def metrics(yp, yt, keyLst, var_s, output_path="./"):
statErr = statError(yp, yt)
statDictLst = [statErr]
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
data = statDictLst[k][statStr]
data = data[~np.isnan(data)]
temp.append(data)
dataBox.append(temp)
median_dict = plotBoxFigMulti(dataBox, label1=keyLst, sharey=False, figsize=(12, 5), path_fig=output_path)
return median_dict
temp = 1
else:
if temp != 0:
temp = temp + 1
ind = np.random.randint(0, ngrid)
print(np.array([maskObs[ind, :], maskDay[ind, :]]))
maskObsDay = maskObs * maskDay
unique, counts = np.unique(maskDay, return_counts=True)
print(np.asarray((unique, counts)).T)
print(counts / ngrid / nt)
fLst = [1, 2, 3]
statLstF = list()
statLstP = list()
maskF = (maskDay >= 1) & (maskDay <= 3)
statP = stat.statError(utils.fillNan(yp, maskF), utils.fillNan(obs, maskF))
statF = stat.statError(utils.fillNan(yf, maskF), utils.fillNan(obs, maskF))
# plot map and time series
import importlib
importlib.reload(plot)
dataGrid = [statP['RMSE'] - statF['RMSE'], statP['Corr'] - statF['Corr']]
prcp = df.getDataTs('APCP_FORA').squeeze()
dataTs = [obs, yp, yf]
crd = df.getGeo()
t = df.getT()
mapNameLst = ['dRMSE', 'dR']
tsNameLst = ['obs', 'prj', 'fore']
plot.plotTsMap(
dataGrid,
dataTs,
# # test error train on different year
trLst = [[20150402, 20160401], [20160401, 20170401], [20170401, 20180401]]
statPLst = list()
statFLst = list()
for k in range(3):
trTrain = trLst[k]
taTrain = utils.time.tRange2Array(trTrain)
taAll = utils.time.tRange2Array([20150402, 20180401])
indTrain, ind2 = utils.time.intersect(taAll, taTrain)
indTest = np.delete(np.arange(len(taAll)), indTrain)
tempYp = ypLst[k][:, indTest]
tempYf = yfLst[k][:, indTest]
tempMask = maskF[:, indTest]
tempObs = obs[:, indTest]
tempStatP = stat.statError(utils.fillNan(tempYp, tempMask),
utils.fillNan(tempObs, tempMask))
tempStatF = stat.statError(utils.fillNan(tempYf, tempMask),
utils.fillNan(tempObs, tempMask))
statPLst.append(tempStatP)
statFLst.append(tempStatF)
# plot map and time series
import importlib
importlib.reload(plot)
dataGrid = [
statPLst[0]['RMSE'] - statFLst[0]['RMSE'],
statPLst[1]['RMSE'] - statFLst[1]['RMSE'],
statPLst[2]['RMSE'] - statFLst[2]['RMSE']
]
prcp = df.getDataTs('APCP_FORA').squeeze()
dataTs = [[obs, ypLst[0], yfLst[0]], [obs, ypLst[1], yfLst[1]],
tRange = [20160501, 20170501]
predLst = list()
outLSTM = os.path.join(pathSMAP['Out_L3_NA'], 'DA', 'CONUSv2f1')
df, pred, obs = master.test(
outLSTM, tRange=tRange, subset=subset, batchSize=100)
predLst.append(pred.squeeze())
for out in outLst:
df, pred, obs = master.test(
out, tRange=tRange, subset=subset, batchSize=100)
predLst.append(pred.squeeze())
obs = obs.squeeze()
# plot box - latency
# if 'post' in doLst:
caseLst = ['Predict'] + ['Nowcast ' + str(nd) + 'd latency' for nd in dLst]
statLst1 = [stat.statError(x, obs) for x in predLst]
keyLst = list(statLst1[0].keys())
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statLst1)):
data = statLst1[k][statStr]
data = data[~np.isnan(data)]
temp.append(data)
dataBox.append(temp)
fig = plot.plotBoxFig(dataBox, keyLst, caseLst, sharey=False)
fig.show()
fig.savefig(os.path.join(saveDir, 'box_latency'))
# figure out how many days observation lead
caseLst.append('All-90-95-DA' + str(nDay))
