def coc(station,
variable_list,
flag_col,
start,
end,
logfile,
diagnostics=False,
plots=False,
idl=False):
for v, variable in enumerate(variable_list):
st_var = getattr(station, variable)
all_filtered = utils.apply_filter_flags(st_var)
# is this needed 13th Nov 2014 RJHD
#reporting_resolution = utils.reporting_accuracy(utils.apply_filter_flags(st_var))
month_ranges = utils.month_starts_in_pairs(start, end)
month_ranges = month_ranges.reshape(-1, 12, 2)
for month in range(12):
hourly_climatologies = np.zeros(24)
hourly_climatologies.fill(st_var.mdi)
# append all e.g. Januaries together
this_month, year_ids, dummy = utils.concatenate_months(
month_ranges[:, month, :], st_var.data, hours=True)
this_month_filtered, dummy, dummy = utils.concatenate_months(
month_ranges[:, month, :], all_filtered, hours=True)
# if fixed climatology period, sort this here
# get as array of 24 hrs.
this_month = np.ma.array(this_month)
this_month = this_month.reshape(-1, 24)
this_month_filtered = np.ma.array(this_month_filtered)
this_month_filtered = this_month_filtered.reshape(-1, 24)
# get hourly climatology for each month
for hour in range(24):
this_hour = this_month[:, hour]
# need to have data if this is going to work!
if len(this_hour.compressed()) > 0:
# winsorize & climatologies - done to match IDL
if idl:
this_hour = utils.winsorize(np.append(
this_hour.compressed(), -999999),
0.05,
idl=idl)
hourly_climatologies[hour] = np.ma.sum(this_hour) / (
len(this_hour) - 1)
else:
this_hour = utils.winsorize(this_hour.compressed(),
0.05,
idl=idl)
hourly_climatologies[hour] = np.ma.mean(this_hour)
if len(this_month.compressed()) > 0:
# can get stations with few obs in a particular variable.
# anomalise each hour over month appropriately
anomalies = this_month - np.tile(hourly_climatologies,
(this_month.shape[0], 1))
anomalies_filtered = this_month_filtered - np.tile(
hourly_climatologies, (this_month_filtered.shape[0], 1))
if len(anomalies.compressed()) >= 10:
iqr = utils.IQR(anomalies.compressed().reshape(
-1)) / 2. # to match IDL
if iqr < 1.5: iqr = 1.5
else:
iqr = st_var.mdi
normed_anomalies = anomalies / iqr
normed_anomalies_filtered = anomalies_filtered / iqr
# get average anomaly for year
year_ids = np.array(year_ids)
monthly_vqvs = np.ma.zeros(month_ranges.shape[0])
monthly_vqvs.mask = [
False for x in range(month_ranges.shape[0])
]
for year in range(month_ranges.shape[0]):
year_locs = np.where(year_ids == year)
this_year = normed_anomalies_filtered[year_locs, :]
if len(this_year.compressed()) > 0:
# need to have data for this to work!
if idl:
monthly_vqvs[year] = utils.idl_median(
this_year.compressed().reshape(-1))
else:
monthly_vqvs[year] = np.ma.median(this_year)
else:
monthly_vqvs.mask[year] = True
# low pass filter
normed_anomalies = coc_low_pass_filter(normed_anomalies,
year_ids, monthly_vqvs,
month_ranges.shape[0])
# copy from distributional_gap.py - refactor!
# get the threshold value
bins, bincenters = utils.create_bins(normed_anomalies, 1.)
hist, binEdges = np.histogram(normed_anomalies, bins=bins)
gaussian = utils.fit_gaussian(bincenters,
hist,
max(hist),
mu=np.mean(normed_anomalies),
sig=np.std(normed_anomalies))
minimum_threshold = round(
1. + utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian))
if diagnostics:
print iqr, minimum_threshold, 1. + utils.invert_gaussian(
FREQUENCY_THRESHOLD, gaussian)
print gaussian
print hist
if plots:
coc_set_up_plot(bincenters,
hist,
gaussian,
variable,
threshold=minimum_threshold,
sub_par="observations")
uppercount = len(
np.where(normed_anomalies > minimum_threshold)[0])
lowercount = len(
np.where(normed_anomalies < -minimum_threshold)[0])
these_flags = station.qc_flags[:, flag_col[v]]
gap_plot_values, tentative_plot_values = [], []
# find the gaps and apply the flags
gap_start = dgc.dgc_find_gap(hist,
binEdges,
minimum_threshold,
gap_size=1) # in DGC it is 2.
these_flags, gap_plot_values, tentative_plot_values =\
coc_find_and_apply_flags(month_ranges[:,month,:],normed_anomalies, these_flags, year_ids, minimum_threshold, gap_start, \
upper = True, plots = plots, gpv = gap_plot_values, tpv = tentative_plot_values)
gap_start = dgc.dgc_find_gap(hist,
binEdges,
-minimum_threshold,
gap_size=1) # in DGC it is 2.
these_flags, gap_plot_values, tentative_plot_values =\
coc_find_and_apply_flags(month_ranges[:,month,:],normed_anomalies, these_flags, year_ids, minimum_threshold, gap_start, \
upper = False, plots = plots, gpv = gap_plot_values, tpv = tentative_plot_values)
station.qc_flags[:, flag_col[v]] = these_flags
if uppercount + lowercount > 1000:
#print "not sorted spurious stations yet"
pass
if plots:
import matplotlib.pyplot as plt
hist, binEdges = np.histogram(tentative_plot_values,
bins=bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters,
plot_hist,
c='orange',
ls='-',
label='tentative',
where='mid')
hist, binEdges = np.histogram(gap_plot_values, bins=bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters,
plot_hist,
'r-',
label='flagged',
where='mid')
import calendar
plt.text(0.1,
0.9,
calendar.month_name[month + 1],
transform=plt.gca().transAxes)
leg = plt.legend(loc='lower center',
ncol=4,
bbox_to_anchor=(0.5, -0.2),
frameon=False,
prop={'size': 13},
labelspacing=0.15,
columnspacing=0.5)
plt.setp(leg.get_title(), fontsize=14)
plt.show()
#plt.savefig(IMAGELOCATION+'/'+station.id+'_ClimatologicalGap_'+str(month+1)+'.png')
flag_locs = np.where(station.qc_flags[:, flag_col[v]] != 0)
# copy flags into attribute
st_var.flags[flag_locs] = 1
if plots or diagnostics:
utils.print_flagged_obs_number(logfile,
"Climatological",
variable,
len(flag_locs[0]),
noWrite=True)
print "where\n"
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 1)[0])
utils.print_flagged_obs_number(logfile,
" Firm Clim",
variable,
nflags,
noWrite=True)
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 2)[0])
utils.print_flagged_obs_number(logfile,
" Tentative Clim",
variable,
nflags,
noWrite=True)
else:
utils.print_flagged_obs_number(logfile, "Climatological", variable,
len(flag_locs[0]))
logfile.write("where\n")
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 1)[0])
utils.print_flagged_obs_number(logfile, " Firm Clim", variable,
nflags)
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 2)[0])
utils.print_flagged_obs_number(logfile, " Tentative Clim",
variable, nflags)
# firm flags match 030220
station = utils.append_history(station, "Climatological Check")
return
def dgc_all_obs(station,
variable,
flags,
start,
end,
plots=False,
diagnostics=False,
idl=False,
windspeeds=False,
GH=False):
'''RJHD addition working on all observations'''
if plots:
import matplotlib.pyplot as plt
st_var = getattr(station, variable)
month_ranges = utils.month_starts_in_pairs(start, end)
month_ranges = month_ranges.reshape(-1, 12, 2)
all_filtered = utils.apply_filter_flags(st_var)
for month in range(12):
if windspeeds == True:
st_var_wind = getattr(station, "windspeeds")
# get monthly averages
windspeeds_month = np.empty([])
for y, year in enumerate(month_ranges[:, month, :]):
if y == 0:
windspeeds_month = np.ma.array(
st_var_wind.data[year[0]:year[1]])
else:
windspeeds_month = np.ma.concatenate(
[windspeeds_month, st_var_wind.data[year[0]:year[1]]])
windspeeds_month_average = dgc_get_monthly_averages(
windspeeds_month, OBS_LIMIT, st_var_wind.mdi, MEAN)
windspeeds_month_mad = utils.mean_absolute_deviation(
windspeeds_month, median=True)
this_month_data = np.array([])
this_month_filtered = np.array([])
this_month_data, dummy, dummy = utils.concatenate_months(
month_ranges[:, month, :], st_var.data, hours=False)
this_month_filtered, dummy, dummy = utils.concatenate_months(
month_ranges[:, month, :], all_filtered, hours=False)
if len(this_month_filtered.compressed()) > OBS_LIMIT:
if idl:
monthly_median = utils.idl_median(
this_month_filtered.compressed().reshape(-1))
else:
monthly_median = np.ma.median(this_month_filtered)
iqr = utils.IQR(this_month_filtered.compressed())
if iqr == 0.0:
# to get some spread if IQR too small
iqr = utils.IQR(this_month_filtered.compressed(),
percentile=0.05)
print "Spurious_stations file not yet sorted"
if iqr != 0.0:
monthly_values = np.ma.array(
(this_month_data.compressed() - monthly_median) / iqr)
bins, bincenters = utils.create_bins(monthly_values, BIN_SIZE)
dummy, plot_bincenters = utils.create_bins(
monthly_values, BIN_SIZE / 10.)
