我有一个问题，我计算模型和参考数据集的不同评估方法-数据集曾经是三维（x，y，t）和netcdf数据。评估方法必须处理四种不同类型的时间聚合：无聚合，整个聚合，带重采样的聚合（每月）和分组聚合（每小时加班，每月超过年）。
我选择的工作马是python-xarray，因为它们在netcdf处理方面非常灵活。
从轻松得分BIAS开始，我有以下代码：
def BIAS_temporal(self,varns_result=None,aggregtime=None,dim_time=None):
""" Compute the BIAS/ME for a timeseries -> e.g. R^3 --> R^2:
BIAS(x,y) = SUM_{dim_t} ( MOD(x,y,dim_t) - OBS(x,y,dim_t) ) """
def bias(x):
return ( x[namex] - x[namey] )
def biasmean(x):
return ( x[namex] - x[namey] ).mean(dim=coordtime)
#
endresult = xarray.Dataset()
for varnsproof,varnsref,varnsres in zip(self.varns_proof,self.varns_ref,varns_result):
# rename the data variables and combine both datasets for evaluation
namex=varnsres+"_x"; namey=varnsres+"_y"
self.DSref.rename({varnsref : namey },inplace=True)
self.DSproof.rename({varnsproof : namex },inplace=True)
DScomb = xarray.merge([self.DSref,self.DSproof])
coordtime = self.MetricCalcProg.ListDimsToXarray(dim_time[varnsref]) #extract the name of time coordinate
#
if aggregtime == 'fullperiod':
DSnew = biasmean(DScomb).to_dataset(name = varnsres)
...
elif aggregtime == '-':
DSnew = bias(DScomb).to_dataset(name = varnsres)
elif "overperiod" in aggregtime:
grpby_method=self.MetricCalcProg.ConvertAggregKey2Groupby(aggregtime)
DSnew = DScomb.groupby(coordtime+'.'+grpby_method).apply(biasmean)
DSnew = DSnew.to_dataset(name = varnsres)
...
elif "overperiod" not in aggregtime:
resamplefreq=self.MetricCalcProg.ConvertAggregKey2Resample(aggregtime)
DSnew = DScomb.resample(time=resamplefreq,keep_attrs=True).apply(biasmean)
DSnew = DSnew.to_dataset(name = varnsres)
...
#
self.DSref.rename({namey : varnsref },inplace=True)
self.DSproof.rename({namex : varnsproof },inplace=True)
unitsnew=GeneralUtils.safe_access_bib('',KeyError,dicttest=self.DSref[varnsref].attrs,dictkey='units',errhandle=False)
if unitsnew is None: unitsnew='-'
longnew="temporal BIAS of "+str(GeneralUtils.safe_access_bib('', KeyError, dicttest=self.DSref[varnsref].attrs, \
dictkey='long_name',errhandle=False))
self.Update_Attributes(Datasetobj=DSnew,variable=varnsres,stdname=varnsres,units=unitsnew,longname=longnew)
endresult = xarray.merge([endresult,DSnew])
return endresult
如您所见，尽管使用了聚合方法“ aggregtime”，但我仍需要定义函数“ bias”和“ biasmean”以始终获得正确的结果。定义另一个类似“ LinearCorrelation”的指标的代码几乎相同：
def LinCorr_temporal(self,varns_result=None,aggregtime=None,dim_time=None):
""" Compute the linear correlation for a timeseries -> e.g. R^3 --> R^2:
LinCorr(x,y) = ???"""
def correl2dtm(x):
a = ( x[namex] - x[namex].mean(dim=coordtime) ) * ( x[namey] - x[namey].mean(dim=coordtime) ) \
/ x[namex].std(dim=coordtime) / x[namey].std(dim=coordtime)
return a.mean(dim=coordtime)
#
endresult = xarray.Dataset()
for varnsproof,varnsref,varnsres in zip(self.varns_proof,self.varns_ref,varns_result):
# ensure that only time stamps are considerd being present in both datasets (otherwise LC>1)
self.DSproof = self.DSproof[varnsproof].where(self.DSref[varnsref].notnull()).to_dataset(name=varnsproof) #harmonize the nan's between both datasets (independent calculations would destroy lincor), part A
self.DSref = self.DSref[varnsref].where(self.DSproof[varnsproof].notnull()).to_dataset(name=varnsref) #harmonize the nan's between both datasets, part B
# rename the data variables and combine both datasets for evaluation
namex=varnsres+"_x"; namey=varnsres+"_y"
self.DSref.rename({varnsref : namey },inplace=True)
self.DSproof.rename({varnsproof : namex },inplace=True)
DScomb = xarray.merge([self.DSref,self.DSproof])
coordtime = self.MetricCalcProg.ListDimsToXarray(dim_time[varnsref]) #extract the name of time coordinate
#
if aggregtime == 'fullperiod':
DSnew = correl2dtm(DScomb).to_dataset(name = varnsres)
...
elif aggregtime == '-':
print(' The Linear Corr. computation makes no sense for each single time step ')
exit()
elif "overperiod" in aggregtime:
grpby_method=self.MetricCalcProg.ConvertAggregKey2Groupby(aggregtime)
DSnew = DScomb.groupby(coordtime+'.'+grpby_method).apply(correl2dtm)
DSnew = DSnew.to_dataset(name = varnsres)
...
elif "overperiod" not in aggregtime:
resamplefreq=self.MetricCalcProg.ConvertAggregKey2Resample(aggregtime)
DSnew = DScomb.resample(time=resamplefreq,keep_attrs=True).apply(correl2dtm)
DSnew = DSnew.to_dataset(name = varnsres)
...
#
self.DSref.rename({namey : varnsref },inplace=True)
self.DSproof.rename({namex : varnsproof },inplace=True)
unitsnew='1'
longnew="temporal lin. correl. of "+str(GeneralUtils.safe_access_bib('', KeyError, dicttest=self.DSref[varnsref].attrs, \
dictkey='long_name',errhandle=False))
self.Update_Attributes(Datasetobj=DSnew,variable=varnsres,stdname=varnsres,units=unitsnew,longname=longnew)
endresult = xarray.merge([endresult,DSnew])
return endresult
如您所见，与“ BIAS”计算相比，只有很小的差异。功能定义不同。聚合时间“-”的例外仅发生在“ LinCorr_temporal”内，并且where语句仅选择存在“非零”数据的那些网格点。
您是否建议将此处定义的两种方法结合在一起，以定义函数“ def bias”和“ def correl2dtm”？我猜到现在为止我还错过了一种编程技术，即动态定义函数。
答案 0 :(得分：2)
函数是python中的对象，因此您可以将它们分配给变量。您可以将它们作为函数参数传递：
def bias(x, namex, namey, coordtime=False):
return (x[namex] - x[namey])
def correl2dtm(x, namex, namey, coordtime):
a = (x[namex] - x[namex].mean(dim=coordtime)) * (x[namey] - x[namey].mean(dim=coordtime)) \
/ x[namex].std(dim=coordtime) / x[namey].std(dim=coordtime)
return a.mean(dim=coordtime)
def temporal(self, func, varns_result=None,aggregtime=None,dim_time=None):
...
DSnew = func(DScomb, namex, namey, coordtime).to_dataset(name = varnsres)
...
# call temporal() with bias function (= old BIAS_temporal)
temporal(bias, ...)
# call temporal() with correl2dtm (= old LinCorr_temporal)
temporal(correl2dtm, ...)