# ---------- allocations ---------- #
MAE = np.empty((N_days, N_fcst, N_grids, N_en))
SPREAD = np.empty((N_days, N_fcst, N_grids, N_en))
CRPS = np.empty((N_days, N_fcst, N_grids, N_en))
print("Computing CRPS ...")
for lead in range(N_fcst):
print("lead = {}".format(lead))
with h5py.File(REFCST_dir + "{}_{}_lead{}.hdf".format(perfix_raw, year, lead), 'r') as h5io:
RAW = h5io[key_raw][:, :EN, ...][..., indx, indy]
with h5py.File(REFCST_dir + "{}_{}_lead{}.hdf".format(perfix_smooth, year, lead), 'r') as h5io:
SMOOTH = h5io[key_smooth][:, :EN, ...][..., indx, indy]
AnEn = W_SL*RAW + (1-W_SL)*SMOOTH
for i, en in enumerate(EN_range):
crps, mae, _ = metrics.CRPS_1d_nan(BCH_obs[:, lead, :], AnEn[:, :en, ...])
MAE[:, lead, :, i] = mae
CRPS[:, lead, :, i] = crps
tuple_save = (MAE, CRPS,)
label_save = ['MAE', 'CRPS',]
du.save_hdf5(tuple_save, label_save, save_dir, '{}_CRPS_vs_EN_{}.hdf'.format(perfix_smooth, year))
fcst_raw = np.empty((N_days, N_fcst, EN, N_grids))
for lead in range(N_fcst):
with h5py.File(
REFCST_dir +
"{}_{}_lead{}.hdf".format(prefix_raw, year_fcst, lead),
'r') as h5io:
RAW = h5io[key_raw][:, :EN, ...]
RAW = RAW[..., ~land_mask_bc]
fcst_ref[:, lead, :, :] = RAW
fcst_raw[:, lead, :, :] = RAW
fcst_ref[fcst_ref < 0] = 0
fcst_raw[fcst_raw < 0] = 0
# ---------- Schaake shuffle ---------- #
print('SimSchaake starts ...')
start_time = time.time()
fcst_ss = sim_schaake(year_analog, fcst_ref, fcst_raw, APCP, ERA5, weights)
print("... Completed. Time = {} sec ".format((time.time() - start_time)))
for l in range(N_fcst):
tuple_save = (fcst_ss[:, l, :, :], )
label_save = [
key_raw,
]
du.save_hdf5(tuple_save, label_save, REFCST_dir,
'{}_SS_{}_lead{}.hdf'.format(prefix_raw, year_fcst, l))
dscale_x = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_x1, ind=0)
dscale_y = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_y1, ind=0)
elif date.month in [3, 4, 5]:
dscale_a = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_a2, ind=0)
dscale_b = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_b2, ind=0)
dscale_x = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_x2, ind=0)
dscale_y = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_y2, ind=0)
elif date.month in [6, 7, 8]:
dscale_a = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_a3, ind=0)
dscale_b = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_b3, ind=0)
dscale_x = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_x3, ind=0)
dscale_y = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_y3, ind=0)
elif date.month in [9, 10, 11]:
dscale_a = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_a4, ind=0)
dscale_b = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_b4, ind=0)
dscale_x = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_x4, ind=0)
dscale_y = pu.feature_to_domain_pct(IN, CLIM, IN_ETOPO, land_mask, size, edge, gap, model_unet_y4, ind=0)
RESULT_a[n, ...] = dscale_a
RESULT_b[n, ...] = dscale_b
RESULT_x[n, ...] = dscale_x
RESULT_y[n, ...] = dscale_y
# append
data.append(REGRID_PCT)
data.append(RESULT_a)
data.append(RESULT_b)
data.append(RESULT_x)
data.append(RESULT_y)
label += ['PCT_REGRID', 'UNET_A', 'UNET_B', 'XNET_A', 'XNET_B']
du.save_hdf5(tuple(data), label, out_dir=save_dir, filename='PRISM_PRED_NCEP_PCT_BC_2016_2018.hdf')
REGRID_P = du.log_trans(REGRID_P)
CLIM_4km_duplicate = du.log_trans(CLIM_4km_duplicate)
CLIM_REGRID_duplicate = du.log_trans(CLIM_REGRID_duplicate)
REGRID_P[REGRID_P<thres] = 0
# ------------------------ #
# save data
dict_save['{}_4km'.format(VAR)] = PRISM_P
dict_save['{}_REGRID'.format(VAR)] = REGRID_P
dict_save['{}_CLIM_4km'.format(VAR)] = CLIM_4km_duplicate
dict_save['{}_CLIM_REGRID'.format(VAR)] = CLIM_REGRID_duplicate
# collecting label
label_save.append('{}_4km'.format(VAR))
label_save.append('{}_REGRID'.format(VAR))
label_save.append('{}_CLIM_4km'.format(VAR))
label_save.append('{}_CLIM_REGRID'.format(VAR))