hist, binEdges = np.histogram(monthly_values, bins=bins)
if GH:
# Use Gauss-Hermite polynomials to add skew and kurtosis to Gaussian fit - January 2015 ^RJHD
initial_values = [
np.max(hist),
np.mean(monthly_values),
np.std(monthly_values),
stats.skew(monthly_values),
stats.kurtosis(monthly_values)
] # norm, mean, std, skew, kurtosis
fit = leastsq(utils.residualsGH, initial_values,
[bincenters, hist,
np.ones(len(hist))])
res = utils.hermite2gauss(fit[0], diagnostics=diagnostics)
plot_gaussian = utils.funcGH(fit[0], plot_bincenters)
# adjust to remove the rising bumps seen in some fits - artefacts of GH fitting?
mid_point = np.argmax(plot_gaussian)
bad, = np.where(
plot_gaussian[mid_point:] < FREQUENCY_THRESHOLD / 10.)
if len(bad) > 0:
plot_gaussian[mid_point:][
bad[0]:] = FREQUENCY_THRESHOLD / 10.
bad, = np.where(
plot_gaussian[:mid_point] < FREQUENCY_THRESHOLD / 10.)
if len(bad) > 0:
plot_gaussian[:mid_point][:bad[
-1]] = FREQUENCY_THRESHOLD / 10.
# extract threshold values
good_values = np.argwhere(
plot_gaussian > FREQUENCY_THRESHOLD)
l_minimum_threshold = round(
plot_bincenters[good_values[0]]) - 1
u_minimum_threshold = 1 + round(
plot_bincenters[good_values[-1]])
else:
gaussian = utils.fit_gaussian(bincenters,
hist,
max(hist),
mu=np.mean(monthly_values),
sig=np.std(monthly_values))
# assume the same threshold value
u_minimum_threshold = 1 + round(
utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian))
l_minimum_threshold = -u_minimum_threshold
plot_gaussian = utils.gaussian(plot_bincenters, gaussian)
if diagnostics:
if GH:
print hist
print res
print iqr, l_minimum_threshold, u_minimum_threshold
else:
print hist
print gaussian
print iqr, u_minimum_threshold, 1. + utils.invert_gaussian(
FREQUENCY_THRESHOLD, gaussian)
if plots:
dgc_set_up_plot(plot_gaussian,
monthly_values,
variable,
threshold=(u_minimum_threshold,
l_minimum_threshold),
sub_par="observations",
GH=GH)
if GH:
plt.figtext(
0.15,
0.67,
'Mean %.2f, S.d. %.2f,\nSkew %.2f, Kurtosis %.2f' %
(res['mean'], res['dispersion'], res['skewness'],
res['kurtosis']),
color='k',
size='small')
uppercount = len(
np.where(monthly_values > u_minimum_threshold)[0])
lowercount = len(
np.where(monthly_values < l_minimum_threshold)[0])
# this needs refactoring - but lots of variables to pass in
if plots or diagnostics: gap_plot_values = np.array([])
if uppercount > 0:
gap_start = dgc_find_gap(hist, binEdges,
u_minimum_threshold)
if gap_start != 0:
for y, year in enumerate(month_ranges[:, month, :]):
this_year_data = np.ma.array(
all_filtered[year[0]:year[1]])
this_year_flags = np.array(flags[year[0]:year[1]])
gap_cleaned_locations = np.where(
((this_year_data - monthly_median) /
iqr) > gap_start)
this_year_flags[gap_cleaned_locations] = 1
flags[year[0]:year[1]] = this_year_flags
if plots or diagnostics:
gap_plot_values = np.append(
gap_plot_values,
(this_year_data[gap_cleaned_locations].
compressed() - monthly_median) / iqr)
if lowercount > 0:
gap_start = dgc_find_gap(hist, binEdges,
l_minimum_threshold)
if gap_start != 0:
for y, year in enumerate(month_ranges[:, month, :]):
this_year_data = np.ma.array(
all_filtered[year[0]:year[1]])
this_year_flags = np.array(flags[year[0]:year[1]])
gap_cleaned_locations = np.where(
np.logical_and(
((this_year_data - monthly_median) / iqr) <
gap_start, this_year_data.mask != True))
this_year_flags[gap_cleaned_locations] = 1
flags[year[0]:year[1]] = this_year_flags
if plots or diagnostics:
gap_plot_values = np.append(
gap_plot_values,
(this_year_data[gap_cleaned_locations].