outLst = [os.path.join(pathCamels['Out'], save_path, x) for x in caseLst]
subset = 'All'
tRange = [19950101, 20000101]
predLst = list()
for out in outLst:
df, pred, obs = master.test(out,
tRange=tRange,
subset=subset,
basinnorm=True,
epoch=200)
# pred=np.maximum(pred,0)
predLst.append(pred)
# plot box
statDictLst = [stat.statError(x.squeeze(), obs.squeeze()) for x in predLst]
# keyLst = list(statDictLst[0].keys())
keyLst = ['Bias', 'RMSE', 'NSE']
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
data = statDictLst[k][statStr]
data = data[~np.isnan(data)]
temp.append(data)
dataBox.append(temp)
# plt.style.use('classic')
plt.rcParams['font.size'] = 14
plt.rcParams['font.family'] = 'Times New Roman'
plt.rcParams["legend.columnspacing"] = 0.1
ypLst = list()
modelName = 'LSTM'
model = train.loadModel(outFolder, nEpoch, modelName=modelName)
yp = train.testModel(model, x, batchSize=100).squeeze()
ypLst.append(
dbCsv.transNorm(yp, rootDB=rootDB, fieldName='SMAP_AM', fromRaw=False))
modelName = 'LSTM-DA'
model = train.loadModel(outFolder, nEpoch, modelName=modelName)
yp = train.testModel(model, x, z=y, batchSize=100).squeeze()
ypLst.append(
dbCsv.transNorm(yp, rootDB=rootDB, fieldName='SMAP_AM', fromRaw=False))
##
statErr1 = stat.statError(ypLst[0], yt)
statErr2 = stat.statError(ypLst[1], yt)
dataGrid = [statErr2['RMSE'], statErr2['RMSE'] - statErr1['RMSE']]
dataTs = [ypLst[0], ypLst[1], yt]
t = df.getT()
crd = df.getGeo()
mapNameLst = ['DA', 'DA-LSTM']
tsNameLst = ['LSTM', 'DA', 'SMAP']
colorMap = None
colorTs = None
plot.plotTsMap(
dataGrid,
dataTs,
crd,
t,
yp1 = train.testModel(model1, x2, c2)
yp1 = camels.transNorm(yp1, 'usgsFlow', toNorm=False).squeeze()
model2 = train.loadModel(outFolder, nEpoch, modelName='DA-1')
yp2 = train.testModel(model2, xz1, c2)
yp2 = camels.transNorm(yp2, 'usgsFlow', toNorm=False).squeeze()
model3 = train.loadModel(outFolder, nEpoch, modelName='DA-7')
yp3 = train.testModel(model3, xz2, c2)
yp3 = camels.transNorm(yp3, 'usgsFlow', toNorm=False).squeeze()
yLst = [yt2, yp1, yp2, yp3]
# plot box
statDictLst = [
stat.statError(yp1, yt2),
stat.statError(yp2, yt2),
stat.statError(yp3, yt2)
]
keyLst = list(statDictLst[0].keys())
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
data = statDictLst[k][statStr]
data = data[~np.isnan(data)]
temp.append(data)
dataBox.append(temp)
fig = plot.plotBoxFig(dataBox, keyLst, ['LSTM', 'DA-1', 'DA-7'], sharey=False)
predLst.append(pred) # the prediction list for all the models
obsLst.append(obs)
np.save(os.path.join(out, 'pred.npy'), pred)
np.save(os.path.join(out, 'obs.npy'), obs)
f = np.load(os.path.join(out, 'x.npy')) # it has been saved previously in the out directory (forcings)
T = (f[:, :, 3] + f[:, :, 4]) / 2 # mean air T for T_residual
T_air = np.expand_dims(T, axis=2)
pred_res = pred - T_air
obs_res = obs - T_air
predLst_res.append(pred_res)
obsLst_res.append(obs_res)
# calculate statistic metrics
# statDict = stat.statError(pred.squeeze(), obs.squeeze())
# statDictLst.append([statDict])
# statDictLst1 = [stat.statError(x.squeeze(), obs.squeeze()) for x, y in predLst]
statDictLst = [stat.statError(x.squeeze(), y.squeeze()) for (x, y) in zip(predLst, obsLst)]