# dictionary to tuple
tuple_etopo = (etopo_4km, etopo_regrid)
tuple_grids = (lon_025, lat_025, lon_4km, lat_4km, land_mask)
# mark labels
label_etopo = ['etopo_4km', 'etopo_regrid']
label_grids = ['lon_025', 'lat_025', 'lon_4km', 'lat_4km', 'land_mask']
# save hdf
tuple_save = tuple(dict_save.values()) + tuple_etopo + tuple_grids
label_all = label_save + label_etopo + label_grids
du.save_hdf5(tuple_save, label_all, out_dir=PRISM_dir, filename='PRISM_{}_features_2015_2020.hdf'.format(VAR))
# =============== CRPS computation =============== #
# allocation
MAE = np.empty((N_days, N_fcst, N_grids))
CRPS = np.empty((N_days, N_fcst, N_grids))
for lead in range(N_fcst):
print("computing lead: {}".format(lead))
with h5py.File(REFCST_dir + "GEFS_QM_{}_lead{}.hdf".format(year, lead),
'r') as h5io:
GEFS_stn = h5io['gefs_qm'][...][..., indx, indy]
crps, mae, _ = metrics.CRPS_1d_nan(BCH_obs[:, lead, :], GEFS_stn)
MAE[:, lead, ...] = mae
CRPS[:, lead, ...] = crps
# save (all lead times, per year, GEFS only)
tuple_save = (
MAE,
CRPS,
)
label_save = [
'MAE',
'CRPS',
]
du.save_hdf5(tuple_save, label_save, save_dir,
'GEFS_CRPS_BCH_{}.hdf'.format(year))
print('Lapse rate correction')
date_ref = 365 + 365
date_list = [
datetime(2018, 1, 1, 0) + timedelta(days=x) for x in range(date_ref)
]
gamma_mon = [
-4.4, -5.9, -7.1, -7.8, -8.1, -8.2, -8.1, -8.1, -7.7, -6.8, -5.5, -4.7
]
TMEAN_fix = np.zeros((date_ref, ) + lon_4km.shape)
TMEAN_correct = np.zeros((date_ref, ) + lon_4km.shape)
delta_etopo = etopo_4km - etopo_regrid
for i, date in enumerate(date_list):
mon_id = date.month - 1
TMEAN_fix[i,
...] = TMEAN_REGRID[i] + gamma_mon[mon_id] * 1e-3 * delta_etopo
gamma_interp = 0.5 * du.interp2d_wraper(lon_025,
lat_025,
jra_gamma[i, ...],
lon_4km,
lat_4km,
method=interp_method)
TMEAN_correct[i, ...] = TMEAN_REGRID[i] + gamma_interp * delta_etopo
TMEAN_correct[:, land_mask] = np.nan
data_save = (lon_4km, lat_4km, TMEAN_correct, TMEAN_fix, land_mask)
label_save = ['lon_4km', 'lat_4km', 'TMEAN_correct', 'TMEAN_fix', 'land_mask']
du.save_hdf5(data_save, label_save, JRA_dir, 'JRA_TMEAN_correct_2018_2020.hdf')
# import 3d (time, lat, lon) features
hdf_io = h5py.File(PRISM_dir + 'PRISM_{}_features_2015_2020.hdf'.format(var), 'r')
PRISM_T = hdf_io['{}_4km'.format(var)][ind_pred, ...]
REGRID_T = hdf_io['{}_REGRID'.format(var)][ind_pred, ...]
hdf_io.close()
# import pre-trained models (import together for saving time)
# UNET
unet = {}
unet['djf'] = keras.models.load_model(model_import_dir+'UNET3_{}_djf.hdf'.format(var))
unet['mam'] = keras.models.load_model(model_import_dir+'UNET3_{}_mam.hdf'.format(var))
unet['jja'] = keras.models.load_model(model_import_dir+'UNET3_{}_jja.hdf'.format(var))
unet['son'] = keras.models.load_model(model_import_dir+'UNET3_{}_son.hdf'.format(var))
for n, date in enumerate(pred_list):
X = (REGRID_T[n, ...], etopo_4km, etopo_regrid)
print(date)
if date.month in [12, 1, 2]:
temp_unet = vu.pred_domain(X, land_mask, unet['djf'], param, method='norm_std')
elif date.month in [3, 4, 5]:
temp_unet = vu.pred_domain(X, land_mask, unet['mam'], param, method='norm_std')
elif date.month in [6, 7, 8]:
temp_unet = vu.pred_domain(X, land_mask, unet['jja'], param, method='norm_std')
elif date.month in [9, 10, 11]:
temp_unet = vu.pred_domain(X, land_mask, unet['son'], param, method='norm_std')
RESULT_UNET[n, ...] = temp_unet
tuple_save = (lon_4km, lat_4km, PRISM_T, REGRID_T, RESULT_UNET)
label_save = ['lon_4km', 'lat_4km', '{}_4km'.format(var), '{}_REGRID'.format(var), 'RESULT_UNET']
du.save_hdf5(tuple_save, label_save, out_dir=save_dir, filename='PRISM_PRED_{}_test.hdf'.format(var))
IND_base = []
for i in range(grid_shape[0]):
for j in range(grid_shape[1]):
if ~bc_in_base[i, j]:
IND.append([i, j])
if ~land_mask[i, j]:
IND_base.append([i, j])
IND = np.array(IND, dtype=np.int)
IND_base = np.array(IND_base, dtype=np.int)
# ---------------------------------- #
# the main loop
print("Main program starts ...")