compressed() - monthly_median) / iqr)
if windspeeds:
this_year_flags[
gap_cleaned_locations] = 2 # tentative flags
slp_average = dgc_get_monthly_averages(
this_month_data, OBS_LIMIT, st_var.mdi,
MEAN)
slp_mad = utils.mean_absolute_deviation(
this_month_data, median=True)
storms = np.where((((windspeeds_month - windspeeds_month_average) / windspeeds_month_mad) > 4.5) &\
(((this_month_data - slp_average) / slp_mad) > 4.5))
if len(storms[0]) >= 2:
storm_1diffs = np.diff(storms)
separations = np.where(storm_1diffs != 1)
#for sep in separations:
if plots:
hist, binEdges = np.histogram(gap_plot_values, bins=bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters,
plot_hist,
'r-',
label='flagged',
where='mid')
import calendar
plt.text(0.1,
0.9,
calendar.month_name[month + 1],
transform=plt.gca().transAxes)
plt.legend(loc='lower center',
ncol=3,
bbox_to_anchor=(0.5, -0.2),
frameon=False,
prop={'size': 13})
plt.show()
#plt.savefig(IMAGELOCATION+'/'+station.id+'_DistributionalGap_'+str(month+1)+'.png')
if diagnostics:
utils.print_flagged_obs_number("",
"Distributional Gap",
variable,
len(gap_plot_values),
noWrite=True)
return flags # dgc_all_obs
def evc(station, variable_list, flag_col, start, end, logfile, diagnostics = False, plots = False, idl = False):
if plots or diagnostics:
import matplotlib.pyplot as plt
import calendar
# very similar to climatological check - ensure that not duplicating
for v, variable in enumerate(variable_list):
st_var = getattr(station, variable)
reporting_resolution = utils.reporting_accuracy(utils.apply_filter_flags(st_var))
reporting_freq = utils.reporting_frequency(utils.apply_filter_flags(st_var))
month_ranges = utils.month_starts_in_pairs(start, end)
month_ranges = month_ranges.reshape(-1,12,2)
month_data_count = np.zeros(month_ranges.shape[0:2])
# for each month
for month in range(12):
# set up hourly climatologies
hourly_clims = np.zeros(24)
hourly_clims.fill(st_var.data.fill_value)
this_month, year_ids, month_data_count[:,month] = utils.concatenate_months(month_ranges[:,month,:], st_var.data, hours = True)
# # extract each year and append together
# year_ids = [] # counter to determine which year each day corresponds to
# for year in range(month_ranges.shape[0]):
# this_year = st_var.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
# if year == 0:
# # store so can access each hour of day separately
# this_month = this_year.reshape(-1,24)
# year_ids = [year for x in range(this_month.shape[0])]
# month_data_count[year,month] = len(this_year.compressed())
# else:
# this_year = this_year.reshape(-1,24)
# this_month = np.ma.concatenate((this_month, this_year), axis = 0)
# year_ids.extend([year for x in range(this_year.shape[0])])
# month_data_count[year,month] = len(this_year.compressed())
# winsorize and get hourly climatology
for h in range(24):
this_hour = this_month[:,h]
if len(this_hour.compressed()) > 100:
# winsorize & climatologies - done to match IDL
if idl:
this_hour_winsorized = utils.winsorize(np.append(this_hour.compressed(), -999999), 0.05, idl = idl)
hourly_clims[h] = np.ma.sum(this_hour_winsorized)/(len(this_hour_winsorized) - 1)
else:
this_hour_winsorized = utils.winsorize(this_hour.compressed(), 0.05, idl = idl)
hourly_clims[h] = np.ma.mean(this_hour_winsorized)
hourly_clims = np.ma.masked_where(hourly_clims == st_var.data.fill_value, hourly_clims)
anomalies = this_month - np.tile(hourly_clims, (this_month.shape[0], 1))
# extract IQR of anomalies (using 1/2 value to match IDL)
if len(anomalies.compressed()) >= 10:
iqr = utils.IQR(anomalies.compressed().reshape(-1)) / 2. # to match IDL
if iqr < 1.5: iqr = 1.5
else:
iqr = st_var.mdi
normed_anomalies = anomalies / iqr
variances = np.ma.zeros(month_ranges.shape[0])
variances.mask = [False for i in range(month_ranges.shape[0])]
rep_accuracies = np.zeros(month_ranges.shape[0])
rep_freqs = np.zeros(month_ranges.shape[0])
variances.fill(st_var.mdi)
rep_accuracies.fill(st_var.mdi)
rep_freqs.fill(st_var.mdi)
year_ids = np.array(year_ids)
# extract variance of normalised anomalies for each year
for y, year in enumerate(range(month_ranges.shape[0])):
year_locs = np.where(year_ids == y)
this_year = normed_anomalies[year_locs,:]
this_year = this_year.reshape(-1)
# end of similarity with Climatological check
if len(this_year.compressed()) >= 30:
variances[y] = utils.mean_absolute_deviation(this_year, median = True)
rep_accuracies[y] = utils.reporting_accuracy(this_year)
rep_freqs[y] = utils.reporting_frequency(this_year)
else:
variances.mask[y] = True
good = np.where(month_data_count[:,month] >= 100)
# get median and IQR of variance for all years for this month
if len(good[0]) >= 10:
median_variance = np.median(variances[good])
iqr_variance = utils.IQR(variances[good]) / 2. # to match IDL
if iqr_variance < 0.01: iqr_variance = 0.01
else:
median_variance = st_var.mdi
iqr_variance = st_var.mdi
# if SLP, then get median and MAD of SLP and windspeed for month
if variable in ["slp", "windspeeds"]:
winds = getattr(station, "windspeeds")
slp = getattr(station, "slp")
# refactor this as similar in style to how target data extracted
for y, year in enumerate(range(month_ranges.shape[0])):
if y == 0:
winds_year = winds.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
winds_month = winds_year.reshape(-1,24)
slp_year = slp.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
slp_month = slp_year.reshape(-1,24)
else:
winds_year = winds.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
winds_year = winds_year.reshape(-1,24)
winds_month = np.ma.concatenate((winds_month, winds_year), axis = 0)
slp_year = slp.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
slp_year = slp_year.reshape(-1,24)
slp_month = np.ma.concatenate((slp_month, slp_year), axis = 0)
median_wind = np.ma.median(winds_month)
median_slp = np.ma.median(slp_month)
wind_MAD = utils.mean_absolute_deviation(winds_month.compressed())
slp_MAD = utils.mean_absolute_deviation(slp_month.compressed())
if diagnostics:
print "median windspeed {} m/s, MAD = {}".format(median_wind, wind_MAD)
print "median slp {} hPa, MAD = {}".format(median_slp, slp_MAD)
# now test to see if variance exceeds expected range
for y, year in enumerate(range(month_ranges.shape[0])):
if (variances[y] != st_var.mdi) and (iqr_variance != st_var.mdi) and \
(median_variance != st_var.mdi) and (month_data_count[y,month] >= DATA_COUNT_THRESHOLD):
# if SLP, then need to test if deep low pressure ("hurricane/storm") present
# as this will increase the variance for this month + year
if variable in ["slp", "windspeeds"]:
iqr_threshold = 6.
# increase threshold if reporting frequency and resolution of this
# year doesn't match average
if (rep_accuracies[y] != reporting_resolution) and \
(rep_freqs[y] != reporting_freq):
iqr_threshold = 8.