statDictLst_res = [stat.statError_res(x.squeeze(), y.squeeze(), z.squeeze(), w.squeeze()) for (x, y, z, w) in
zip(predLst, obsLst, predLst_res, obsLst_res)]
### save this file too
# median and STD calculation
count = 0
mdstd = np.zeros([len(statDictLst_res[0]),3])
for i in statDictLst_res[0].values():
median = np.nanmedian((i)) # abs(i)
STD = np.nanstd((i)) # abs(i)
mean = np.nanmean((i)) #abs(i)
k = np.array([[median,STD, mean]])
mdstd[count] = k
count = count +1
mdstd = pd.DataFrame(mdstd, index=statDictLst_res[0].keys(), columns=['median', 'STD','mean'])
if retrained==True:
latC, lonC = dfC.getGeo()
# case = '090303'
for case in caseLst:
testName = subsetPattern.format(case, 3)
errLst = list()
for k in levLst:
if k in [0, 1]:
# if k in [-1]:
subset = 'ecoReg_{}_L{}_v2f1'.format(case, k)
else:
subset = subsetPattern.format(case, k)
outName = subset + '_Forcing'
out = os.path.join(pathSMAP['Out_L3_NA'], 'ecoRegionCase', outName)
df, yp, yt = master.test(out, tRange=tRange, subset=testName)
errLst.append(stat.statError(yp[:, :, 0], yt[:, :, 0]))
# plot box
cLst = 'ygbr'
keyLst = ['RMSE', 'Corr']
dataBox = list()
for key in keyLst:
temp = list()
for err in errLst:
temp.append(err[key])
dataBox.append(temp)
fig = plot.plotBoxFig(dataBox,
' ',
figsize=(8, 6),
colorLst=cLst,
sharey=False)
optTrain = default.update(default.optTrainSMAP, nEpoch=100)
out = os.path.join(cDir, 'output', 'CONUSv4f1_multi')
masterDict = wrapMaster(out, optData, optModel, optLoss, optTrain)
# train
# train(masterDict)
# runTrain(masterDict, cudaID=2, screen='LSTM-multi')
# test
df, yp, yt, sigma = test(out, tRange=[20160401, 20170401], subset='CONUSv4f1')
# plot ts MAP
dataGrid = list()
dataTs = list()
for k in range(2):
statErr = stat.statError(yp[:, :, k], yt[:, :, k])
dataGrid.append(statErr['RMSE'])
dataTs.append([yp[:, :, k], yt[:, :, k]])
t = df.getT()
crd = df.getGeo()
mapNameLst = ['RMSE ', 'RMSE']
tsNameLst = ['LSTM', 'SMAP']
plot.plotTsMap(dataGrid,
dataTs,
lat=crd[0],
lon=crd[1],
t=t,
mapNameLst=mapNameLst,
tsNameLst=tsNameLst,
multiTS=True,
isGrid=True)
daylen = xt.shape[1]
Pred = np.full(yt.shape, np.nan)
for ii in range(ngage):
xdata = x[ii, :, :]
ydata = y[ii, :, :]
regmodel = LinearRegression().fit(xdata, ydata)
xtest = xt[ii, :, :]
ypred = regmodel.predict(xtest)
Pred[ii, :, 0] = ypred.squeeze()
pred = camels.transNorm(Pred, 'usgsFlow', toNorm=False)
obs = camels.transNorm(yt, 'usgsFlow', toNorm=False)
gageid = 'All'
pred = camels.basinNorm(pred, gageid=gageid, toNorm=False)
obs = camels.basinNorm(obs, gageid=gageid, toNorm=False)
# plot box
statDictLst = [stat.statError(pred.squeeze(), obs.squeeze())]
keyLst = ['Bias', 'RMSE', 'NSE']
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
data = statDictLst[k][statStr]
data = data[~np.isnan(data)]
temp.append(data)
dataBox.append(temp)
# plt.style.use('classic')
plt.rcParams['font.size'] = 14
plt.rcParams['font.family'] = 'Times New Roman'
plt.rcParams["legend.columnspacing"] = 0.1
plt.rcParams["legend.handletextpad"] = 0.2
batchSize=100)
df2, yf2, obs2 = master.test(out,
tRange=[20160401, 20180401],
subset=subset,
batchSize=100)
out = os.path.join(pathSMAP['Out_L3_NA'], 'DA', 'CONUSv2f1_LSTM2015')
df1, yp1, obs1 = master.test(out,
tRange=[20150402, 20180401],