start_time = time.time()
OUT = SL_search(
IND,
IND_base,
era_3hq,
Z,
facet_h,
facet_m,
facet_l,
W_facet,
)
print("{} secs for all locs".format(time.time() - start_time))
# ---------------------------------- #
# save
tuple_save = (OUT, )
label_save = ['IND']
du.save_hdf5(tuple_save, label_save, save_dir, 'S40_mon{}.hdf'.format(month))
for lead in range(N_fcst):
for d, date in enumerate(date_list):
# ini date + lead time
date_true = date + timedelta(hours=FCSTs[lead])
if date_true.month-1 in month_around:
flag_pick[mon, lead, d] = 1.0
else:
flag_pick[mon, lead, d] = 0.0
MEAN = np.empty((12, N_fcst, N_grids))
for lead in range(N_fcst):
MEAN[:, lead, :] = climo_mean(ERA5_obs[:, lead, :], flag_pick[:, lead, :])
# ---------- Duplicate to 2016-2020 ---------- #
N_days = 366 + 365*3
date_base = datetime(2016, 1, 1)
date_list = [date_base + timedelta(days=x) for x in np.arange(N_days, dtype=np.float)]
MEAN_BCH = np.empty((N_days, N_fcst, N_grids))
for i, date in enumerate(date_list):
mon_ = date.month-1
MEAN_BCH[i, :, :] = MEAN[mon, ...]
tuple_save = (MEAN_BCH,)
label_save = ['MEAN_BCH']
du.save_hdf5(tuple_save, label_save, save_dir, 'BCH_climo-mean_ERA5.hdf')
param['size'] = 96
# loop over variables and seasons
VARS = ['TMEAN']
for VAR in VARS:
# import 3d (time, lat, lon) features
with h5py.File(JRA_dir + 'JRA_{}_features_2015_2020.hdf'.format(VAR), 'r') as hdf_io:
REGRID_T = hdf_io['{}_REGRID'.format(VAR)][ind_pred, ...]
shape_3d = REGRID_T.shape
RESULT_CLEAN = np.zeros(shape_3d)
RESULT_025 = np.zeros((shape_3d[0],)+lon_025.shape)
for n in range(shape_3d[0]):
#start_time = time.time()
print('\t{}'.format(n))
X = (REGRID_T[n, ...], etopo_regrid)
temp_unet = vu.pred_domain(X, land_mask, CGAN, param, method='norm_std')
#temp_025 = du.nearest_wraper(lon_4km, lat_4km, temp_unet, lon_025, lat_025)
temp_025 = du.interp2d_wraper(lon_4km, lat_4km, temp_unet, lon_025, lat_025, method=interp_method)
temp_4km = du.interp2d_wraper(lon_025, lat_025, temp_025, lon_4km, lat_4km, method=interp_method)
RESULT_025[n, ...] = temp_025
RESULT_CLEAN[n, ...] = temp_4km
#print("--- %s seconds ---" % (time.time() - start_time))
RESULT_CLEAN[:, land_mask] = np.nan
tuple_save = (lon_4km, lat_4km, RESULT_CLEAN, RESULT_025, etopo_4km, etopo_regrid)
label_save = ['lon_4km', 'lat_4km', '{}_REGRID'.format(VAR), '{}_025'.format(VAR), 'etopo_4km', 'etopo_regrid']
du.save_hdf5(tuple_save, label_save, out_dir=JRA_dir, filename='JRA_{}_clean_{}.hdf'.format(VAR, year))
CLIM_elev[i, en, ..., 1] = etopo_025
with h5py.File(
REFCST_dir +
"{}_final_dress_SS_{}_lead{}.hdf".format(TYPE, year, lead),
'r') as h5io:
H15_SL = h5io['AnEn'][:N_sample, :EN, ...]