if diagnostics:
print np.abs(variances[y] - median_variance) / iqr_variance, variances[y] , median_variance , iqr_variance , iqr_threshold, month+1, year+start.year
if np.abs((variances[y] - median_variance) / iqr_variance) > iqr_threshold:
# check for storms
winds_month = winds.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
slp_month = slp.data[month_ranges[year,month][0]:month_ranges[year,month][1]]
storm = False
if (len(winds_month.compressed()) >= 1) and (len(slp_month.compressed()) >= 1):
# find max wind & min SLP
# max_wind_loc = np.where(winds_month == np.max(winds_month))[0][0]
# min_slp_loc = np.where(slp_month == np.min(slp_month))[0][0]
# if these are above thresholds and within one day of each other,
# then it likely was a storm
# print "fix this in case of multiple max/min locations"
# if (np.abs(max_wind_loc - min_slp_loc) <= 24) and \
# (((np.max(winds_month) - median_wind) / wind_MAD) > MAD_THRESHOLD) and \
# (((median_slp - np.min(slp_month)) / slp_MAD) > MAD_THRESHOLD):
# locations where winds greater than threshold
high_winds, = np.where((winds_month - median_wind)/wind_MAD > MAD_THRESHOLD)
# and where SLP less than threshold
low_slps, = np.where((median_slp - slp_month)/slp_MAD > MAD_THRESHOLD)
# if any locations match, then it's a storm
match_loc = high_winds[np.in1d(high_winds, low_slps)]
if len(match_loc) > 0:
storm = True
else:
print "write spurious"
# check the SLP first difference series
# to ensure a drop down and climb out of minimum SLP/or climb up and down from maximum wind speed
if variable == "slp":
diffs = np.diff(slp_month.compressed())
elif variable == "windspeeds":
diffs = np.diff(winds_month.compressed())
negs, poss = 0,0
biggest_neg, biggest_pos = 0,0
for diff in diffs:
if diff > 0:
if negs > biggest_neg: biggest_neg = negs
negs = 0
poss += 1
else:
if poss > biggest_pos: biggest_pos = poss
poss = 0
negs += 1
if (biggest_neg < 10) and (biggest_pos < 10) and not storm:
# not a hurricane, so mask
station.qc_flags[month_ranges[year,month,0]:month_ranges[year,month,1], flag_col[v]] = 1
if plots or diagnostics:
print "No Storm or Hurricane in %i %i - flagging\n" % (month+1, y+start.year)
else:
logfile.write("No Storm or Hurricane in %i %i - flagging\n" % (month+1, y+start.year))
else:
# hurricane
if plots or diagnostics:
print "Storm or Hurricane in %i %i - not flagging\n" % (month+1, y+start.year)
else:
logfile.write("Storm or Hurricane in %i %i - not flagging\n" % (month+1, y+start.year))
if plots:
# plot showing the pressure, pressure first differences and the wind speeds
plot_times = utils.times_hours_to_datetime(station.time.data[month_ranges[year,month][0]:month_ranges[year,month][1]], start)
evc_plot_slp_wind(plot_times, slp_month, diffs, median_slp, slp_MAD, winds_month, median_wind, wind_MAD)
else:
iqr_threshold = 8.
if (rep_accuracies[y] != reporting_resolution) and \
(rep_freqs[y] != reporting_freq):
iqr_threshold = 10.
if np.abs(variances[y] - median_variance) / iqr_variance > iqr_threshold:
if diagnostics:
print "flagging {} {}".format(year+start.year,calendar.month_name[month+1])
# remove the data
station.qc_flags[month_ranges[year,month,0]:month_ranges[year,month,1], flag_col[v]] = 1
if plots:
plot_variances = (variances - median_variance) / iqr_variance
plot_variances = np.ma.masked_where(month_data_count[:,month] < DATA_COUNT_THRESHOLD,plot_variances)
evc_plot_hist(plot_variances, iqr_threshold, "Variance Check - %s - %s" % (variable, calendar.month_name[month+1]))
flag_locs = np.where(station.qc_flags[:, flag_col[v]] != 0)
if plots or diagnostics:
utils.print_flagged_obs_number(logfile, "Variance", variable, len(flag_locs[0]), noWrite = True)
else:
utils.print_flagged_obs_number(logfile, "Variance", variable, len(flag_locs[0]))
# copy flags into attribute
st_var.flags[flag_locs] = 1
# matches 030660 for T, D and SLP 21/8/2014
station = utils.append_history(station, "Excess Variance Check")
return # evc
def coc(station, variable_list, flag_col, start, end, logfile, diagnostics = False, plots = False, idl = False):
for v, variable in enumerate(variable_list):
st_var = getattr(station, variable)
all_filtered = utils.apply_filter_flags(st_var)
# is this needed 13th Nov 2014 RJHD
#reporting_resolution = utils.reporting_accuracy(utils.apply_filter_flags(st_var))
month_ranges = utils.month_starts_in_pairs(start, end)
month_ranges = month_ranges.reshape(-1,12,2)
for month in range(12):
hourly_climatologies = np.zeros(24)
hourly_climatologies.fill(st_var.mdi)
# append all e.g. Januaries together
this_month, year_ids, dummy = utils.concatenate_months(month_ranges[:,month,:], st_var.data, hours = True)
this_month_filtered, dummy, dummy = utils.concatenate_months(month_ranges[:,month,:], all_filtered, hours = True)
# if fixed climatology period, sort this here
# get as array of 24 hrs.
this_month = np.ma.array(this_month)
this_month = this_month.reshape(-1,24)
this_month_filtered = np.ma.array(this_month_filtered)
this_month_filtered = this_month_filtered.reshape(-1,24)
# get hourly climatology for each month
for hour in range(24):
this_hour = this_month[:,hour]
# need to have data if this is going to work!
if len(this_hour.compressed()) > 0:
# winsorize & climatologies - done to match IDL
if idl:
this_hour = utils.winsorize(np.append(this_hour.compressed(), -999999), 0.05, idl = idl)
hourly_climatologies[hour] = np.ma.sum(this_hour)/(len(this_hour) - 1)
else:
this_hour = utils.winsorize(this_hour.compressed(), 0.05, idl = idl)
hourly_climatologies[hour] = np.ma.mean(this_hour)
if len(this_month.compressed()) > 0:
# can get stations with few obs in a particular variable.
# anomalise each hour over month appropriately
anomalies = this_month - np.tile(hourly_climatologies, (this_month.shape[0],1))
anomalies_filtered = this_month_filtered - np.tile(hourly_climatologies, (this_month_filtered.shape[0],1))
if len(anomalies.compressed()) >= 10:
iqr = utils.IQR(anomalies.compressed().reshape(-1))/2. # to match IDL
if iqr < 1.5: iqr = 1.5
else:
iqr = st_var.mdi
normed_anomalies = anomalies / iqr
normed_anomalies_filtered = anomalies_filtered / iqr
# get average anomaly for year
year_ids = np.array(year_ids)
monthly_vqvs = np.ma.zeros(month_ranges.shape[0])
monthly_vqvs.mask = [False for x in range(month_ranges.shape[0])]
for year in range(month_ranges.shape[0]):
year_locs = np.where(year_ids == year)
this_year = normed_anomalies_filtered[year_locs,:]
if len(this_year.compressed()) > 0:
# need to have data for this to work!
if idl:
monthly_vqvs[year] = utils.idl_median(this_year.compressed().reshape(-1))
else:
monthly_vqvs[year] = np.ma.median(this_year)
else:
monthly_vqvs.mask[year] = True
# low pass filter
normed_anomalies = coc_low_pass_filter(normed_anomalies, year_ids, monthly_vqvs, month_ranges.shape[0])
# copy from distributional_gap.py - refactor!
# get the threshold value
bins, bincenters = utils.create_bins(normed_anomalies, 1.)
hist, binEdges = np.histogram(normed_anomalies, bins = bins)
gaussian = utils.fit_gaussian(bincenters, hist, max(hist), mu=np.mean(normed_anomalies), sig = np.std(normed_anomalies))
minimum_threshold = round(1. + utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian))
if diagnostics:
print iqr, minimum_threshold, 1. + utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian)
print gaussian
print hist
if plots:
coc_set_up_plot(bincenters, hist, gaussian, variable, threshold = minimum_threshold, sub_par = "observations")
uppercount = len(np.where(normed_anomalies > minimum_threshold)[0])
lowercount = len(np.where(normed_anomalies < -minimum_threshold)[0])
these_flags = station.qc_flags[:, flag_col[v]]
gap_plot_values, tentative_plot_values = [], []
# find the gaps and apply the flags
gap_start = dgc.dgc_find_gap(hist, binEdges, minimum_threshold, gap_size = 1) # in DGC it is 2.
