = np.array(
[pt[LON], pt[LAT], pt[ELEV], pt[TDI], pt[get_lst_varname(mth)]])
rslt = r.gwr_anomaly(
robjects.FloatVector(ngh_stns[LON]),
robjects.FloatVector(ngh_stns[LAT]),
robjects.FloatVector(ngh_stns[ELEV]),
robjects.FloatVector(ngh_stns[TDI]),
robjects.FloatVector(ngh_stns[get_lst_varname(mth)]),
robjects.FloatVector(ngh_wgt), robjects.Matrix(ngh_obs_cntr),
robjects.FloatVector(a_pt))
fit_anom = np.array(rslt.rx('fit_anom'))
nrow = np.array(rslt.rx('fit_nrow'))[0]
ncol = np.array(rslt.rx('fit_ncol'))[0]
fit_anom = np.reshape(fit_anom, (nrow, ncol), order='F')
interp_anom = np.array(rslt.rx('pt_anom')).ravel()
interp_vals = interp_anom + pt[get_norm_varname(mth)]
return interp_vals
def interp(self, pt, stns_rm=None):
'''
Interpolate monthly temperature normals and daily
temperatures to a single point location.
Parameters
----------
pt : structured array
A structured array containing the point's latitude, longitude,
elevation, topographic dissection index, and average land skin
temperatures for each month. An empty point can be initialized
with build_empty_pt()
stns_rm : ndarray or str, optional
An array of station ids or a single station id for stations that
should not be considered neighbors for the specific point.
Returns
----------
tair_daily : ndarray
A 1-D array of interpolated daily temperatures.
tair_norms : ndarray
A 1-D array of size 12 with the interpolated
monthly temperature normals.
tair_se : ndarray
A 1-D array of size 12 with the kriging standard
errors for the interpolated monthly temperature
normals.
'''
tair_daily = np.zeros(self.ndays)
tair_norms = np.zeros(12)
tair_se = np.zeros(12)
for mth in np.arange(1, 13):
tair_mean, tair_var = self.krig_tair.krig(pt, mth, stns_rm=stns_rm)
std_err, ci = self.krig_tair.std_err_ci(tair_mean, tair_var)
pt[get_norm_varname(mth)] = tair_mean
tair_norms[mth - 1] = tair_mean
tair_se[mth - 1] = std_err
tair_daily[self.mth_masks[mth]] = self.gwr_tair.gwr_mth(
pt, mth, stns_rm=stns_rm)
return tair_daily, tair_norms, tair_se
def build_empty_pt():
ptDtype = [(LON, np.float64), (LAT, np.float64), (ELEV, np.float64),
(TDI, np.float64), (CLIMDIV, np.float64), (MASK, np.float64)]
ptDtype.extend([("tmin%02d" % mth, np.float64)
for mth in np.arange(1, 13)])
ptDtype.extend([("tmax%02d" % mth, np.float64)
for mth in np.arange(1, 13)])
ptDtype.extend([(get_norm_varname(mth), np.float64)
for mth in np.arange(1, 13)])
ptDtype.extend([(get_optim_varname(mth), np.float64)
for mth in np.arange(1, 13)])
ptDtype.extend([(get_lst_varname(mth), np.float64)
for mth in np.arange(1, 13)])
ptDtype.extend([(get_optim_anom_varname(mth), np.float64)
for mth in np.arange(1, 13)])
a_pt = np.empty(1, dtype=ptDtype)
return a_pt[0]
def gwr_predict(self, pt, mth, nnghs=None, stns_rm=None):
if nnghs is None:
# Get the nnghs to use from the optimal values at surrounding stations
nnghs = self.__get_nnghs(pt, mth, stns_rm)
self.stn_slct.set_ngh_stns(pt[LAT],
pt[LON],
nnghs,
load_obs=False,
stns_rm=stns_rm)
nghs = self.stn_slct.ngh_stns
ngh_lon = ri.FloatSexpVector(nghs[LON])
ngh_lat = ri.FloatSexpVector(nghs[LAT])
ngh_elev = ri.FloatSexpVector(nghs[ELEV])
ngh_tdi = ri.FloatSexpVector(nghs[TDI])
ngh_lst = ri.FloatSexpVector(nghs[get_lst_varname(mth)])
ngh_tair = ri.FloatSexpVector(nghs[get_norm_varname(mth)])
pt_svp = ri.FloatSexpVector(
(pt[LON], pt[LAT], pt[ELEV], pt[TDI], pt[get_lst_varname(mth)]))
ngh_wgt = ri.FloatSexpVector(self.stn_slct.ngh_wgt)
rslt = self.r_gwr_func(ngh_lon, ngh_lat, ngh_elev, ngh_tdi, ngh_lst,
ngh_tair, ngh_wgt, pt_svp)
tair_mean = rslt[0]
tair_var = rslt[1]
bad_interp = rslt[2]
if bad_interp != 0:
print "".join(
["ERROR: ",
str(bad_interp), " bad interp: ",
str(pt)])
return tair_mean, tair_var
def __init__(self, stn_da):
'''
Parameters
----------
stnda : twx.db.StationSerialDataDb
A StationSerialDataDb object pointing to the
database from which observations will be loaded.
'''
self.stn_da = stn_da
mask_stns = np.isnan(self.stn_da.stns[BAD])
self.stn_slct = StationSelect(self.stn_da,
stn_mask=mask_stns,
rm_zero_dist_stns=True)
self.vnames_norm = [get_norm_varname(mth) for mth in np.arange(1, 13)]
self.vnames_lst = [get_lst_varname(mth) for mth in np.arange(1, 13)]
self.df_stns = pd.DataFrame(self.stn_da.stns)
self.df_stns.index = self.df_stns[STN_ID]
# Calculate annual means for monthly LST and Tair normals
self.df_stns['lst'] = self.df_stns[self.vnames_lst].mean(axis=1)
self.df_stns['norm'] = self.df_stns[self.vnames_norm].mean(axis=1)
def krig(self, pt, mth, nnghs=None, vario_params=None, stns_rm=None):
'''
Run moving window regression kriging to interpolate a
temperature normal for a specific month to a single point location.