# noisy AnEn preprocess
H15_SL[H15_SL < 0] = 0
H15_SL = np.log(H15_SL + 1)
RAW[..., ~land_mask_bc] = H15_SL
RAW[..., land_mask_bc] = 0
for i in range(N_sample):
X = np.concatenate((RAW[i, ..., None], CLIM_elev[i, ...]), axis=-1)
temp_ = model.predict([X])
cgan_raw[i, ...] = temp_[-1][..., 0]
cgan_raw[cgan_raw < 0] = 0
cgan_raw = np.exp(cgan_raw) - 1 # <--- de-normalized
cgan_raw[..., land_mask_bc] = np.nan
tuple_save = (cgan_raw, )
label_save = [
'cnn_pred',
]
du.save_hdf5(tuple_save, label_save, REFCST_dir,
'{}_CNN_{}_lead{}.hdf'.format(TYPE, year, lead))
count += L_day
print('Interpolation')
for i in range(days_ref):
if i % 200 == 0:
print('\tday index: {}'.format(i))
temp_interp = du.interp2d_wraper(lon_025,
lat_025,
jra_var[i, ...],
lon_4km,
lat_4km,
method=interp_method)
# land mask applied
temp_interp[land_mask] = np.nan
jra_interp[i, ...] = temp_interp
print('Feature engineering')
if VAR in ['TMAX', 'TMIN', 'TMEAN']:
print('\tK to C')
jra_var = jra_var - 273.15
jra_interp = jra_interp - 273.15
data_save = (lon_4km, lat_4km, jra_interp, TMEAN_4km, etopo_regrid,
land_mask)
label_save = [
'lon_4km', 'lat_4km', '{}_REGRID'.format(VAR), 'TMEAN_4km',
'etopo_regrid', 'land_mask'
]
du.save_hdf5(data_save, label_save, JRA_dir,
'JRA_{}_features_US_2015_2020.hdf'.format(VAR))
X_3d_valid = (T_REGRID[ind_valid, ...][ind_valid_temp, ...],
T_CLIM_4km[ind_valid, ...][ind_valid_temp, ...])
I, C, OUT = vu.baseline_estimator(X_3d_train, X_2d, Y, X_3d_valid,
X_2d, land_mask)
OUT_with_clim[ind_valid_temp, ...] = OUT
C_with_clim[..., N_with_clim * i:N_with_clim * (i + 1)] = C
I_with_clim[..., i] = I
if clim:
# save as hdf
tuple_save = (lon_4km, lat_4km, C_no_clim, I_no_clim, OUT_no_clim,
C_with_clim, I_with_clim, OUT_with_clim)
label_save = [
'lon_4km', 'lat_4km', 'C_no_clim', 'I_no_clim', 'OUT_no_clim',
'C_with_clim', 'I_with_clim', 'OUT_with_clim'
]
du.save_hdf5(tuple_save,
label_save,
out_dir=save_dir,
filename='BASELINE_PRISM_{}_2018_2020.hdf'.format(var))
else:
# save as hdf
tuple_save = (lon_4km, lat_4km, C_no_clim, I_no_clim, OUT_no_clim)
label_save = [
'lon_4km', 'lat_4km', 'C_no_clim', 'I_no_clim', 'OUT_no_clim'
]
du.save_hdf5(tuple_save,
label_save,
out_dir=save_dir,
filename='BASELINE_PRISM_{}_2018_2020.hdf'.format(var))
# print("SG filter starts ...")
# start_time2 = time.time()
AnEn_grid = np.empty((L_fcst_days, EN) + bc_shape)
# AnEn_grid[...] = 0.0
# AnEn_SG = np.empty((L_fcst_days, EN)+bc_shape)
# AnEn_SG[...] = np.nan
for i in range(L_fcst_days):
for j in range(EN):
AnEn_grid[i, j, ~land_mask_bc] = AnEn[i, ..., j]
