Evaluation of global fine-resolution precipitation products and their uncertainty quantification in

18 Pages·2016·3.45 MB·English

Hydrol.EarthSyst.Sci.,20,903–920,2016
www.hydrol-earth-syst-sci.net/20/903/2016/
doi:10.5194/hess-20-903-2016
©Author(s)2016.CCAttribution3.0License.
Evaluation of global fine-resolution precipitation products and their
uncertainty quantification in ensemble discharge simulations
W.Qi1,2,C.Zhang1,G.Fu2,C.Sweetapple2,andH.Zhou1
1SchoolofHydraulicEngineering,DalianUniversityofTechnology,Dalian116024,China
2CentreforWaterSystems,CollegeofEngineering,MathematicsandPhysicalSciences,UniversityofExeter,
NorthParkRoad,HarrisonBuilding,Exeter,EX44QF,UK
Correspondenceto:C.Zhang([email protected])
Received:9July2015–PublishedinHydrol.EarthSyst.Sci.Discuss.:10September2015
Revised:11December2015–Accepted:2February2016–Published:26February2016
Abstract. The applicability of six fine-resolution precipi- drological models contribute most of the uncertainty in ex-
tation products, including precipitation radar, infrared, mi- treme discharges. Interactions between precipitation prod-
crowaveandgauge-basedproducts,usingdifferentprecipita- ucts and hydrological models can have the similar magni-
tion computation recipes, is evaluated using statistical and tude of contribution to discharge uncertainty as the hydro-
hydrological methods in northeastern China. In addition, logicalmodels.Abetterprecipitationproductdoesnotguar-
a framework quantifying uncertainty contributions of pre- antee a better discharge simulation because of interactions.
cipitation products, hydrological models, and their interac- It is also found that a good discharge simulation depends
tions to uncertainties in ensemble discharges is proposed. on a good coalition of a hydrological model and a precipi-
The investigated precipitation products are Tropical Rain- tation product, suggesting that, although the satellite-based
fall Measuring Mission (TRMM) products (TRMM3B42 precipitationproductsarenotasaccurateasthegauge-based
and TRMM3B42RT), Global Land Data Assimilation Sys- products, they could have better performance in discharge
tem(GLDAS)/Noah,AsianPrecipitation–Highly-Resolved simulationswhenappropriatelycombinedwithhydrological
Observational Data Integration Towards Evaluation of Wa- models. This information is revealed for the first time and
terResources(APHRODITE),PrecipitationEstimationfrom verybeneficialforprecipitationproductapplications.
Remotely Sensed Information using Artificial Neural Net-
works (PERSIANN), and a Global Satellite Mapping of
Precipitation (GSMAP-MVK+) product. Two hydrological
models of different complexities, i.e. a water and energy 1 Introduction
budget-based distributed hydrological model and a physi-
cally based semi-distributed hydrological model, are em- Knowledgeofprecipitationplaysanimportantroleintheun-
ployedtoinvestigatetheinfluenceofhydrologicalmodelson derstandingofthewatercycle,andthusinthemanagement
simulated discharges. Results show APHRODITE has high of water resources (Sellers, 1997; Sorooshian et al., 2005;
accuracy at a monthly scale compared with other products, Wang et al., 2005; Ebert et al., 2007; Buarque et al., 2011;
and GSMAP-MVK+ shows huge advantage and is better Tapiadoretal.,2012;Yongetal.,2012;GaoandLiu,2013;
thanTRMM3B42inrelativebias(RB),Nash–Sutcliffecoef- Peng et al., 2014a, b). However, precipitation data are not
ficientofefficiency(NSE),rootmeansquareerror(RMSE), availableinmanyregions,particularlymountainousdistricts
correlation coefficient (CC), false alarm ratio, and critical andruralareasindevelopingcountries.Forexample,north-
success index. These findings could be very useful for vali- eastChina,whichplaysanimportantroleinfoodproduction
dation,refinement,andfuturedevelopmentofsatellite-based to support the country’s population and is also an industrial
products (e.g. NASA Global Precipitation Measurement). region with many heavy industries, frequently suffers from
Althoughlargeuncertaintyexistsinheavyprecipitation,hy- drought,posingathreattoregionalsustainabledevelopment.
Insuchareas,duetoinsufficientgaugeobservations,alterna-
PublishedbyCopernicusPublicationsonbehalfoftheEuropeanGeosciencesUnion.
904 W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification
tiveprecipitationdataarerequiredforefficientmanagement by Beven and Binley (1992). This approach outputs proba-
ofwaterresources. bility distributions of model parameters conditioned on ob-
In recent years, implementation of gauge-based and re- serveddata,andtheuncertaintiesinmodelinputsarerepre-
motesatellite-basedprecipitationproductshasbecomepop- sented by uncertain parameters. Similar to GLUE, Hong et
ular, particularly for ungauged catchments (Artan et al., al.(2006)proposedaMonteCarlo-basedmethodtoquantify
2007;Jiangetal.,2012;Lietal.,2013;MüllerandThomp- uncertainty in hydrological simulations using satellite pre-
son, 2013; Maggioni et al., 2013; Xue et al., 2013; Kneis cipitation data, in which flow simulation uncertainty is rep-
et al., 2014; Meng et al., 2014; Ochoa et al., 2014). Nu- resentedbyensemblesimulationresults.