these_flags, gap_plot_values, tentative_plot_values =\
coc_find_and_apply_flags(month_ranges[:,month,:],normed_anomalies, these_flags, year_ids, minimum_threshold, gap_start, \
upper = True, plots = plots, gpv = gap_plot_values, tpv = tentative_plot_values)
gap_start = dgc.dgc_find_gap(hist, binEdges, -minimum_threshold, gap_size = 1) # in DGC it is 2.
these_flags, gap_plot_values, tentative_plot_values =\
coc_find_and_apply_flags(month_ranges[:,month,:],normed_anomalies, these_flags, year_ids, minimum_threshold, gap_start, \
upper = False, plots = plots, gpv = gap_plot_values, tpv = tentative_plot_values)
station.qc_flags[:, flag_col[v]] = these_flags
if uppercount + lowercount > 1000:
#print "not sorted spurious stations yet"
pass
if plots:
import matplotlib.pyplot as plt
hist, binEdges = np.histogram(tentative_plot_values, bins = bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters, plot_hist, c='orange', ls='-', label = 'tentative', where='mid')
hist, binEdges = np.histogram(gap_plot_values, bins = bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters, plot_hist, 'r-', label = 'flagged', where='mid')
import calendar
plt.text(0.1,0.9,calendar.month_name[month+1], transform = plt.gca().transAxes)
leg=plt.legend(loc='lower center',ncol=4, bbox_to_anchor=(0.5,-0.2),frameon=False,prop={'size':13},labelspacing=0.15,columnspacing=0.5)
plt.setp(leg.get_title(), fontsize=14)
plt.show()
#plt.savefig(IMAGELOCATION+'/'+station.id+'_ClimatologicalGap_'+str(month+1)+'.png')
flag_locs = np.where(station.qc_flags[:, flag_col[v]] != 0)
# copy flags into attribute
st_var.flags[flag_locs] = 1
if plots or diagnostics:
utils.print_flagged_obs_number(logfile, "Climatological", variable, len(flag_locs[0]), noWrite = True)
print "where\n"
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 1)[0])
utils.print_flagged_obs_number(logfile, " Firm Clim", variable, nflags, noWrite = True)
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 2)[0])
utils.print_flagged_obs_number(logfile, " Tentative Clim", variable, nflags, noWrite = True)
else:
utils.print_flagged_obs_number(logfile, "Climatological", variable, len(flag_locs[0]))
logfile.write("where\n")
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 1)[0])
utils.print_flagged_obs_number(logfile, " Firm Clim", variable, nflags)
nflags = len(np.where(station.qc_flags[:, flag_col[v]] == 2)[0])
utils.print_flagged_obs_number(logfile, " Tentative Clim", variable, nflags)
# firm flags match 030220
station = utils.append_history(station, "Climatological Check")
return
def detect(station,
neighbour,
variable,
flags,
neighbour_count,
start,
end,
distance=0,
diagnostics=False,
plots=False):
'''
Detect which observations are outliers
:param MetVar station: station object (target)
:param MetVar neighbour: station object (neighbour)
:param string variable: which variable to process
:param array flags: array to store how many neighbours thing each obs is bad
:param array neighbour_count: how many neighbours present at each obs
:param datetime start: start of dataset
:param datetime end: end of dataset
:param int distance: separation of target and neighbour
:param bool diagnostics: extra output
:param bool plots: make figures
:returns: None
'''
FILTERING_FLAG_COL = {
"temperatures": [0, 1, 4, 5, 8, 12, 16, 20, 27, 41, 44, 58],
"dewpoints":
[0, 2, 4, 6, 8, 9, 13, 17, 21, 28, 30, 31, 32, 42, 45, 59],
"slp": [0, 3, 4, 7, 11, 15, 19, 23, 29, 43, 46, 60],
"windspeeds": [0, 4, 10, 14, 18, 22, 56, 62, 63, 64]
} # not used, but ready for it.
st_var = getattr(station, variable)
neigh_var = getattr(neighbour, variable)
# filter by flags - not all (no Climatological [24,25], or Odd cluster [54,55,56,57]), T record check not in D,
total_flags = np.sum(station.qc_flags[:, FILTERING_FLAG_COL[variable]],
axis=1)
st_filtered = np.ma.masked_where(total_flags == 1, st_var.data)
neigh_filtered = np.ma.masked_where(total_flags == 1, neigh_var.data)
# match the observation times
match = np.where(
np.logical_and((st_filtered.data != st_var.mdi),
(neigh_filtered.data != neigh_var.mdi)))
month_ranges = utils.month_starts_in_pairs(start, end).reshape(
-1, 12, 2) # in year-long sets of pairs.
if len(match[0]) >= 100:
neighbour_count[match] += 1 # number of neighbours with data present
differences = np.ma.zeros(len(st_filtered))
differences.fill(st_var.mdi)
differences.mask = True
differences[
match] = st_filtered.data[match] - neigh_filtered.data[match]
differences.mask[match] = False
all_iqrs = np.zeros(len(differences))
# get monthly IQR values
for month in range(12):
this_month, dummy1, dummy2 = utils.concatenate_months(
month_ranges[:, month, :], differences, hours=False)
if len(this_month.compressed()) > 4:
iqr = utils.IQR(this_month.compressed())
if iqr <= 2.: iqr = 2.
else:
iqr = 2.
# and copy back into the array
for year in month_ranges[:, month, :]:
all_iqrs[year[0]:year[1]] = iqr
if plots:
plot_target_neigh_diffs_dist(differences, min(all_iqrs))
dubious = np.ma.where(np.ma.abs(differences) > 5. * all_iqrs)
if len(dubious[0]) >= 1.:
if variable == "slp":
# check if they are storms
positive = np.ma.where(differences > 5. * iqr)
negative = np.ma.where(differences < -5. * iqr)
# if majority negative (2/3) and separation > 100
if (distance > 100.) and (
float(len(positive[0])) / len(dubious[0]) < 0.333):
if len(positive[0]) > 0:
flags[positive] += 1
if len(negative[0]) > 0:
neighbour_count[match] -= 1
else:
flags[dubious] += 1
else:
flags[dubious] += 1
return # detect
def dgc_all_obs(station, variable, flags, start, end, plots = False, diagnostics = False, idl = False, windspeeds = False, GH = False):
'''RJHD addition working on all observations'''
if plots:
import matplotlib.pyplot as plt
st_var = getattr(station, variable)
month_ranges = utils.month_starts_in_pairs(start, end)
month_ranges = month_ranges.reshape(-1,12,2)
all_filtered = utils.apply_filter_flags(st_var)
for month in range(12):
if windspeeds == True:
st_var_wind = getattr(station, "windspeeds")
# get monthly averages
windspeeds_month = np.empty([])
for y, year in enumerate(month_ranges[:,month,:]):
if y == 0:
windspeeds_month = np.ma.array(st_var_wind.data[year[0]:year[1]])
else:
windspeeds_month = np.ma.concatenate([windspeeds_month, st_var_wind.data[year[0]:year[1]]])
windspeeds_month_average = dgc_get_monthly_averages(windspeeds_month, OBS_LIMIT, st_var_wind.mdi, MEAN)
windspeeds_month_mad = utils.mean_absolute_deviation(windspeeds_month, median=True)
this_month_data = np.array([])
this_month_filtered = np.array([])
this_month_data, dummy, dummy = utils.concatenate_months(month_ranges[:,month,:], st_var.data, hours = False)
this_month_filtered, dummy, dummy = utils.concatenate_months(month_ranges[:,month,:], all_filtered, hours = False)
if len(this_month_filtered.compressed()) > OBS_LIMIT:
if idl:
monthly_median = utils.idl_median(this_month_filtered.compressed().reshape(-1))
else:
monthly_median = np.ma.median(this_month_filtered)
iqr = utils.IQR(this_month_filtered.compressed())
if iqr == 0.0:
# to get some spread if IQR too small
iqr = utils.IQR(this_month_filtered.compressed(), percentile = 0.05)
print "Spurious_stations file not yet sorted"
if iqr != 0.0:
monthly_values = np.ma.array((this_month_data.compressed() - monthly_median) / iqr)
bins, bincenters = utils.create_bins(monthly_values, BIN_SIZE)
dummy, plot_bincenters = utils.create_bins(monthly_values, BIN_SIZE/10.)