# # smoothings
# temp_ = AnEn_grid[i, j, ...]
# temp_barnes = ana.sg2d(temp_, window_size=9, order=3, derivative=None) # <-- copied
# temp_barnes[~land_mask_bc] = temp_[~land_mask_bc]
# temp_barnes = ana.sg2d(temp_barnes, window_size=9, order=3, derivative=None) # <-- copied
# temp_barnes[land_mask_bc] = np.nan
# AnEn_SG[i, j, ...] = temp_barnes
AnEn_grid[..., land_mask_bc] = np.nan
# print("... Completed. Time = {} sec ".format((time.time() - start_time2)))
# tuple_save = (AnEn_grid, AnEn_SG)
# label_save = ['AnEn', 'AnEn_SG']
tuple_save = (AnEn_grid, )
label_save = [
'AnEn',
]
du.save_hdf5(tuple_save, label_save, REFCST_dir,
'BASE_final_{}_lead{}.hdf'.format(year_fcst, lead))
clim_4km[i,
...] = CLIM[var][mon_id, ...] / mon_days_366[mon_id]
clim_regrid[i,
...] = CLIM_REGRID[var][mon_id,
...] / mon_days_366[mon_id]
else:
clim_4km[i,
...] = CLIM[var][mon_id, ...] / mon_days_365[mon_id]
clim_regrid[i,
...] = CLIM_REGRID[var][mon_id,
...] / mon_days_365[mon_id]
clim_4km = du.log_trans(clim_4km)
clim_regrid = du.log_trans(clim_regrid)
clim_regrid[clim_regrid < thres] = 0
# save data
data_save = (lon_4km, lat_4km, lon_ncep[domain_ind[0]:domain_ind[1],
domain_ind[2]:domain_ind[3]],
lat_ncep[domain_ind[0]:domain_ind[1],
domain_ind[2]:domain_ind[3]], var_4km, clim_4km,
clim_regrid, ncep_var[:, domain_ind[0]:domain_ind[1],
domain_ind[2]:domain_ind[3]], etopo_4km,
etopo_regrid, land_mask)
label_save = [
'lon_4km', 'lat_4km', 'lon_ncep', 'lat_ncep', '{}_REGRID'.format(var),
'{}_CLIM_4km'.format(var), '{}_CLIM_REGRID'.format(var),
'{}_originals'.format(var), 'etopo_4km', 'etopo_regrid', 'land_mask'
]
du.save_hdf5(data_save, label_save, NCEP_dir,
'NCEP_{}_features_BC_2016_2020.hdf'.format(var))
unet['djf'],
param,
method='norm_std')
elif date.month in [3, 4, 5]:
temp_unet = vu.pred_domain(X,
land_mask,
unet['mam'],
param,
method='norm_std')
elif date.month in [6, 7, 8]:
temp_unet = vu.pred_domain(X,
land_mask,
unet['jja'],
param,
method='norm_std')
elif date.month in [9, 10, 11]:
temp_unet = vu.pred_domain(X,
land_mask,
unet['son'],
param,
method='norm_std')
RESULT_UNET[n, ...] = temp_unet
tuple_save = (lon_4km, lat_4km, REGRID_T, RESULT_UNET)
label_save = ['lon_4km', 'lat_4km', '{}_REGRID'.format(var), 'RESULT_UNET']
du.save_hdf5(tuple_save,
label_save,
out_dir=save_dir,
filename='ERA_PRED_{}_2018_2020.hdf'.format(var))
bc_inds[2]:bc_inds[3]]
era_pct_daily = np.sum(np.reshape(era_pct, (N_days, 8, 48, 112)), axis=1)
ERA5 += (era_pct_daily[..., indx, indy], )
ERA5_obs = np.concatenate(ERA5, axis=0)
# ========== BCH ========== #
OBS = ()
for key in stn_code:
with pd.HDFStore(BACKUP_dir + 'BCH_PREC_QC_1D_2016_2020.hdf',
'r') as hdf_io:
pd_temp = hdf_io[key]
pd_temp.index = pd_temp['datetime']
pd_temp = pd_temp['2016-01-01':'2019-12-31']
obs_ = pd_temp['PREC_HOUR_QC'].values
OBS += (obs_[:, None], )
BCH_obs = np.concatenate(OBS, axis=-1)
# ! <--- data cleaning
BCH_obs[BCH_obs > 300] = np.nan
for i in range(len(stn_code)):
BCH_obs[:, i] = bu.clean_no_response(BCH_obs[:, i], tol_window=60 * 8)
tuple_save = (ERA5_obs, BCH_obs, flag_pick)
label_save = ['ERA5_obs', 'BCH_obs', 'stn_flag']
du.save_hdf5(tuple_save, label_save, save_dir, 'BCH_ERA5_1D_pairs.hdf')
obs = obs[flag_nonan]
fcst = fcst[flag_nonan]
L = np.sum(flag_nonan)
o_bar_ = np.mean(obs)
o_bar[d] = o_bar_
for n in range(N_boost):
ind_bagging = np.random.choice(L, size=L, replace=True)
obs_ = obs[ind_bagging]
fcst_ = fcst[ind_bagging]