merous precipitation products have been developed to esti- In addition to individual contributions from hydrological
materainfall,forexample:TropicalRainfallMeasuringMis- modelsandprecipitationdata,theinteractionsbetweenpre-
sion(TRMM)products(Huffmanetal.,2007),GlobalLand cipitation products and hydrological models also contribute
Data Assimilation System (GLDAS) precipitation products touncertaintyinsimulateddischarges.However,tothebest
(Kato et al., 2007), Asian Precipitation – Highly-Resolved of our knowledge, the previous studies have not quantified
Observational Data Integration Towards Evaluation of Wa- therespectivecontributionsofprecipitationproducts,hydro-
ter Resources (APHRODITE) (Xie et al., 2007; Yatagai et logical models, and their interactions to the total discharge
al., 2012), Precipitation Estimation from Remotely Sensed simulationuncertainty.
InformationusingArtificialNeuralNetworks(PERSIANN) The overall objectives of this paper are (1) to inves-
(Sorooshian et al., 2000, 2002), and Global Satellite Map- tigate the applicability of six fine-resolution precipitation
pingofPrecipitationproduct(GSMAP)(Kubotaetal.,2007; products using both statistical and hydrological evalua-
Aonashietal.,2009). tion methods in a small river basin in northeast China;
There are uncertainties in these products. Several studies (2) to propose a framework to quantify the contributions
have been carried out to analyse the uncertainty of TRMM of various uncertainties from precipitation products, hy-
inhigh-latituderegions(Yongetal.,2010,2012,2014;Chen drological models, and their interactions to uncertainty
etal.,2013a;ZhaoandYatagai,2014),butstudiesinnorth- in simulated discharges. The precipitation products inves-
east China are few. Evaluation of GLDAS data has gener- tigated are TRMM3B42, TRMM3B42RT, GLDAS/Noah
allybeenlimitedtotheUnitedStatesandotherobservation- (GLDAS_Noah025SUBP_3H),APHRODITE,PERSIANN,
rich regions of the world (Kato et al., 2007); assessments andGSMAP-MVK+.Twohydrologicalmodelsofdifferent
andapplicationsinotherregionsarerare(Wangetal.,2011; complexities–awater-andenergy-budget-baseddistributed
Zhou et al., 2013). The APHRODITE, PERSIANN, and hydrologicalmodel(WEB-DHM)(Wangetal.,2009a,b,c)
GSMAP products are seldom evaluated in northeast China and a physically based semi-distributed hydrological model
using basin-scale gauge data (Zhou et al., 2008). Owing to TOPMODEL (Beven and Kirkby, 1979) – were employed
the high heterogeneity of rainfall across a variety of spa- to investigate the influence of hydrological models on dis-
tiotemporalscales,theuncertaintycharacteristicsofprecipi- charge simulations. The respective uncertainties from pre-
tationproductsarevariable(Asadullahetal.,2008;Dinkuet cipitation products, hydrological models, and the combined
al.,2008;Nikolopoulosetal.,2010;Panetal.,2010).Thus, uncertainties from the interactions between products and
innortheastChina,itisessentialtocompletelyevaluatethe modelsarequantifiedusingaglobalsensitivityanalysisap-
applicabilityoftheseprecipitationproducts.Inaddition,itis proach, i.e. the analysis of variance approach (ANOVA). A
also worth comparing the performance of different precipi- riverbasinwithaseriesof8-yeardataisusedtodemonstrate
tationcomputationrecipes:forexample,theartificialneural themethodology.
networkfunctionusedinPERSIANN,thehistogrammatch- Thepaperisorganizedasfollows.Section2introducesthe
ingapproachusedinTRMM3B42,andthecloudmotionvec- study region, precipitation products, hydrological models,
torsusedinGSMAP-MVK+,becausetheinter-comparison and the proposed framework. Section 3 presents the statis-
couldrevealthestrategiesthatcouldbeusedtoobtainmore ticalevaluationresults.Hydrologicalevaluationsandtheim-
accurateprecipitationdata. plementationoftheproposedframeworkaregiveninSect.4.
Researchers have implemented precipitation products in DiscussionisgiveninSect.5.Summaryandconclusionsare
discharge simulations and reported discharge uncertainties presentedinSect.6.
(Hong et al., 2006; Pan et al., 2010; Serpetzoglou et al.,
2010). Also, many uncertainty analysis approaches have
beenintroducedtoquantifytheuncertainty(BevenandBin- 2 Materialsandmethodology
ley, 1992; Freer et al., 1996; Kuczera and Parent, 1998;
Beven and Freer, 2001b; Peters et al., 2003; Heidari et al., 2.1 Biliubasin
2006; Kuczera et al., 2006; Tolson and Shoemaker, 2007;
Blasone et al., 2008; Vrugt et al., 2009). In these prior ap- Biliubasin(2814km2),locatedinthecoastalregionbetween
proaches,oneofthepopularmethodsisthegeneralizedlike- theChinaBohaiSeaandtheChinaHuanghaiSea,coverslon-
lihooduncertaintyestimation(GLUE)technique,introduced gitudes122.29to122.92◦Eandlatitudes39.54to40.35◦N.