subset=subset,
batchSize=100)
df2, yp2, obs2 = master.test(out,
tRange=[20160401, 20180401],
subset=subset,
batchSize=100)
statF = stat.statError(yf2.squeeze(), obs2.squeeze())
statP = stat.statError(yp2.squeeze(), obs2.squeeze())
t = df1.getT()
lat, lon = df1.getGeo()
dataTS = [obs1.squeeze(), yp1.squeeze(), yf1.squeeze()]
tBar = np.datetime64('2016-04-01')
def funcMap():
gridF, uy, ux = utils.grid.array2grid(statF['RMSE'], lat=lat, lon=lon)
gridP, uy, ux = utils.grid.array2grid(statP['RMSE'], lat=lat, lon=lon)
figM, axM = plt.subplots(1, 2, figsize=(10, 4))
axplot.mapGrid(axM[0], uy, ux, gridF, vRange=[0, 0.1], cmap=plt.cm.jet)
axM[0].set_title('Temporal Test RMSE of LSTM-DI')
axplot.mapGrid(axM[1], uy, ux, gridP, vRange=[0, 0.1], cmap=plt.cm.jet)
axM[1].set_title('Temporal Test RMSE of LSTM')
maskObs = 1 * ~np.isnan(obs.squeeze())
maskDay = np.zeros(maskObs.shape).astype(int)
ngrid, nt = maskObs.shape
for j in range(ngrid):
temp = 0
for i in range(nt):
maskDay[j, i] = temp
if maskObs[j, i] == 1:
temp = 1
else:
if temp != 0:
temp = temp + 1
ind = np.random.randint(0, ngrid)
maskObsDay = maskObs * maskDay
maskF = (maskDay >= 1) & (maskDay <= 3)
statP = stat.statError(yp, obs)
statLst = [
stat.statError(utils.fillNan(x, maskF), utils.fillNan(obs, maskF))
for x in yfLst
]
# if 'post' in doLst:
caseLst = ['Predict'] + [str(nd) + 'd latency' for nd in dLst]
keyLst = list(statLst[0].keys())
dataBox = list()
for iS in range(len(keyLst)):
key = keyLst[iS]
temp = list()
temp.append(statP[key])
print(key, np.nanmedian(statP[key]))
for k in range(len(statLst)):
temp = 0
for i in range(nt):
maskDay[j, i] = temp
if maskObs[j, i] == 1:
temp = 1
else:
if temp != 0:
temp = temp + 1
ind = np.random.randint(0, ngrid)
maskObsDay = maskObs * maskDay
unique, counts = np.unique(maskObsDay, return_counts=True)
maskF = (maskDay >= 1) & (maskDay <= 3)
statPLst = list()
statFLst = list()
for k in range(3):
statP = stat.statError(utils.fillNan(ypLst[k], maskF),
utils.fillNan(obs, maskF))
statF = stat.statError(utils.fillNan(yfLst[k], maskF),
utils.fillNan(obs, maskF))
statPLst.append(statP)
statFLst.append(statF)
cropFile = r'/mnt/sdb/Data/Crop/cropRate_CONUSv2f1.csv'
cropRate = pd.read_csv(cropFile, dtype=np.float, header=None).values
# croprate - 0 corn, 4 soybean, 22 spring wheat, 23 winter wheat
dataGrid = [
(statPLst[0]['RMSE'] - statFLst[0]['RMSE']) / statPLst[0]['RMSE'],
(statPLst[1]['RMSE'] - statFLst[1]['RMSE']) / statPLst[1]['RMSE'],
(statPLst[2]['RMSE'] - statFLst[2]['RMSE']) / statPLst[2]['RMSE'],
]
prcp = df.getDataTs('APCP_FORA').squeeze()
dataTs = [[obs, ypLst[0], yfLst[0]], [obs, ypLst[1], yfLst[1]],
subset='CONUSv4f1',
tRange=[sd, ed])
obs = df.getData(varT='SMAP_AM', doNorm=True, rmNan=False)
modelName = 'LSTM-DA-' + str(k)
model = train.loadModel(outFolder, nEpoch, modelName=modelName)
yP = train.testModel(model, (x, obs), batchSize=100).squeeze()
ypLst.append(
dbCsv.transNorm(yP,
rootDB=rootDB,
fieldName='SMAP_AM',
fromRaw=False))
if 'post' in doLst:
statDictLst = list()
for k in range(0, len(ypLst)):
statDictLst.append(stat.statError(ypLst[k], yT))
keyLst = ['RMSE', 'ubRMSE', 'Bias', 'Corr']
caseLst = ['LSTM']
for k in dLst:
caseLst.append('DA-' + str(k))
# plot box
dataBox = list()