prob_true_, prob_pred_ = reliability_diagram(
obs_, fcst_, hist_bins)
brier_ = brier_score_loss(obs_, fcst_)
prob_true[d, :, n] = prob_true_
prob_pred[d, :, n] = prob_pred_
brier[d, n] = brier_
hist_bins_ = np.mean(prob_pred[d, ...], axis=1)
use_, _ = np.histogram(fcst, bins=np.array(list(hist_bins_) + [1.0]))
use[d, :] = use_
tuple_save = (brier, prob_true, prob_pred, use, o_bar)
label_save = ['brier', 'pos_frac', 'pred_value', 'use', 'o_bar']
du.save_hdf5(tuple_save, label_save, save_dir,
'{}_Calib_loc{}.hdf'.format(prefix_out, r))
data_mask = np.squeeze(temp_io.dataset_mask())
N_lon = temp_io.width
N_lat = temp_io.height
temp_io.close()
lon, lat = np.meshgrid(np.arange(bounds[0], bounds[0] + dx * N_lon, dx),
np.arange(bounds[1], bounds[1] + dy * N_lat, dy))
# loop over files
PRISM_PCT = np.empty((len(file_dir), N_lat, N_lon))
datenum = []
for i, temp_dir in enumerate(file_dir):
temp_file = glob(temp_dir + '*.bil')[0]
#print(temp_file)
temp_io = rasterio.open(temp_file, 'r')
temp_data = np.squeeze(temp_io.read())
temp_io.close()
# vals
temp_data[data_mask == 0] = np.nan
temp_data = np.flipud(temp_data)
PRISM_PCT[i, ...] = temp_data
# datenum
date_str = basename(temp_file)[-16:-8] # get date string from filename
datenum += du.dt_to_sec([datetime.strptime(date_str, '%Y%m%d')])
# save hdf5
tuple_save = (PRISM_PCT, np.array(datenum), lon, lat)
label_save = ['PRISM_{}'.format(var), 'datenum', 'lon', 'lat']
du.save_hdf5(tuple_save,
label_save,
out_dir=PRISM_dir,
filename='PRISM_{}_2015_2020.hdf'.format(var))
prism_025[i, ...] = temp_025
prism_regrid[i, ...] = temp_regrid
# collecting fields
dict_4km[VAR] = prism_4km
dict_025[VAR] = prism_025
dict_regrid[VAR] = prism_regrid
# collecting label
label_4km.append(VAR + '_4km')
label_025.append(VAR + '_025')
label_regrid.append(VAR + '_REGRID')
# dictionary to tuple
tuple_4km = tuple(dict_4km.values())
tuple_025 = tuple(dict_025.values())
tuple_regrid = tuple(dict_regrid.values())
tuple_etopo = (etopo_4km, etopo_025, etopo_regrid)
tuple_grids = (lon_025, lat_025, lon_4km, lat_4km, land_mask)
# mark labels
label_etopo = ['etopo_4km', 'etopo_025', 'etopo_regrid']
label_grids = ['lon_025', 'lat_025', 'lon_4km', 'lat_4km', 'land_mask']
# save hdf
tuple_save = tuple_4km + tuple_025 + tuple_regrid + tuple_etopo + tuple_grids
label_save = label_4km + label_025 + label_regrid + label_etopo + label_grids
du.save_hdf5(tuple_save,
label_save,
out_dir=PRISM_dir,
filename='PRISM_regrid_clim.hdf')
land_mask_025 = hdf_io['land_mask_025'][...]
land_mask_terrain_025 = hdf_io['land_mask_terrain_025'][...]
land_mask = hdf_io['land_mask'][...]
land_mask_terrain_4km = hdf_io['land_mask_terrain'][...]
etopo_4km = hdf_io['etopo_4km'][...]
etopo_regrid = hdf_io['etopo_regrid'][...]
land_mask_025[:, 155:] = True
land_mask_025[152:, :] = True
land_mask_025[:5, :] = True
land_mask_terrain_025[:, 155:] = True
land_mask_terrain_025[145:, :] = True
land_mask_terrain_025[:5, :] = True
# save
tuple_save = (lon_4km, lat_4km, lon_025, lat_025, etopo_4km, etopo_regrid,
land_mask, land_mask_025, land_mask_terrain_4km,
land_mask_terrain_025)
label_save = [
'lon_4km', 'lat_4km', 'lon_025', 'lat_025', 'etopo_4km', 'etopo_regrid',
'land_mask', 'land_mask_025', 'land_mask_terrain', 'land_mask_terrain_025'
]
du.save_hdf5(tuple_save,
label_save,
out_dir=PRISM_dir,
filename='land_mask_NA.hdf')
with nc.Dataset(filenames[VAR][file_num], 'r') as nc_io:
temp_var[j, ...] = np.squeeze(nc_io.variables[nc_keys[VAR]][...])