Hydrol.EarthSyst.Sci.,20,903–920,2016 www.hydrol-earth-syst-sci.net/20/903/2016/
W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification 905
Table1.Precipitationproducts.
Product Spatial Temporal Arealcoverage Startdate Type
resolution resolution
TRMM3B42 0.25◦ 3h Global50◦N–S 1Jan1998 PR+IR+MW+gauge+HM
TRMM3B42RT 0.25◦ 3h Global50◦N–S 1Mar2000 IR+MW
GLDAS/Noah 0.25◦ 3h Global90◦N–60◦S 24Feb2000 IR+MW+gauge
GSMAP-MVK+ 0.1◦ 1h Global60◦N–S 1Mar2000 IR+MW+CMV
PRRSIANN 0.25◦ 3h Global60◦N–S 1Mar2000 PR+IR+MW+ANN
◦
◦ 60–150 E, 1Jan1961
APHRODITE 0.25 1day ◦ ◦ gauge
15 S–55 N to2007
PR:precipitationradar;IR:infraredestimation;MW:microwaveestimation;HM:histogrammatching;CMV:cloudmotionvectors;ANN:artificialneural
network.
Figure1.Biliubasin:(a)thelocationofLiaoningprovincewithinChina;(b)thelocationofBiliubasinwithinLiaoningprovince;(c)the
◦
distributionsofraingauges,dischargegauge,automaticweatherstations,digitalelevationmodel,anddiagrammatic0.25 precipitationcells;
and(d)diagrammaticdescriptionofdownscalingthe0.25◦precipitationcellsto300m×300mcells,andretrievingthe300m×300mcells
locatedwithinthebasinboundary.
This basin is characterized by a snow, winter dry, and hot 2.2 Precipitationproducts
summerclimate(Koppenclimateclassification)andtheaver-
ageannualtemperatureis10.6◦C.Summer(JulytoSeptem-
The selected precipitation products are shown in Table 1.
ber)isthemajorrainyseason.Thereare11rainfallstations
These data are all freely available. In these selected pre-
and one discharge gauge which have historical data from
cipitationproducts,APHRODITEiswhollybasedongauge
January 2000 to December 2007. The average elevation is
data; TRMM3B42 and GLDAS are remote satellite estima-
240m. The gauge distribution in Biliu is shown in Fig. 1.
tions with gauge data corrections; while others are remote
The basin slopes vary from 0 to 38◦. Land-use data are ob-
satelliteestimationswithoutgaugedatacorrections.Remote-
tained from the USGS (http://edc2.usgs.gov/glcc/glcc.php).
basedprecipitationestimationhasmanyweaknesses;e.g.mi-
The land-use types have been reclassified to SiB2 land-use
crowave estimation could miss convective rainfall and ty-
typesforthisstudy(Sellersetal.,1996).Therearesixland-
phoonrainbecauseofitssparsetimeintervalresolution;in-
usetypes,withbroadleafandneedleleaftreesandshortveg-
fraredestimationhasahighertimeintervalresolution,butit
etationbeingthemaintypes.Soildataareobtainedfromthe
cannot penetrate thick clouds. Ground rain-gauge-based in-
FoodandAgricultureOrganization(FAO,2003)Globaldata
terpolationproductsarelimitedbyinterpolationalgorithms,
product, and there are two types of soil in the basin: clay
gaugedensity,andgaugedataquality(Xieetal.,2007).The
loamLuvisolsandloamPhaeozems.
detailsofdatasourcesusedineachprecipitationproductcan
www.hydrol-earth-syst-sci.net/20/903/2016/ Hydrol.EarthSyst.Sci.,20,903–920,2016
906 W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification
befoundinTable1.Thedetailedintroductionsoftheseprod- tioncriteriaareusedinrainfallamounterrorassessment:cor-
uctsareasfollows. relation coefficient (CC), root mean square error (RMSE),
TRMM is a joint mission between NASA and Japan Nash–Sutcliffe coefficient of efficiency (NSE), and relative
Aerospace Exploration Agency designed to monitor and bias(RB).Thesearecalculatedasfollows:
study tropical rainfall (Kummerow et al., 2000; Huffman et
1
al., 2007). Three instruments – a visible infrared radiome- Pn (cid:0)X −X (cid:1)22
ter, a TRMM microwave imager, and a precipitation radar - pi oi
areemployedtoobtainaccurateprecipitationestimation.The RMSE=i=1  (1)
 n 
TRMMprecipitationradaristhefirstspace-basedprecipita-
tion radar and operates between 35◦N and 35◦S. Outside
thisband,themicrowaveimagerisusedbetween40◦Nand Pn (cid:0)X −X (cid:1)2
pi oi
40◦S, and the visible infrared radiometer data are used be- NSE=1−i=1 (2)
tween50◦Nto50◦S.Usuallytheprecipitationradariscon- Pn (cid:0)X −X (cid:1)2
sidered to give the most accurate estimation from satellite, pi o
i=1
anddatafromitareoftenusedforcalibrationofpassivemi-
n n
crowavedatafromotherinstruments(Ebertetal.,2007).The PXpi−PXoi
post-real-timeproductusedinthisstudyistheTRMM3B42, RB= i=1 i=1 ×100%, (3)
n
which utilizes three data sources: the TRMM combined in- P
X
oi
strumentestimationusingdatafrombothTRMMprecipita- i=1
tionradarandthemicrowaveimager;theGPCPmonthlyrain
where X represents observed data; X represents esti-
gaugeanalysisdevelopedbytheGlobalPrecipitationClima- oi pi
mateddata;nisthetotalnumberofdatapoints.Aperfectfit
tology Center; and the Climate Assessment and Monitoring
shouldhaveCCandNSEvaluesof1.ThelowertheRMSE
Systemmonthlyraingaugeanalysis.TRMM3B42appliesan
andRB,thebettertheestimation.Thesecomparisoncriteria
infrared to rain rate relationship using histogram matching,
havebeenusedbymanystudies(Ebertetal.,2007;Wanget
while TRMM3B42RT merges microwave and infrared pre-
al.,2011;Yongetal.,2012),sotheyareusedinthisstudy.