cmap = plt.cm.jet
cLst = cmap(np.linspace(0, 1, len(caseLst)))
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
temp.append(statDictLst[k][statStr])
dataBox.append(temp)
maskDay[j, i] = temp
if maskObs[j, i] == 1:
temp = 1
else:
if temp != 0:
temp = temp + 1
ind = np.random.randint(0, ngrid)
maskObsDay = maskObs * maskDay
maskF = (maskDay >= 1) & (maskDay <= 3)
# figure out train and test time index
tR0 = [20150402, 20180401]
tA0 = utils.time.tRange2Array(tR0)
nt = len(tA0)
tTrainLst = list()
tTestLst = list()
for k in range(len(yrLst)):
tR = tRangeLst[k]
tA = utils.time.tRange2Array(tR)
ind0 = np.array(range(nt))
ind1, ind2 = utils.time.intersect(tA0, tA)
tTestLst.append(np.delete(ind0, ind1))
tTrainLst.append(ind1)
# calculate stat
for k in range(len(yrLst)):
yfTemp = utils.fillNan(yfLst[k], maskF)
yfTemp = yfTemp[:, tTestLst[k]]
statP = stat.statError(yfTemp, utils.fillNan(obs, maskF))
statF = stat.statError(yfTemp, utils.fillNan(obs, maskF))
if iEns == 0:
predLst = predtempLst
else:
for ii in range(len(outLst)):
predLst[ii] = np.concatenate([predLst[ii], predtempLst[ii]],
axis=2)
# predLst: List of all experiments with shape: Ntime*Nbasin*Nensemble
# get the ensemble mean from simulations of different seeds
ensLst = []
for ii in range(len(outLst)):
temp = np.nanmean(predLst[ii], axis=2, keepdims=True)
ensLst.append(temp)
# plot boxplots for different experiments
statDictLst = [stat.statError(x.squeeze(), obsAll.squeeze()) for x in ensLst]
keyLst = ["NSE", "KGE"] # which metric to show
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
data = statDictLst[k][statStr]
data = data[~np.isnan(data)]
temp.append(data)
dataBox.append(temp)
plt.rcParams["font.size"] = 14
labelname = ["PUR", "PUR-FDC", "PUR-1/3FDC"]
xlabel = ["NSE", "KGE"]
fig = plot.plotBoxFig(dataBox, xlabel, labelname, sharey=False, figsize=(6, 5))
for i in range(nt):
maskDay[j, i] = temp
if maskObs[j, i] == 1:
temp = 1
else:
if temp != 0:
temp = temp + 1
maskObsDay = maskObs * maskDay
fLst = [1, 2, 3]
statLst = list()
for nf in fLst:
maskF = maskDay == nf
temp = list()
for yf in yfLst:
statErr = stat.statError(utils.fillNan(yf, maskF),
utils.fillNan(obs, maskF))
temp.append(statErr)
statLst.append(temp)
# load result from RK
dirRK = r'D:\\data\\Koster17\\'
fileNameLst = ['rmse_lead_{}.dat'.format(x) for x in [1, 2, 3]]
tempLst = list()
for k in range(3):
# lon lat are identical. Tested
temp = np.loadtxt(os.path.join(dirRK, fileNameLst[k]))
tempLst.append(temp[:, 2])
RKlon = temp[:, 0]
RKlat = temp[:, 1]
lat, lon = df.getGeo()
errLst = list()
temp = 1
else:
if temp != 0:
temp = temp + 1
ind = np.random.randint(0, ngrid)
print(np.array([maskObs[ind, :], maskDay[ind, :]]))
maskObsDay = maskObs * maskDay
unique, counts = np.unique(maskDay, return_counts=True)
print(np.asarray((unique, counts)).T)
print(counts / ngrid / nt)
fLst = [1, 2, 3]
statLstF = list()
statLstP = list()
maskF = (maskDay >= 1) & (maskDay <= 3)
statP = stat.statError(utils.fillNan(yp, maskF), utils.fillNan(obs, maskF))
statF = stat.statError(utils.fillNan(yf, maskF), utils.fillNan(obs, maskF))
for nf in fLst:
xp = np.full([ngrid, nt], np.nan)
xf = np.full([ngrid, nt], np.nan)
y = np.full([ngrid, nt], np.nan)
xf[maskObsDay == nf] = yf[maskObsDay == nf]