era_sfp[i, ...] = f(temp_var, axis=0)
print('Lapse rate with linear regression')
sorter = np.array(
[19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0])
for i in range(days_ref):
print(i)
for j in range(grid_shape[0]):
for k in range(grid_shape[1]):
temp_lev = np.array(A + B * era_sfp[i, j, k])
temp_height = 1000 * metpy.calc.pressure_to_height_std(
temp_lev * units.Pa).__array__()
temp_temp = era_var[i, :, j, k]
era_height[i, :, j, k] = temp_height
era_lev[i, :, j, k] = temp_lev
ind1 = np.searchsorted(temp_height, temp_height[0] + 1500)
ind2 = np.searchsorted(temp_height, temp_height[0] + 2500)
era_gamma[i, j, k] = linear_slope(temp_height[ind1:ind2],
temp_temp[ind1:ind2])
data_save = (lon_025, lat_025, era_var, era_sfp, era_gamma, era_height,
era_lev)
label_save = [
'lon_025', 'lat_025', 'era_var', 'era_sfp', 'era_gamma', 'era_height',
'era_lev'
]
du.save_hdf5(data_save, label_save, ERA_dir, 'ERA_TMEAN_GAMMA_2018_2020.hdf')
obs = obs[flag_nonan]
fcst = fcst[flag_nonan]
L = np.sum(flag_nonan)
o_bar_ = np.mean(obs)
o_bar[d] = o_bar_
for n in range(N_boost):
ind_bagging = np.random.choice(L, size=L, replace=True)
obs_ = obs[ind_bagging]
fcst_ = fcst[ind_bagging]
prob_true_, prob_pred_ = reliability_diagram(
obs_, fcst_, hist_bins)
brier_ = brier_score_loss(obs_, fcst_)
prob_true[d, :, n] = prob_true_
prob_pred[d, :, n] = prob_pred_
brier[d, n] = brier_
hist_bins_ = np.mean(prob_pred[d, ...], axis=1)
use_, _ = np.histogram(fcst, bins=np.array(list(hist_bins_) + [1.0]))
use[d, :] = use_
tuple_save = (brier, prob_true, prob_pred, use, o_bar)
label_save = ['brier', 'pos_frac', 'pred_value', 'use', 'o_bar']
du.save_hdf5(tuple_save, label_save, save_dir,
'GEFS_99th_Calib_loc{}.hdf'.format(r))
with h5py.File(ERA_dir + 'PT_3hr_{}.hdf'.format(year), 'r') as h5io:
era_pct = h5io['era_025'][...]
PCT_history['{}'.format(year)] = era_pct
for year in range(2000, 2020):
print("Processing year: {}".format(year))
N_days = (datetime(year + 1, 1, 1) - datetime(year, 1, 1)).days
ERA_fcst = np.zeros((N_days, N_fcst, 160, 220))
for day in range(N_days):
for t, fcst_temp in enumerate(FCSTs):
# fcsted (targeted) date
date_true = datetime(
year, 1, 1) + timedelta(days=day) + timedelta(hours=fcst_temp)
# handling cross years
year_true = int(date_true.year)
ind_true = int(
(date_true - datetime(year_true, 1, 1)).total_seconds() / 60 /
60 / freq)
#
ERA_fcst[day, t,
...] = PCT_history['{}'.format(year_true)][ind_true, ...]
ERA_fcst[ERA_fcst < 1e-5] = 0.0
tuple_save = (ERA_fcst, )
label_save = ['era_fcst']
filename = 'ERA5_GEFS-fcst_{}.hdf'.format(year)
du.save_hdf5(tuple_save, label_save, ERA_dir, filename)
o_bar = np.mean(BCH_binary)
hist_bins_base = np.mean(prob_pred_base, axis=1)
hist_bins_bcnn = np.mean(prob_pred_bcnn, axis=1)
hist_bins_sl = np.mean(prob_pred_sl, axis=1)
hist_bins_scnn = np.mean(prob_pred_scnn, axis=1)
use_base, _ = np.histogram(BASE_prob,
bins=np.array(list(hist_bins_base) + [1.0]))
use_bcnn, _ = np.histogram(BCNN_prob,
bins=np.array(list(hist_bins_bcnn) + [1.0]))
use_sl, _ = np.histogram(SL_prob,
bins=np.array(list(hist_bins_sl) + [1.0]))
use_scnn, _ = np.histogram(SCNN_prob,
bins=np.array(list(hist_bins_scnn) + [1.0]))
tuple_save = (prob_true_base, prob_true_bcnn, prob_true_sl, prob_true_scnn,
prob_pred_base, prob_pred_bcnn, prob_pred_sl, prob_pred_scnn,
brier_base, brier_bcnn, brier_sl, brier_scnn, use_base,
use_bcnn, use_sl, use_scnn, o_bar)
label_save = [
'prob_true_base', 'prob_true_bcnn', 'prob_true_sl', 'prob_true_scnn',
'prob_pred_base', 'prob_pred_bcnn', 'prob_pred_sl', 'prob_pred_scnn',
'brier_base', 'brier_bcnn', 'brier_sl', 'brier_scnn', 'use_base',
'use_bcnn', 'use_sl', 'use_scnn', 'o_bar'
]
du.save_hdf5(tuple_save, label_save, save_dir,
'Accum_Calib_lead{}_loc{}.hdf'.format(lead_, r))
MEAN = MEAN_BCH[flag_pick, ...]