cipitationestimation.
Probability distributions by occurrence and volume are
PERSIANN is a product that, using an artificial neural
alsoanalysed,whichcanprovideuswiththeinformationon
network function, estimates precipitation by combining in-
thefrequencyandontheproducterrordependenceonprecip-
fraredprecipitationestimationandtheTRMMcombinedin-
itationintensity(Chenetal.,2013a,b).Thecriticalsuccess
strumentestimation(whichassimilateswithTRMMprecipi-
tationradarandmicrowavedata).GSMAP-MVK+usesmi- index(CSI),probabilityofdetection(POD),andfalsealarm
ratio (FAR) are used to quantify the ability of precipitation
crowave and infrared precipitation data together and com-
productstodetectobservedrainfallevents.Thesearedefined
bines cloud motion vectors to generate fine-resolution pre-
asfollows:
cipitationestimation.
The Global Land Data Assimilation System (GLDAS)
H
project is an extension of the existing and more mature CSI= (4)
H+M+F
North American Land Data Assimilation System (Rodell et
H
al.,2004).Itintegratessatellite-andground-baseddatasets POD= (5)
H+M
for parameterizing, forcing, and constraining a few offline
F
landsurfacemodelsforgeneratingoptimalfieldsoflandsur- FAR= , (6)
H+F
face states and fluxes. At present, GLDAS drives four land
surface models: Mosaic (Koster and Suarez, 1992), Noah
where H is the total number of hits; M is the total number
(Chenetal.,1996;Bettsetal.,1997;Korenetal.,1999;Ek,
ofmisses;F isthetotalnumberoffalsealarms(Ebertetal.,
2003), the Community Land Model (Dai et al., 2003), and
2007; Su et al., 2008). A perfect detection should have CSI
theVariableInfiltrationCapacitymodel(Liangetal.,1994).
andPODvaluesequalto1andaFARvalueof0.
Amongthem,theGLDAS/NoahLandSurfaceModelprod-
uct(GLDAS_NOAH025SUBP_3H) hasa3h 0.25◦×0.25◦
2.4 Hydrologicalmodelsanddata
resolution, which is desirable for basin-scale research. The
GLDASprecipitationdatacombinemicrowaveandinfrared 2.4.1 WEB-DHM
data,andalsoassimilategaugeobservations.
The distributed biosphere hydrological model, WEB-DHM
2.3 Criteriaforaccuracyassessment (Wang et al., 2009a, b, c), was developed by coupling
a simple biosphere scheme (Sellers et al., 1986) with a
Uncertainties of precipitation products are evaluated on the geomorphology-based hydrological model (Yang, 1998) to
basis of basin-averaged rainfall observations. Four evalua- describe water, energy, and CO fluxes at a basin scale.
2
Hydrol.EarthSyst.Sci.,20,903–920,2016 www.hydrol-earth-syst-sci.net/20/903/2016/
W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification 907
WEB-DHMhasbeenusedinseveralevaluationsandappli-
cations(Wangetal.,2010a,b,2012;Shresthaetal.,2014).
WEB-DHMinputdataincludeprecipitation,temperature,
downwardsolarradiation,long-waveradiation,airpressure,
wind speed, and humidity. With the exception of precipita-
tion,allinputdataareobtainedfromautomaticweathersta-
tions.TherearethreeautomaticweatherstationsnearBiliu,
and observations from these are obtained from the China
Meteorological Data Sharing Service System (downloaded
from http://cdc.cma.gov.cn/home.do). Hourly precipitation
data are downscaled from daily rain gauge observations us-
ing a stochastic method (Wang et al., 2011). Hourly tem-
peraturesarecalculatedfromdailymaximumandminimum
Figure2.Diagrammaticflowchartoftheproposedframeworkfor
temperaturesusingatemperaturemodel(PartonandLogan,
quantification of uncertainty contributions to ensemble discharges
1981). The estimated temperatures are also further evalu-
simulatedusingprecipitationproductsonthebasisoftheanalysis
atedusingdailyaveragetemperature.Downwardsolarradi-
ofvariance(ANOVA)approach.