hist, binEdges = np.histogram(monthly_values, bins = bins)
if GH:
# Use Gauss-Hermite polynomials to add skew and kurtosis to Gaussian fit - January 2015 ^RJHD
initial_values = [np.max(hist), np.mean(monthly_values), np.std(monthly_values), stats.skew(monthly_values), stats.kurtosis(monthly_values)] # norm, mean, std, skew, kurtosis
fit = leastsq(utils.residualsGH, initial_values, [bincenters, hist, np.ones(len(hist))])
res = utils.hermite2gauss(fit[0], diagnostics = diagnostics)
plot_gaussian = utils.funcGH(fit[0], plot_bincenters)
# adjust to remove the rising bumps seen in some fits - artefacts of GH fitting?
mid_point = np.argmax(plot_gaussian)
bad, = np.where(plot_gaussian[mid_point:] < FREQUENCY_THRESHOLD/10.)
if len(bad) > 0: plot_gaussian[mid_point:][bad[0]:] = FREQUENCY_THRESHOLD/10.
bad, = np.where(plot_gaussian[:mid_point] < FREQUENCY_THRESHOLD/10.)
if len(bad) > 0: plot_gaussian[:mid_point][:bad[-1]] = FREQUENCY_THRESHOLD/10.
# extract threshold values
good_values = np.argwhere(plot_gaussian > FREQUENCY_THRESHOLD)
l_minimum_threshold = round(plot_bincenters[good_values[0]]) - 1
u_minimum_threshold = 1 + round(plot_bincenters[good_values[-1]])
else:
gaussian = utils.fit_gaussian(bincenters, hist, max(hist), mu = np.mean(monthly_values), sig = np.std(monthly_values))
# assume the same threshold value
u_minimum_threshold = 1 + round(utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian))
l_minimum_threshold = -u_minimum_threshold
plot_gaussian = utils.gaussian(plot_bincenters, gaussian)
if diagnostics:
if GH:
print hist
print res
print iqr, l_minimum_threshold, u_minimum_threshold
else:
print hist
print gaussian
print iqr, u_minimum_threshold, 1. + utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian)
if plots:
dgc_set_up_plot(plot_gaussian, monthly_values, variable, threshold = (u_minimum_threshold, l_minimum_threshold), sub_par = "observations", GH = GH)
if GH:
plt.figtext(0.15, 0.67, 'Mean %.2f, S.d. %.2f,\nSkew %.2f, Kurtosis %.2f' %(res['mean'], res['dispersion'], res['skewness'], res['kurtosis']), color='k', size='small')
uppercount = len(np.where(monthly_values > u_minimum_threshold)[0])
lowercount = len(np.where(monthly_values < l_minimum_threshold)[0])
# this needs refactoring - but lots of variables to pass in
if plots or diagnostics: gap_plot_values = np.array([])
if uppercount > 0:
gap_start = dgc_find_gap(hist, binEdges, u_minimum_threshold)
if gap_start != 0:
for y, year in enumerate(month_ranges[:,month,:]):
this_year_data = np.ma.array(all_filtered[year[0]:year[1]])
this_year_flags = np.array(flags[year[0]:year[1]])
gap_cleaned_locations = np.where(((this_year_data - monthly_median) / iqr) > gap_start)
this_year_flags[gap_cleaned_locations] = 1
flags[year[0]:year[1]] = this_year_flags
if plots or diagnostics: gap_plot_values = np.append(gap_plot_values, (this_year_data[gap_cleaned_locations].compressed() - monthly_median)/iqr)
if lowercount > 0:
gap_start = dgc_find_gap(hist, binEdges, l_minimum_threshold)
if gap_start != 0:
for y, year in enumerate(month_ranges[:,month,:]):
this_year_data = np.ma.array(all_filtered[year[0]:year[1]])
this_year_flags = np.array(flags[year[0]:year[1]])
gap_cleaned_locations = np.where(np.logical_and(((this_year_data - monthly_median) / iqr) < gap_start, this_year_data.mask != True))
this_year_flags[gap_cleaned_locations] = 1
flags[year[0]:year[1]] = this_year_flags
if plots or diagnostics: gap_plot_values = np.append(gap_plot_values, (this_year_data[gap_cleaned_locations].compressed() - monthly_median)/iqr)
if windspeeds:
this_year_flags[gap_cleaned_locations] = 2 # tentative flags
slp_average = dgc_get_monthly_averages(this_month_data, OBS_LIMIT, st_var.mdi, MEAN)
slp_mad = utils.mean_absolute_deviation(this_month_data, median=True)
storms = np.where((((windspeeds_month - windspeeds_month_average) / windspeeds_month_mad) > 4.5) &\
(((this_month_data - slp_average) / slp_mad) > 4.5))
if len(storms[0]) >= 2:
storm_1diffs = np.diff(storms)
separations = np.where(storm_1diffs != 1)
#for sep in separations:
if plots:
hist, binEdges = np.histogram(gap_plot_values, bins = bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters, plot_hist, 'r-', label = 'flagged', where='mid')
import calendar
plt.text(0.1,0.9,calendar.month_name[month+1], transform = plt.gca().transAxes)
plt.legend(loc='lower center',ncol=3, bbox_to_anchor=(0.5,-0.2),frameon=False,prop={'size':13})
plt.show()
#plt.savefig(IMAGELOCATION+'/'+station.id+'_DistributionalGap_'+str(month+1)+'.png')
if diagnostics:
utils.print_flagged_obs_number("", "Distributional Gap", variable, len(gap_plot_values), noWrite=True)
return flags # dgc_all_obs
def dgc_all_obs(station,
variable,
flags,
start,
end,
logfile,
plots=False,
diagnostics=False,
idl=False,
windspeeds=False,
GH=False,
doMonth=False):
'''RJHD addition working on all observations'''
if plots:
import matplotlib.pyplot as plt
month_ranges = utils.month_starts_in_pairs(start, end)
month_ranges = month_ranges.reshape(-1, 12, 2)
# extract variable
st_var = getattr(station, variable)
# apply flags (and mask incomplete year if appropriate)
all_filtered = utils.apply_filter_flags(st_var,
doMonth=doMonth,
start=start,
end=end)
st_var_complete_year = copy.deepcopy(st_var)
if doMonth:
# restrict the incomplete year if appropriate - keep other flagged obs.
full_year_end = utils.get_first_hour_this_year(start, end)
st_var_complete_year.data.mask[full_year_end:] = True
for month in range(12):
# if requiring wind data, extract data and find monthly averages
if windspeeds == True:
st_var_wind = getattr(station, "windspeeds")
if doMonth:
# restrict the incomplete year if appropriate
st_var_wind.data.mask[full_year_end:] = True
# get monthly averages
windspeeds_month = np.empty([])
for y, year in enumerate(month_ranges[:, month, :]):
if y == 0:
windspeeds_month = np.ma.array(
st_var_wind.data[year[0]:year[1]])
else:
windspeeds_month = np.ma.concatenate(
[windspeeds_month, st_var_wind.data[year[0]:year[1]]])
windspeeds_month_average = dgc_get_monthly_averages(
windspeeds_month, OBS_LIMIT, st_var_wind.mdi, MEAN)
windspeeds_month_mad = utils.mean_absolute_deviation(
windspeeds_month, median=True)
# pull data from each calendar month together
this_month_data, dummy, dummy = utils.concatenate_months(
month_ranges[:, month, :], st_var.data, hours=False)
this_month_filtered, dummy, dummy = utils.concatenate_months(
month_ranges[:, month, :], all_filtered, hours=False)
this_month_complete, dummy, dummy = utils.concatenate_months(
month_ranges[:, month, :], st_var_complete_year.data, hours=False)
# if enough clean and complete data for this calendar month find the median and IQR
if len(this_month_filtered.compressed()) > OBS_LIMIT:
if idl:
monthly_median = utils.idl_median(
this_month_filtered.compressed().reshape(-1))
else:
monthly_median = np.ma.median(this_month_filtered)
iqr = utils.IQR(this_month_filtered.compressed())
if iqr == 0.0:
# to get some spread if IQR too small
iqr = utils.IQR(this_month_filtered.compressed(),
percentile=0.05)
print "Spurious_stations file not yet sorted"
# if have an IQR, anomalise using median and standardise using IQR
if iqr != 0.0:
monthly_values = np.ma.array(
(this_month_data.compressed() - monthly_median) / iqr)
complete_values = np.ma.array(
(this_month_complete.compressed() - monthly_median) / iqr)
# use complete years only for the histogram - aiming to find outliers.
bins, bincenters = utils.create_bins(complete_values, BIN_SIZE)
dummy, plot_bincenters = utils.create_bins(
complete_values, BIN_SIZE / 10.)
hist, binEdges = np.histogram(complete_values, bins=bins)
"""
Change to monthly updates Oct 2017
Thought about changing distribution to use filtered values
But this changes the test beyond just dealing with additional months
Commented out lines below would be alternative.
"""
# bins, bincenters = utils.create_bins(filtered_values, BIN_SIZE)
# dummy, plot_bincenters = utils.create_bins(filtered_values, BIN_SIZE/10.)
# hist, binEdges = np.histogram(filtered_values, bins = bins)
# used filtered (incl. incomplete year mask) to determine the distribution.
if GH:
# Use Gauss-Hermite polynomials to add skew and kurtosis to Gaussian fit - January 2015 ^RJHD
# Feb 2019 - if large amounts off centre, can affect initial values
# switched to median and MAD
initial_values = [
np.max(hist),
np.median(complete_values),
utils.mean_absolute_deviation(complete_values,
median=True),
stats.skew(complete_values),
stats.kurtosis(complete_values)
] # norm, mean, std, skew, kurtosis
fit = leastsq(utils.residualsGH, initial_values,
[bincenters, hist,
np.ones(len(hist))])
res = utils.hermite2gauss(fit[0], diagnostics=diagnostics)
plot_gaussian = utils.funcGH(fit[0], plot_bincenters)
# adjust to remove the rising bumps seen in some fits - artefacts of GH fitting?
mid_point = np.argmax(plot_gaussian)
bad, = np.where(
plot_gaussian[mid_point:] < FREQUENCY_THRESHOLD / 10.)
if len(bad) > 0:
plot_gaussian[mid_point:][
bad[0]:] = FREQUENCY_THRESHOLD / 10.
bad, = np.where(
plot_gaussian[:mid_point] < FREQUENCY_THRESHOLD / 10.)
if len(bad) > 0:
plot_gaussian[:mid_point][:bad[
-1]] = FREQUENCY_THRESHOLD / 10.
# extract threshold values
good_values = np.argwhere(
plot_gaussian > FREQUENCY_THRESHOLD)
l_minimum_threshold = round(
plot_bincenters[good_values[0]]) - 1
u_minimum_threshold = 1 + round(
plot_bincenters[good_values[-1]])
if diagnostics:
print hist
print res
print iqr, l_minimum_threshold, u_minimum_threshold
# or just a standard Gaussian
else:
gaussian = utils.fit_gaussian(
bincenters,
hist,
max(hist),
mu=np.median(complete_values),
sig=utils.mean_absolute_value(complete_values))
# assume the same threshold value
u_minimum_threshold = 1 + round(
utils.invert_gaussian(FREQUENCY_THRESHOLD, gaussian))
l_minimum_threshold = -u_minimum_threshold
plot_gaussian = utils.gaussian(plot_bincenters, gaussian)
if diagnostics:
print hist
print gaussian
print iqr, u_minimum_threshold, 1. + utils.invert_gaussian(
FREQUENCY_THRESHOLD, gaussian)
if plots:
dgc_set_up_plot(plot_gaussian,
complete_values,
variable,
threshold=(u_minimum_threshold,
l_minimum_threshold),
sub_par="observations",
GH=GH)
if GH:
plt.figtext(
0.15,
0.67,
'Mean %.2f, S.d. %.2f,\nSkew %.2f, Kurtosis %.2f' %
(res['mean'], res['dispersion'], res['skewness'],
res['kurtosis']),
color='k',
size='small')
# now trying to find gaps in the distribution
uppercount = len(
np.where(monthly_values > u_minimum_threshold)[0])
lowercount = len(
np.where(monthly_values < l_minimum_threshold)[0])
# this needs refactoring - but lots of variables to pass in
if plots or diagnostics: gap_plot_values = np.array([])
# do one side of distribution and then other
if uppercount > 0:
gap_start = dgc_find_gap(hist, binEdges,
u_minimum_threshold)
if gap_start != 0:
# if found a gap, then go through each year for this calendar month
# and flag observations further from middle
for y, year in enumerate(month_ranges[:, month, :]):
# not using filtered - checking all available data
this_year_data = np.ma.array(
st_var.data[year[0]:year[1]])
this_year_flags = np.array(flags[year[0]:year[1]])
gap_cleaned_locations = np.ma.where(
((this_year_data - monthly_median) /
iqr) > gap_start)
this_year_flags[gap_cleaned_locations] = 1
flags[year[0]:year[1]] = this_year_flags
if plots or diagnostics:
gap_plot_values = np.append(
gap_plot_values,
(this_year_data[gap_cleaned_locations].