xp[maskObsDay == nf] = yp[maskObsDay == nf]
y[maskObsDay == nf] = obs[maskObsDay == nf]
statLstF.append(stat.statError(xf, y))
statLstP.append(stat.statError(xp, y))
# plot box - forecast
matplotlib.rcParams.update({'font.size': 11})
matplotlib.rcParams.update({'lines.linewidth': 2})
matplotlib.rcParams.update({'lines.markersize': 12})
maskDay = np.zeros(maskObs.shape).astype(int)
ngrid, nt = maskObs.shape
for j in range(ngrid):
temp = 0
for i in range(nt):
maskDay[j, i] = temp
if maskObs[j, i] == 1:
temp = 1
else:
if temp != 0:
temp = temp + 1
ind = np.random.randint(0, ngrid)
maskObsDay = maskObs * maskDay
unique, counts = np.unique(maskObsDay, return_counts=True)
maskF = (maskDay >= 1) & (maskDay <= 3)
statP = stat.statError(utils.fillNan(yp, maskF), utils.fillNan(obs, maskF))
statF = stat.statError(utils.fillNan(yf, maskF), utils.fillNan(obs, maskF))
maskObsDay = maskObs * maskDay
print(np.array([maskObs[ind, :], maskDay[ind, :]]))
print(np.asarray((unique, counts)).T)
print(counts / ngrid / nt)
# see result for different seasons
tRangeLst = [[20160401, 20160701], [20160701, 20161001], [20161001, 20170101],
[20170101, 20170401], [20170401, 20170701], [20170701, 20171001],
[20171001, 20180101], [20180101, 20180401]]
tAllA = utils.time.tRange2Array(tAllR)
statPLst = list()
statFLst = list()
# %% load data and stat
kcLst = [7, 8, 13]
tRange = [20160401, 20180401]
statLst = list()
statRefLst = list()
for kc in kcLst:
tempLst = list()
for k in range(1, 18):
testName = subsetLst[kc - 1]
if k != kc:
outName = 'ecoRegion{:02d}{:02d}_v2f1_Forcing'.format(kc, k)
else:
outName = 'ecoRegion{:02d}_v2f1_Forcing'.format(kc)
out = os.path.join(pathSMAP['Out_L3_NA'], 'ecoRegion', outName)
df, yp, yt = master.test(out, tRange=tRange, subset=testName)
temp = stat.statError(yp[:, :, 0], yt[:, :, 0])
tempLst.append(temp)
if k == kc:
statRefLst.append(temp)
statLst.append(tempLst)
# %% plot box
keyLst = stat.keyLst
ecoLst = ['{:02d}'.format(x) for x in range(1, 18)]
caseLst = ['{:02d}'.format(x) for x in [7, 8, 13]]
for k in range(len(caseLst)):
dataBox = list()
key = 'RMSE'
for ii in range(len(ecoLst)):
temp = list()
np.save(os.path.join(out, 'obs.npy'), obs)
f = np.load(
os.path.join(out, 'x.npy')
) # it has been saved previously in the out directory (forcings)
T = (f[:, :, 3] + f[:, :, 4]) / 2 # mean air T for T_residual
T_air = np.expand_dims(T, axis=2)
pred_res = pred - T_air
obs_res = obs - T_air
predLst_res.append(pred_res)
obsLst_res.append(obs_res)
# calculate statistic metrics
# statDict = stat.statError(pred.squeeze(), obs.squeeze())
# statDictLst.append([statDict])
# statDictLst1 = [stat.statError(x.squeeze(), obs.squeeze()) for x, y in predLst]
statDictLst = [
stat.statError(x.squeeze(), y.squeeze())
for (x, y) in zip(predLst, obsLst)
]
statDictLst_res = [
stat.statError_res(x.squeeze(), y.squeeze(), z.squeeze(), w.squeeze())
for (x, y, z, w) in zip(predLst, obsLst, predLst_res, obsLst_res)
]
# median and STD calculation
count = 0
mdstd = np.zeros([len(statDictLst_res[0]), 3])
for i in statDictLst_res[0].values():
median = np.nanmedian((i)) # abs(i)
STD = np.nanstd((i)) # abs(i)
mean = np.nanmean((i)) #abs(i)
k = np.array([[median, STD, mean]])
cDir = os.path.dirname(os.path.abspath(__file__))