ation is estimated from sunshine duration, temperature, and
humidity using a hybrid model (Yang et al., 2006). Long-
wave radiation is obtained from the GLDAS/Noah (Rodell tionsofTOPMODELanditsmathematicalformulationcan
et al., 2004). Air pressure is estimated according to altitude befoundinBevenandKirkby(1979).TOPMODELhasbeen
(Yang et al., 2006). These meteorological data are then in- popularlyutilizedinresearchacrosstheworld(Blazkovaand
terpolatedto300m×300mmodelcellsthroughaninverse- Beven, 1997; Cameron et al., 1999; Hossain and Anagnos-
distanceweightingapproach.Becauseoftheelevationdiffer- tou,2005;Bastolaetal.,2008;Gallartetal.,2008;Bouilloud
encesamongmodelcellsandmeteorologicalgauges,thein- et al., 2010; Qi et al., 2013), because of its relatively sim-
terpolatedsurfaceairtemperaturesarefurthermodifiedwith plemodelstructure.TheinputdataofTOPMODELmainly
a lapse rate of 6.5Kkm−1. Gauge rainfall data are also in- include basin-averaged precipitation and topographic data
terpolatedto300m×300mmodelcellsandbasin-averaged whichcanbeestimatedfromDEM.
gauge rainfall data are calculated on the basis of interpola-
tionresults.Inadditiontotheabove,theleafareaindexand 2.5 Theproposedframework
fraction of photosynthetically active radiation data are ob-
tainedfromlevel-4MODISglobalproductMOD11A2.The Figure2showsthediagrammaticflowchartoftheproposed
digital elevation model (DEM) is from the NASA SRTM frameworkforquantificationofuncertaintycontributionsto
(Shuttle Radar Topographic Mission) with a resolution of ensembledischargessimulatedusingprecipitationproducts.
30m×30m.Weresampledtheresolutionto300minmodel This framework includes four parts: (a) selection of pre-
calculationtoreducecomputationcost,whilethemodelpro- cipitation products; (b) selection of hydrological models;
cessedfinerDEMs(30mgrid)togeneratesub-gridparame- (c) ensemble of discharge simulations using the hydrologi-
ters(suchashillslopeangleandlength). calmodelsandprecipitationproducts;and(d)quantification
of individual and interactive contributions using the analy-
2.4.2 TOPMODEL sis of variance (ANOVA) approach including contributions
fromprecipitationproducts,hydrologicalmodels,andinter-
TOPMODEL is a physically based, variable-contributing actions between models and products. Because the spatial
area model of basin hydrology which attempts to combine resolution of selected precipitation products does not cor-
the advantages of a simple lumped parameter model with respond with WEB-DHM model cells, the following pro-
distributed effects (Beven and Kirkby, 1979). Fundamental cedures were carried out for basin-averaged rainfall calcu-
to TOPMODEL’s parameterization are three assumptions: lations: (1) resampling 0.25 or 0.1◦ precipitation product
(1) saturated-zone dynamics can be approximated by suc- grids into 300m×300m cells (the grid size used in WEB-
cessive steady-state representations; (2) hydrological gradi- DHMsimulations);(2)calculatingbasin-averagedprecipita-
ents of the saturated zone can be approximated by the local tion using 300m precipitation product grids located within
topographic surfaceslope; and (3) thetransmissivity profile thebasinboundary.Diagrammaticdescriptionsofthesepro-
whose form declines exponentially with increasing vertical cedures are shown in Fig. 1d. Because WEB-DHM needs
depth of the water table or storage is spatially constant. On hourly input data, for the 3h resolution precipitation prod-
the basis of the above-mentioned assumptions, the index of ucts,weassumedrainfallisuniformlydistributedwithineach
hydrologicalsimilarityisrepresentedasthetopographicin- 3h period. For daily resolution products, we used the same
dex, ln(a/tanβ), for which a is the area per unit contour approach as downscaling observed precipitation data. This
lengthandβ isthelocalslopeangle.Moredetaileddescrip- downscaling approach may affect uncertainty in simulated
www.hydrol-earth-syst-sci.net/20/903/2016/ Hydrol.EarthSyst.Sci.,20,903–920,2016
908 W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification
discharge. However, Wang et al. (2011) have already suc- and K is the number of hydrological models (two in this
cessfullyappliedthedownscalingapproach,andshownthat study). Then the variation fraction η2 is calculated as fol-
theinfluenceisnegligible. lows:
The total ensemble uncertainty Y is the variance of dis-
charges. To relate Y to the uncertainty sources, the super- η2 = 1XI SSAi (14)
scriptsj andk inYj,k representacombinationofprecipita- precipitation I i=1 SSTi
tionproductj andhydrologicalmodelk:
Yj,k =Pj+Mk+PMj,k, (7) ηm2odel= I1XI SSSSBTi (15)
i=1 i
uwcht,erMe Preprreepsreenstesntthsetehfefecetffoecftktohfhjytdhroploregciicpailtamtioodnelp,raondd- ηi2nteraction= I1XI SSSSTIi . (16)
PM represents the interaction effect. In this study, j varies i=1 i
fromonetosix,andk variesfromonetotwo.Detailsofthe
η2 has a value between 0 and 1, which represent 0 and
quantificationareexplainedinthefollowsections.