Parameters
----------
pt : structured array
A structured array containing the point's latitude, longitude,
elevation, topographic dissection index, and average land skin
temperatures for each month. An empty point can be initialized
with build_empty_pt()
mth : int, optional
The month for which to interpolate a temperature normal
nnghs : int, optional
The number of neighboring stations to use. If not provided, nnghs
will be determined from the previously determined optimal number
at surrounding neighboring stations.
vario_params : tuple, optional
A tuple of size 3 (nugget, sill, range). If not provided, the
variogram parameters will be determined from previously fit
variogram parameters at neighboring stations
stns_rm : ndarray or str, optional
An array of station ids or a single station id for stations that
should not be considered neighbors for the specific point.
Returns
----------
tair_mean : float
The interpolated temperature normal.
tair_var : float
The kriging prediction variance for the
temperature normal.
'''
if nnghs is None:
# Get the nnghs to use from the optimal values at surrounding stations
nnghs = self.__get_nnghs(pt, mth, stns_rm)
self.stn_slct.set_ngh_stns(pt[LAT],
pt[LON],
nnghs,
load_obs=False,
stns_rm=stns_rm)
if vario_params is None:
nug, psill, vrange = self.__get_vario_params(pt, mth)
else:
nug, psill, vrange = vario_params
nghs = self.stn_slct.ngh_stns
ngh_lon = ri.FloatSexpVector(nghs[LON])
ngh_lat = ri.FloatSexpVector(nghs[LAT])
ngh_elev = ri.FloatSexpVector(nghs[ELEV])
ngh_tdi = ri.FloatSexpVector(nghs[TDI])
ngh_lst = ri.FloatSexpVector(nghs[get_lst_varname(mth)])
ngh_tair = ri.FloatSexpVector(nghs[get_norm_varname(mth)])
pt_svp = ri.FloatSexpVector(
(pt[LON], pt[LAT], pt[ELEV], pt[TDI], pt[get_lst_varname(mth)]))
nug = ri.FloatSexpVector([nug])
psill = ri.FloatSexpVector([psill])
vrange = ri.FloatSexpVector([vrange])
ngh_wgt = ri.FloatSexpVector(self.stn_slct.ngh_wgt)
rslt = self.r_krig_func(ngh_lon, ngh_lat, ngh_elev, ngh_tdi, ngh_lst,
ngh_tair, ngh_wgt, pt_svp, nug, psill, vrange)
tair_mean = rslt[0]
tair_var = rslt[1]
bad_interp = rslt[2]
if bad_interp != 0:
print "".join(
["ERROR: ",
str(bad_interp), " bad interp: ",
str(pt)])
return tair_mean, tair_var
def get_krig_params(self, pt, mth, rm_stnid=None):
'''
Get the moving window regression kriging variogram
parameters for a specific point and month. Currently
assumes an exponential variogram
Parameters
----------
pt : structured array
A structured array containing the point's latitude, longitude,
elevation, topographic dissection index, and average land skin
temperatures for each month. An empty point can be initialized
with build_empty_pt()
mth : int
The specific month as an integer (1-12)
rm_stnid : ndarray or str, optional
An array of station ids or a single station id for stations that
should not be considered neighbors for the specific point.
Returns
----------
nug : float
Exponential variogram nugget.
psill : float
Exponential variogram partial sill.
rng : float
Exponential variogram range.
'''
# First determine the nnghs to use based on smoothed weighted average of
# the optimal nnghs bandwidth at each station point.
self.stn_slct.set_ngh_stns(pt[LAT],
pt[LON],
DFLT_INIT_NNGHS,
load_obs=False)
indomain_mask = np.isfinite(
self.stn_slct.ngh_stns[get_optim_varname(mth)])
domain_stns = self.stn_slct.ngh_stns[indomain_mask]
if domain_stns.size == 0:
raise Exception(
"Cannot determine the optimal # of neighbors to use!")
n_wgt = self.stn_slct.ngh_wgt[indomain_mask]
nnghs = np.int(
np.round(
np.average(domain_stns[get_optim_varname(mth)],
weights=n_wgt)))
# Now use the optimal nnghs to get the krig params for this mth
self.stn_slct.set_ngh_stns(pt[LAT], pt[LON], nnghs, load_obs=False)
nghs = self.stn_slct.ngh_stns
ngh_lon = ri.FloatSexpVector(nghs[LON])
ngh_lat = ri.FloatSexpVector(nghs[LAT])
ngh_elev = ri.FloatSexpVector(nghs[ELEV])
ngh_tdi = ri.FloatSexpVector(nghs[TDI])
ngh_lst = ri.FloatSexpVector(nghs[get_lst_varname(mth)])
ngh_tair = ri.FloatSexpVector(nghs[get_norm_varname(mth)])
ngh_wgt = ri.FloatSexpVector(self.stn_slct.ngh_wgt)
ngh_dists = ri.FloatSexpVector(self.stn_slct.ngh_dists)
rslt = self.r_func(ngh_lon, ngh_lat, ngh_elev, ngh_tdi, ngh_lst,
ngh_tair, ngh_wgt, ngh_dists)
nug = rslt[0]
psill = rslt[1]
rng = rslt[2]
return nug, psill, rng