compressed() - monthly_median) / iqr)
if len(gap_cleaned_locations[0]) > 0:
print "Upper {}-{} - {} obs flagged".format(
y + start.year, month,
len(gap_cleaned_locations[0]))
print gap_cleaned_locations, this_year_data[
gap_cleaned_locations]
if lowercount > 0:
gap_start = dgc_find_gap(hist, binEdges,
l_minimum_threshold)
if gap_start != 0:
# if found a gap, then go through each year for this calendar month
# and flag observations further from middle
for y, year in enumerate(month_ranges[:, month, :]):
this_year_data = np.ma.array(
st_var.data[year[0]:year[1]])
this_year_flags = np.array(flags[year[0]:year[1]])
gap_cleaned_locations = np.ma.where(
np.logical_and(
((this_year_data - monthly_median) / iqr) <
gap_start, this_year_data.mask != True))
# add flag requirement for low pressure bit if appropriate
this_year_flags[gap_cleaned_locations] = 1
flags[year[0]:year[1]] = this_year_flags
if plots or diagnostics:
gap_plot_values = np.append(
gap_plot_values,
(this_year_data[gap_cleaned_locations].
compressed() - monthly_median) / iqr)
if len(gap_cleaned_locations[0]) > 0:
print "Lower {}-{} - {} obs flagged".format(
y + start.year, month,
len(gap_cleaned_locations[0]))
print gap_cleaned_locations, this_year_data[
gap_cleaned_locations]
# if doing SLP then do extra checks for storms
if windspeeds:
windspeeds_year = np.ma.array(
st_var_wind.data[year[0]:year[1]])
this_year_flags[
gap_cleaned_locations] = 2 # tentative flags
slp_average = dgc_get_monthly_averages(
this_month_data, OBS_LIMIT, st_var.mdi,
MEAN)
slp_mad = utils.mean_absolute_deviation(
this_month_data, median=True)
# need to ensure that this_year_data is less than slp_average, hence order of test
storms, = np.ma.where((((windspeeds_year - windspeeds_month_average) / windspeeds_month_mad) > MAD_THRESHOLD) &\
(((slp_average - this_year_data) / slp_mad) > MAD_THRESHOLD))
# using IDL terminology
if len(storms) >= 2:
# use the first difference series to find when there are gaps in
# contiguous sequences of storm observations - want to split up into
# separate storm events
storm_1diffs = np.diff(storms)
separations, = np.where(storm_1diffs != 1)
# expand around storm signal so that all low SLP values covered, and unflagged
if len(separations) >= 1:
print " multiple storms in {} {}".format(
y + start.year, month)
# if more than one storm signal that month, then use intervals
# in the first difference series to expand around the first interval alone
storm_start = 0
storm_finish = separations[0] + 1
first_storm = dgc_expand_storms(
storms[storm_start:storm_finish],
len(this_year_data))
final_storms = copy.deepcopy(
first_storm)
for j in range(len(separations)):
# then do the rest in a loop
if j + 1 == len(separations):
# final one
this_storm = dgc_expand_storms(
storms[separations[j] +
1:],
len(this_year_data))
else:
this_storm = dgc_expand_storms(
storms[separations[j] +
1:separations[j +
1] +
1],
len(this_year_data))
final_storms = np.append(
final_storms, this_storm)
else:
# else just expand around the signal by 6 hours either way
final_storms = dgc_expand_storms(
storms, len(this_year_data))
else:
final_storms = storms
if len(storms) >= 1:
print "Tropical Storm signal in {} {}".format(
y + start.year, month)
this_year_flags[final_storms] = 0
# and write flags back into array
flags[year[0]:year[1]] = this_year_flags
if plots:
hist, binEdges = np.histogram(gap_plot_values, bins=bins)
plot_hist = np.array([0.01 if h == 0 else h for h in hist])
plt.step(bincenters,
plot_hist,
'r-',
label='flagged',
where='mid')
import calendar
plt.text(0.1,
0.9,
calendar.month_name[month + 1],
transform=plt.gca().transAxes)
plt.legend(loc='lower center',
ncol=3,
bbox_to_anchor=(0.5, -0.2),
frameon=False,
prop={'size': 13})
plt.show()
#plt.savefig(IMAGELOCATION+'/'+station.id+'_DistributionalGap_'+str(month+1)+'.png')
nflags, = np.where(flags != 0)
utils.print_flagged_obs_number(logfile,
"Distributional Gap All",
variable,
len(nflags),
noWrite=diagnostics)
return flags # dgc_all_obs
def detect(station, neighbour, variable, flags, neighbour_count, start, end, distance = 0, diagnostics = False, plots = False):
'''
Detect which observations are outliers
:param MetVar station: station object (target)
:param MetVar neighbour: station object (neighbour)
:param string variable: which variable to process
:param array flags: array to store how many neighbours thing each obs is bad
:param array neighbour_count: how many neighbours present at each obs
:param datetime start: start of dataset
:param datetime end: end of dataset
:param int distance: separation of target and neighbour
:param bool diagnostics: extra output
:param bool plots: make figures
:returns: None
'''
FILTERING_FLAG_COL = {"temperatures":[0,1,4,5,8,12,16,20,27,41,44,58],
"dewpoints":[0,2,4,6,8,9,13,17,21,28,30,31,32,42,45,59],
"slp":[0,3,4,7,11,15,19,23,29,43,46,60],
"windspeeds":[0,4,10,14,18,22,56,62,63,64]} # not used, but ready for it.
st_var = getattr(station, variable)
neigh_var = getattr(neighbour, variable)
# filter by flags - not all (no Climatological or Odd cluster), T record check not in D,
total_flags = np.sum(station.qc_flags[:,FILTERING_FLAG_COL[variable]], axis = 1)
st_filtered = np.ma.masked_where(total_flags == 1, st_var.data)
neigh_filtered = np.ma.masked_where(total_flags == 1, neigh_var.data)
# match the observation times
match = np.where(np.logical_and((st_filtered.data != st_var.mdi), (neigh_filtered.data != neigh_var.mdi)))
month_ranges = utils.month_starts_in_pairs(start, end).reshape(-1,12,2) # in year-long sets of pairs.
if len(match[0]) >= 100:
neighbour_count[match] += 1 # number of neighbours with data present
differences = np.ma.zeros(len(st_filtered))
differences.fill(st_var.mdi)
differences.mask = True
differences[match] = st_filtered.data[match] - neigh_filtered.data[match]
differences.mask[match] = False
all_iqrs = np.zeros(len(differences))
# get monthly IQR values
for month in range(12):
this_month, dummy1, dummy2 = utils.concatenate_months(month_ranges[:,month,:], differences, hours = False)
if len(this_month.compressed()) > 4:
iqr = utils.IQR(this_month.compressed())
if iqr <= 2.: iqr = 2.
else:
iqr = 2.
# and copy back into the array
for year in month_ranges[:,month,:]:
all_iqrs[year[0]:year[1]] = iqr
if plots:
plot_target_neigh_diffs_dist(differences, min(iqr))
dubious = np.ma.where(np.ma.abs(differences) > 5. * all_iqrs)
if len(dubious[0]) >= 1.:
if variable == "slp":
# check if they are storms
positive = np.ma.where(differences > 5. * iqr)
negative = np.ma.where(differences < -5. * iqr)
# if majority negative (2/3) and separation > 100
if (distance > 100.) and (float(len(positive[0]))/len(dubious[0]) < 0.333):
if len(positive[0]) > 0:
flags[positive] += 1
if len(negative[0]) > 0:
neighbour_count[match] -= 1
else:
flags[dubious] += 1
else:
flags[dubious] += 1
return # detect