100% contributions to the overall uncertainty of simulated
dischargesrespectively.I equals15inthisstudy.Asshown
2.5.1 Subsamplingapproach
inEqs.(14)–(16),thesubsamplingapproachisnecessarybe-
ANOVAcouldunderestimatevariancewhenthesamplesize cause it guarantees that every contributor has the same de-
is small (Bosshard et al., 2013). To reduce the effect of the nominator I. This same denominator makes sure that the
samplesize,Bosshardetal.(2013)proposedasubsampling inter-comparison among precipitation contribution, model
method, which was used in this paper. In the subsampling contribution,andinteractioncontributionisfreeofinfluence
method,thesuperscriptj inEq.(7)isreplacedwithg(h,i). fromthesamplingnumberofprecipitationproductsandhy-
AccordingtoBosshardetal.(2013),ineachsubsamplingit- drologicalmodels.
erationi,datafromtwoproductsshouldbeselectedoutofall
thesixproducts,andthus15combinationscanbeobtained.
3 Statisticalevaluations
Therefore,thesuperscriptgbecomesa2×15matrix:
(cid:18) (cid:19) 3.1 Dailyandmonthlyscales
1 1 ... 1 2 2 ... 4 4 5
g= . (8)
2 3 ... 6 3 4 ... 5 6 6
Comparison of precipitation product data and gauge obser-
2.5.2 Uncertaintycontributiondecomposition vations at a daily scale is shown in Fig. 3. Observations
are shown on the x axis and precipitation product data are
Based on the ANOVA theory (Bosshard et al., 2013), total shown on the y axis. Four criteria, RMSE, CC, NSE, and
errorvariance(SST)canbedividedintosumsofsquaresdue RB, are also shown. GSMAP-MVK+ is the best product
totheindividualeffectsas and PERSIANN has the poorest performance with respect
to RMSE and NSE. GSMAP-MVK+ is also the best with
SST=SSA+SSB+SSI, (9)
respect to CC, while GLDAS has the poorest performance
where SSA is the error contribution of precipitation prod- withaCCvalueof0.55.WithrespecttoRB,APHRODITE
ucts, SSB is the error contribution of hydrological models, performs best and GSMAP-MVK+ the second best, while
andSSIistheerrorcontributionoftheirinteractions. TRMM3B42RT the least best with an RB value of −38%.
Thetermscanbeestimatedusingthesubsamplingproce- None of the products can outperform others in terms of all
dureasfollows: thestatisticalcriteria.Thismaybeduetothedifferentlimi-
tationsofsatellitesensorsandinversealgorithmsofprecipi-
SST =XH XK (cid:16)Yg(h,i),k−Yg(o,i),o(cid:17)2 (10) tationproducts.Thissituationshowsthattheselectionofthe
i
bestprecipitationproductsisdifficult.
h=1k=1
TRMM3B42RT and TRMM3B42 underestimate precipi-
SSA =K·XH (cid:16)Yg(h,i),o−Yg(o,i),o(cid:17)2 (11) tation amounts. This underestimation may be because con-
i
vectiverainfallalwayshappensinsummerinnortheastChina
h=1
(Shou and Xu, 2007a, b; Yuan et al., 2010), and indicates
XK (cid:16) (cid:17)2
SSB =H· Yg(o,i),k−Yg(o,i),o (12) thelimitationofTRMMalgorithmsinhigh-latituderegions
i
k=1 withconvectiverainfall.Thistypeofrainfallhasalargerain-
SSIi=XH XK (cid:16)Yg(h,i),k−Yg(h,i),o−Yg(o,i),k+Yg(o,i),o(cid:17)2, (13) fnaoltl abmeocuanpttuwreidthibnyamshicorrotwtiamvee ipmeraigoedr.anTdh,isthteyrpeefooref,rcaainn--
h=1k=1 fall may also have a thick cloud that is impenetrable by
wheresymbolo indicatesaveragingoveraparticularindex; infrared (Ebert et al., 2007). Thus microwave and infrared
H isthenumberofprecipitationproducts(sixinthisstudy) estimation could underestimate rainfall. Compared with
Hydrol.EarthSyst.Sci.,20,903–920,2016 www.hydrol-earth-syst-sci.net/20/903/2016/
W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification 909
Figure3.Scatterplotsofbasin-averagedprecipitationproductsversusgaugeobservationsatadailyscale.
Figure4.Scatterplotsofbasin-averagedprecipitationproductsversusgaugeobservationsatamonthlyscale.