# =============== Allocation and MAE computation =============== #
# N_days
if year % 4 == 0:
N_days = 366
else:
N_days = 365
# other params
N_fcst = 54
EN = 75
N_grids = BCH_obs.shape[-1]
MAE = np.empty((N_days, N_fcst, N_grids))
print("Computing MAE ...")
for lead in range(N_fcst):
print("lead = {}".format(lead))
MAE[:, lead, ...] = np.abs(BCH_obs[:, lead, :] - MEAN[:, lead, :])
tuple_save = (MAE, )
label_save = [
'MAE',
]
du.save_hdf5(tuple_save, label_save, save_dir,
'CLIM_MAE_BCH_{}.hdf'.format(year))
# loop over variables and seasons
VARS = ['TMEAN']
for VAR in VARS:
# import 3d (time, lat, lon) features
with h5py.File(JRA_dir + 'JRA_{}_features_US_2015_2020.hdf'.format(VAR), 'r') as hdf_io:
PRISM_T = hdf_io['{}_4km'.format(VAR)][...]
REGRID_T = hdf_io['{}_REGRID'.format(VAR)][...]
shape_3d = REGRID_T.shape
RESULT_CLEAN = np.zeros(shape_3d)
RESULT_025 = np.zeros((shape_3d[0],)+lon_025.shape)
for n in range(shape_3d[0]):
print('\t{}'.format(n))
X = (REGRID_T[n, ...], etopo_regrid)
temp_unet = vu.pred_domain(X, land_mask, CGAN, param, method='norm_std')
temp_025 = du.interp2d_wraper(lon_4km, lat_4km, temp_unet, lon_025, lat_025, method=interp_method)
temp_4km = du.interp2d_wraper(lon_025, lat_025, temp_025, lon_4km, lat_4km, method=interp_method)
RESULT_025[n, ...] = temp_025
RESULT_CLEAN[n, ...] = temp_4km
RESULT_CLEAN[:, land_mask] = np.nan
tuple_save = (lon_4km, lat_4km, PRISM_T, RESULT_CLEAN, RESULT_025, etopo_4km, etopo_regrid)
label_save = ['lon_4km', 'lat_4km', '{}_4km'.format(VAR), '{}_REGRID'.format(VAR), '{}_025'.format(VAR), 'etopo_4km', 'etopo_regrid']
du.save_hdf5(tuple_save, label_save, out_dir=JRA_dir, filename='JRA_US_{}_clean_2015_2020.hdf'.format(VAR))
# ========== Region subsets ========== #
south = [-130, -121, 48.75, 50.25]
north = [-127.5, -110, 53, 60]
loc_id = [] # 0 van isl, 1 south, 2 rocky, 3 north
for i in range(len(stn_code)):
stn_lat_temp = stn_lat[i]
stn_lon_temp = stn_lon[i]
if du.check_bounds(stn_lon_temp, stn_lat_temp, south):
loc_id.append(1)
elif du.check_bounds(stn_lon_temp, stn_lat_temp, north):
loc_id.append(3)
else:
loc_id.append(2)
loc_id = np.array(loc_id)
flag_sw = loc_id == 1
flag_si = loc_id == 2
flag_n = loc_id == 3
# ========== Save ========== #
tuple_save = (stn_lon, stn_lat, flag_sw, flag_si, flag_n)
label_save = ['stn_lon', 'stn_lat', 'flag_sw', 'flag_si', 'flag_n']
du.save_hdf5(tuple_save, label_save, save_dir, 'BCH_wshed_groups.hdf')
# 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_
BS[flag_, lead, :] = metrics.BS_binary_1d_nan(obs_flag, pred_flag)
# save (all lead times, per year, per experiment)
tuple_save = (BS, BCH_90th)
label_save = ['BS', 'stn_90th']
du.save_hdf5(tuple_save, label_save, save_dir,
'{}_BS_BCH_{}.hdf'.format(prefix_out, year))
Z_l[land_mask] = 0 Z_m = gaussian_filter(etopo_025, 5 / np.pi) Z_m[land_mask] = 0 Z_h = np.copy(etopo_025) Z_h[land_mask] = 0 facet_h = facet(Z_h) facet_m = facet(Z_m) facet_l = facet(Z_l) facet_h[land_mask] = np.nan facet_m[land_mask] = np.nan facet_l[land_mask] = np.nan sigma_facet = window_stdev_slow(etopo_025, radius=5) W_facet = sigma_facet / np.nanmax(sigma_facet) # W_025 = 0.6*(sigma025-15) # W_025[W_025>0.6] = 0.6 # W_025[W_025<0.2] = 0.2 # W_025[land_mask] = np.nan W_025 = 0.8 * (sigma025 - 15) W_025[W_025 > 0.8] = 0.8 W_025[W_025 < 0.2] = 0.2 W_025[land_mask] = np.nan tuple_save = (facet_h, facet_m, facet_l, W_facet, W_025) label_save = ['facet_h', 'facet_m', 'facet_l', 'W_facet', 'W_SL'] du.save_hdf5(tuple_save, label_save, save_dir, 'NA_SL_info.hdf')
