for lead in range(N_fcst):
print("computing lead: {}".format(lead))
with h5py.File(
REFCST_dir + "{}_{}_lead{}.hdf".format(prefix_raw, year, lead),
'r') as h5io:
AnEn_ = h5io[key_raw][:, :EN, ...]
if type_ind == 0 or type_ind == 1:
# id 0 and 1 are flattened grid points, reshape them to 2d.
AnEn_full[..., ~land_mask_bc] = AnEn_
AnEn_stn = AnEn_full[..., indx, indy]
else:
AnEn_stn = AnEn_[..., indx, indy]
# cnn outputs can be negative, fix it here.
AnEn_stn = ana.cnn_precip_fix(AnEn_stn)
# extracting the 90-th threshold for initializaiton time + lead time
for mon in range(12):
flag_ = flag_pick[:, lead] == mon
# stn obs
obs_ = BCH_obs[flag_, lead, :]
# fcst
pred_ = AnEn_stn[flag_, ...]
# station-wise threshold
thres_ = BCH_90th[mon, :]
# Brier Score ( ">=" is applied)
obs_flag = obs_ >= thres_
obs_flag[np.isnan(obs_)] = np.nan
pred_flag = pred_ >= thres_
for lead in range(N_fcst):
with h5py.File(
REFCST_dir + "{}_{}_lead{}.hdf".format(prefix_raw, y, lead),
'r') as h5io:
AnEn_ = h5io[key_raw][:, :EN, ...]
if type_ind == 0 or type_ind == 1:
# id 0 and 1 are flattened grid points, reshape them to 2d.
AnEn_full[..., ~land_mask_bc] = AnEn_
AnEn_stn[flag_pick, lead, ...] = AnEn_full[..., indx, indy]
else:
# cnn outputs can be negative, fix it here.
AnEn_stn[flag_pick, lead,
...] = ana.cnn_precip_fix(AnEn_[..., indx, indy])
# ========== Calibration ========== #
# params
N_bins = 15
N_boost = 100
hist_bins = np.linspace(0, 1, N_bins)
N_lead_day = 7 # number of forecasted days
# 3-hr lead times to days
fcst_leads_ini = np.arange(0, 72 * 3 + 3, 3, dtype=np.float)
# date_base = datetime(2017, 1, 1, 0) # 2017 as a reference
DAYS = []
for lead in fcst_leads_ini:
date_temp = date_base + timedelta(hours=lead)