TRMM3B42RT, TRMM3B42 provides an improvement in RMSE, CC, NSE, and RB. GLDAS/Noah is the poorest in
RB. This improvement may be attributed to the assimila- termsofRMSEandNSE.WithrespecttoCC,GLDASand
tion with gauge data and histogram matching. Compared TRMM3B42 are equally poor, with CC values of 0.81. The
with APHRODITE and GSMAP-MVK+, TRMM3B42 has results also show that PERSIANN overestimates precipita-
lowaccuracyasrepresentedbyRB.Thisimpliesthatthere- tionamount,whileLietal.(2013)foundPERSIANNunder-
trieval algorithm used by TRMM3B42 still needs to be im- estimates rainfall in south China. This may be attributed to
proved with respect to RB. The reason why APHRODITE thedifferentlatitudesofthestudyregions.
outperforms TRMM3B42 is that APHRODITE is a gauge- Figure 5 shows time series of average monthly precipita-
based product. The fact that GSMAP-MVK+ outperforms tiondataagainstgaugeobservationsduringtheperiod2000–
TRMM3B42 in terms of RB may be due to the cloud 2007. Each curve represents a different precipitation prod-
motion vectors it uses. Compared with GSMAP-MVK+, uct.GLDASdata(Fig.5a)seriouslyunderestimatehighrain-
GLDAS/Noah precipitation shows low accuracy in all the fall. Similarly, TRMM3B42RT underestimates peak precip-
criteria even though they use similar data sources: IR and itation intensity also. Comparatively, APHRODITE, PER-
MW. SIANN,TRMM3B42,andGSMAP-MVK+havebetterper-
Comparison of precipitation product data and gauge ob- formances.
servations at a monthly scale is shown in Fig. 4. Here,
the APHRODITE product (Fig. 4d) performs best based on
www.hydrol-earth-syst-sci.net/20/903/2016/ Hydrol.EarthSyst.Sci.,20,903–920,2016
910 W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification
Figure5.Timeseriesplotsofbasin-averagedprecipitationproductvaluesversusgaugeobservationsatamonthlyscale.
Figure7.Probabilitydistributionsofthesixprecipitationproducts
Figure6.Inter-annualbasin-averagedmonthlyprecipitation.
byoccurrence(CDFc)andvolume(CDFv).
3.2 Inter-annualevaluations
3.3 Probabilitydistributionevaluations
Figure 6 shows the inter-annual average monthly precipita-
Figure7showsthecumulativeprobabilitydistributionfunc-
tion. Each curve represents a different data product. PER-
tion(CDF)byoccurrence(CDFc)andbyvolume(CDFv)for
SIANN overestimates in all the 12 months, while others
precipitationproducts.Probabilitiesareshownonthey axis,
underestimate, especially during the summer. This may re- and the x axis shows rainfall intensity with a 1mmday−1
sult from the artificial neural network function and limita-
intervallogspace.
tions of infrared and microwave estimation. APHRODITE
PERSIANN is the best by both occurrence and volume.
data are relatively close to observations. Compared with
However, for CDFc, TRMM3B42RT is the least best, and,
TRMM3B42RT,TRMM3B42isbetter,whichindicatesthat
forCDFv,TRMM3B42RT,andGLDAS/Noaharecompara-
the gauge corrections and histogram-matching used by
ble and worse than others. All precipitation products over-
TRMM3B42impactpositivelyonaccuracy.Duringthesum-
estimateoccurrenceandvolumeprobabilitiesexceptrainfall
mer, discrepancies between products become larger. With a intensitiesoflargerthan63and53mmday−1foroccurrence
decrease of rainfall magnitude, the discrepancies between
andvolumeprobabilities,respectively.Thismaybebecause
products reduce. This information implies that the differ-
theprecipitationproductsoverestimatetheintensityofsome
ences in precipitation estimation are related to precipitation
heavyrainfall(recalltheresultsinSect.3.1).Theresultsdif-
magnitudes:thelargertherainfallmagnitudes,thegreaterthe
ferfromthoseofLietal.(2013),inwhichPERSIANNhas
differences.
thepoorestperformance.Thismayresultfromdifferencesin
studyregion(inthestudyofLietal.(2013),southChinawas
studied).
Hydrol.EarthSyst.Sci.,20,903–920,2016 www.hydrol-earth-syst-sci.net/20/903/2016/
W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification 911
Figure8.Falsealarmratio,probabilityofdetection,andcriticalsuccessindexforthesixprecipitationproducts.
3.4 Contingencystatistics on rainfall magnitude, and implies that microwave and in-
fraredprecipitationestimationmayhaverelativelystrongde-
pendenceonrainfallmagnitudeintermsofprobabilityofde-
Figure8showsthefalsealarmratio,probabilityofdetection,
tectionandcriticalsuccessindex.
andcriticalsuccessindexforeachprecipitationproduct.
PERSIANN has the highest false alarm ratio among the
products, while TRMM3B42RT has the lowest. The false 4 Hydrologicalevaluations
alarm ratio of TRMM3B42 is larger than TRMM3B42RT,
which indicates that the gauge corrections and histogram- 4.1 Assessmentofhydrologicalmodels
matching used by TRMM3B42 do not provide positive ef-
fectsonfalsealarmratioandmaygiverisetouncertaintyin WEB-DHM was calibrated against observed discharges of
false alarm ratio. GSMAP-MVK+ has a lower false alarm Biliu. Six main parameters were selected to calibrate using
ratiothanTRMM3B42. a trial and error approach due to the model’s computational
No obvious trends are observed for the false alarm ratio burden. Model parameter multipliers were calibrated, simi-
overall(comparedwiththeprobabilityofdetectionandcrit- lar to the study by Wang et al. (2011). The trial and error
ical success index), which means the false alarm ratio de- approach has two steps. First, all the multiplier values are
pendence on rainfall magnitude is weak. However, Chen et set to 1 which represents the default parameter values from
al.(2013a)foundthefalsealarmratiosofTRMM3B42and theFoodandAgricultureOrganization(FAO,2003)andthe
TRMM3B42RTtoincreasewithanincreaseinrainfallinten- SiB2 model. Second, the multiplier values are varied until
sity. The differences are attributed mainly to observed data. acceptable discharge simulation accuracy is obtained. The
InthestudyofChenetal.(2013a),nationalraingaugedata calibrated parameter values are listed in Table 2. The sim-
were employed, whereas in this study more detailed basin ulated daily, monthly, and inter-annual results are shown in
dataareused. Fig.9a,c,ande.