out = os.path.join(cDir, 'output', 'CONUSv4f1')
rootDB = os.path.join(cDir, 'data')
nEpoch = 100
tRange = [20160401, 20170401]
# load data
df, yp, yt = master.test(
out, tRange=[20160401, 20170401], subset='CONUSv4f1', epoch=100, reTest=True)
yp = yp.squeeze()
yt = yt.squeeze()
# calculate stat
statErr = stat.statError(yp, yt)
dataGrid = [statErr['RMSE'], statErr['Corr']]
dataTs = [yp, yt]
t = df.getT()
crd = df.getGeo()
mapNameLst = ['RMSE', 'Correlation']
tsNameLst = ['LSTM', 'SMAP']
# plot map and time series
plot.plotTsMap(
dataGrid,
dataTs,
lat=crd[0],
lon=crd[1],
t=t,
mapNameLst=mapNameLst,
dfC = dbCsv.DataframeCsv(rootDB=rootDB, subset='CONUSv2f1', tRange=tRange)
latC, lonC = dfC.getGeo()
errLstAll = list()
for case in caseLst:
testName = subsetPattern.format(case, 3)
errLst = list()
for k in levLst:
if k in [0, 1]:
subset = 'ecoReg_{}_L{}_v2f1'.format(case, k)
else:
subset = subsetPattern.format(case, k)
outName = subset + '_Forcing'
out = os.path.join(pathSMAP['Out_L3_NA'], 'ecoRegionCase', outName)
df, yp, yt = master.test(out, tRange=tRange, subset=testName)
err = stat.statError(yp[:, :, 0], yt[:, :, 0])
errLst.append(err)
errLstAll.append(errLst)
# plot box
cLst = 'ygbr'
keyLst = ['RMSE', 'Corr']
for key in keyLst:
dataBox = list()
for errLst in errLstAll:
temp = list()
for err in errLst:
temp.append(err[key])
dataBox.append(temp)
fig = figplot.boxPlot(dataBox,
label1=caseLabLst,
model = train.loadModel(outFolder, nEpoch, modelName=modelName)
yP = train.testModel(model, (x, obs), c, batchSize=100).squeeze()
ypLstmLst.append(
dbCsv.transNorm(yP,
rootDB=rootDB,
fieldName='SMAP_AM',
fromRaw=False))
if 'post' in doLst:
# stat
ypLst = [ypLstmLst, ypAnnLst]
statDictLst = list()
for i in range(0, len(ypLst)):
tempLst = list()
for j in range(0, len(ypLst[i])):
tempLst.append(stat.statError(ypLst[i][j], yT))
statDictLst.append(tempLst)
keyLst = list(tempLst[0].keys())
# plot box
dataBox = list()
caseLst1 = keyLst
caseLst2 = ['LSTM', 'LSTM-DA']
for iS in range(len(keyLst)):
statStr = keyLst[iS]
dataBox = list()
for iS in range(len(keyLst)):
statStr = keyLst[iS]
temp = list()
for k in range(len(statDictLst)):
temp.append(statDictLst[k][statStr])
dfz2 = camels.DataframeCsv(subset='all', tRange=[20141227, 20091227])
z2 = dfz2.getDataObs(doNorm=True, rmNan=False)
df2 = camels.DataframeCsv(subset='all', tRange=[20100101, 20150101])
x2 = df2.getDataTS(varLst=camels.forcingLst, doNorm=True, rmNan=True)
c2 = df2.getDataConst(varLst=camels.attrLstSel, doNorm=True, rmNan=True)
yt2 = df2.getDataObs(doNorm=False, rmNan=False).squeeze()
model = train.loadModel(outFolder, 100, modelName='test')
yp1 = train.testModel(model, x1, c1)
yp1 = camels.transNorm(yp1, 'usgsFlow', toNorm=False).squeeze()
yp2 = train.testModel(model, x2, c2)
yp2 = camels.transNorm(yp2, 'usgsFlow', toNorm=False).squeeze()
statErr1 = stat.statError(yp1, yt2)
statErr2 = stat.statError(yp2, yt2)
dataMap = [statErr2['Corr'], statErr1['Corr'] - statErr2['Corr']]
dataTs = [yt2, yp2]
t = df2.getT()
crd = df2.getGeo()
mapNameLst = ['Test Corr', 'Train Corr - Test Corr']
tsNameLst = ['USGS', 'LSTM']
colorMap = None
colorTs = None
import imp
imp.reload(plot)
plot.plotTsMap(dataMap,
dataTs,
lat=crd[:, 0],