Among all selected products, GLDAS/Noah has the low- TOPMODEL uses basin-averaged parameter values, and
est probability of detection and critical success index dur- theseparametervaluesareestimatedbyexperienceorobser-
ing periods of high rainfall intensity, while APHRODITE vation. However, these methods do not give precise param-
retains a high probability of detection and critical success eter values. Therefore, the parameter values are considered
index. This is because APHRODITE uses gauge observa- as uncertain and provided with ranges based on experience
tions, and implies that the APHRODITE algorithm is ef- (BevenandKirkby,1979;BevenandFreer,2001a,b;Peters
fective.PERSIANNhascomparableprobabilityofdetection etal.,2003).SixparametersofTOPMODELwerecalibrated
with APHRODITE. The critical success index of GSMAP- using the dynamically dimensioned search algorithm (Tol-
MVK+ is also comparable with APHRODITE. Compared son and Shoemaker, 2007), and the results are given in Ta-
withTRMM3B42RT,TRMM3B42hasgreaterprobabilityof ble3.Thesimulateddaily,monthly,andinter-annualresults
detection and comparable critical success index. This infor- areshowninFig.9b,d,andf.
mation implies that the retrieval algorithm of TRMM3B42 NotethattheparametersofTOPMODELandWEB-DHM
provides positive effects on probability of detection, but no were calibrated using observed precipitation data, and the
obviouspositiveimpactsoncriticalsuccessindex. accuracyofsimulateddischargeswasvalidatedusinggauge
Decreasing trends are observed for all products in terms observations.Comparisonwiththerainfall–runoffmodelpa-
ofprobabilityofdetectionandcriticalsuccessindex,match- rametervaluesreportedforthecasestudycatchmentinprevi-
ing the results of Chen et al. (2013a) for TRMM3B42 and ousresearchshowsthattheparametervaluesareappropriate
TRMM3B42RT. This indicates that probability of detection (Qietal.,2013,2015,2016).
andcriticalsuccessindexhaverelativelystrongdependence
www.hydrol-earth-syst-sci.net/20/903/2016/ Hydrol.EarthSyst.Sci.,20,903–920,2016
912 W.Qietal.:Evaluationofglobalfine-resolutionprecipitationproductsandtheiruncertaintyquantification
Table2.WEB-DHMparameters.
Symbol(unit) Briefdescription Basin-averaged
value
Ks(mmh−1) Saturatedhydraulicconductivityforsoilsurface 26.43
Anik Hydraulicconductivityanisotropyratio 11.49
Sstmax(mm) Maximumsurfacewaterstorage 42.75
Kg(mmh−1) Hydraulicconductivityforgroundwater 0.36
α vanGenuchtenparameter 0.01
n vanGenuchtenparameter 1.88
Figure9.ObservedandsimulatedflowsusingWEB-DHMandTOPMODELfrom2000to2007:(a),(c),and(e)aredaily,monthly,and
inter-annualsimulationsusingWEB-DHMrespectively;(b),(d),and(f)aredaily,monthly,andinter-annualsimulationsusingTOPMODEL
respectively.
4.2 Daily-scaledischarges These time periods were not considered for accuracy com-
parison.
In the case of WEB-DHM simulations, the best
Figures 10 and 11 display scatter plots of discharges dur- NSE(0.41)correspondswithAPHRODITE(Fig.10d),while
ing the period 2000–2007 simulated using WEB-DHM and thebestvalueforRB(1%)correspondswithGLDAS/Noah.
TOPMODEL against gauge observations at a daily scale. InthecaseofTOPMODELsimulations,thebestNSE(0.41)
Twocriteria,NSEandRB,areshown.Itshouldbeennoted corresponds with APHRODITE, and the best value for
that the start dates are different for precipitation products, RB(−24%)correspondswithAPHRODITEalso.Although
and observed data were used when product data are not the best NSE is the same for both WEB-DHM and TOP-
available: from 1 January 2000 to 29 February 2000 for MODEL simulations and the corresponding product is also
TRMM3B42RT, GSMAP-MVK+, and PERSIANN; from thesame,thereisalargedifferenceinthebestRBvalues.At
1 January 2000 to 23 February 2000 for GLDAS/Noah. thedaily-scaleprecipitationamountevaluation,theleastbest
Hydrol.EarthSyst.Sci.,20,903–920,2016 www.hydrol-earth-syst-sci.net/20/903/2016/