assessingtheapplicabilityofrandomforest,stochasticgradient...

Research ArticleAssessing the Applicability of RandomForest Stochastic GradientBoosted Model and Extreme Learning Machine Methods to theQuantitative Precipitation Estimation of the Radar Data A CaseStudy to Gwangdeoksan Radar South Korea in 2018

Ju-Young Shin Yonghun Ro Joo-Wan Cha Kyu-Rang Kim and Jong-Chul Ha

Applied Meteorology Research Division National Institute of Meteorological Sciences 33 Seohobuk-ro Seogwipo-siJeju-do 63568 Republic of Korea

Correspondence should be addressed to Yonghun Ro royh1koreakr

Received 12 March 2019 Revised 23 July 2019 Accepted 23 August 2019 Published 7 October 2019

Guest Editor Jongho Kim

Copyright copy 2019 Ju-Young Shin et alis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Machine learning algorithms should be tested for use in quantitative precipitation estimation models of rain radar data in SouthKorea because such an application can provide a more accurate estimate of rainfall than the conventional ZR relationship-basedmodel e applicability of random forest stochastic gradient boosted model and extreme learning machine methods toquantitative precipitation estimation models was investigated using case studies with polarization radar data fromGwangdeoksanradar station Various combinations of input variable sets were tested and results showed that machine learning algorithms can beapplied to build the quantitative precipitation estimation model of the polarization radar data in South Korea e machinelearning-based quantitative precipitation estimation models led to better performances than ZR relationship-based modelsparticularly for heavy rainfall events e extreme learning machine is considered the best of the algorithms used based onevaluation criteria

1 Introduction

Quantitative precipitation estimation (QPE) using remotesensing data has been widely used to investigate the spatialcharacteristics of precipitation events [1 2]is method canbe used to obtain rainfall estimation at ungauged locationscloud characteristics and areal rainfall depth [3ndash6] espatial resolution of rain radar data is the finest of all theseWhile the spatial resolution of satellite images is greater thanapproximately 10 km the spatial resolution of rain radardata is approximately 1 km [7ndash9] Because of the spatialresolution of rain radar data it is often applied into rainfall-runoff modeling particularly in terms of flash flood andurban flood modeling [10 11] e accurate forecast of theseextreme hydrological events can mitigate damages on thehydraulic infrastructure and prevent the crisis of water-related disaster on human life e accurate QPE of radar

data is the key for the accurate forecast of extreme hydro-logical events

Reflectivity and rainfall rate (ZR) relationship-basedmodels have been used broadly for QPEmodels of rain radardata [12ndash14] Because ZR relationship can be changed basedon the characteristics of the rainfall event and the radarinstrument used various methodologies are applied to buildZR relationship-based QPE models and correct their esti-mations [15ndash18] However the ZR relationship-based modelstill has high uncertainty in a rainfall estimation [19ndash21]

Machine learning (ML) algorithms have been widelyemployed to create functional relationships for naturalphenomena and data processing Many ML algorithms weredeveloped and employed to model a function in fields suchas meteorology hydrology and agriculture Applications ofML algorithms can provide accurate models of naturalphenomena [22ndash25] and thus can be good candidates for

HindawiAdvances in MeteorologyVolume 2019 Article ID 6542410 17 pageshttpsdoiorg10115520196542410

QPE of rain radar data Recently random forest (RF)stochastic gradient boosted model (GBM) and extremelearning machine (ELM) have been actively employed as MLalgorithms [26ndash28] ese advanced ML algorithms whichhave been tested recently would increase our capacity tobuild QPE model Chiang et al [29] proposed a QPE modelusing a recurrent neural network and three-dimensionalradar data ey reported that the ML-based model pro-duced more accurate estimations than the ZR relation-basedmodel Yu et al [30] attempted to develop quantitativeprecipitation forecast (QPF) models of rain radar data usingRF and support vector regression eir proposed meth-odology focused on QPFmodels for typhoons in Taiwan andperformed well

To the best of our knowledge advanced ML algorithmseg RF GBM and ELM have not been employed for QPE ofrain radar data in South Korea is should be resolvedbecause applying ML algorithms may provide more accuraterainfall estimation of rain radar data than the conventionalZR relation-based model erefore this study investigatedthe applicability of the ML algorithms for QPE usingGwangdeoksan radar station South Korea as a case study inorder to enhance performance of QPE in radar data RFGBM and ELM are the ML algorithms used their appli-cability is investigated using four rainfall events and theirperformances for the QPE model are compared is studycan provide fundamental information on the developmentof QPE model using ML algorithms in South Korea Par-ticularly the characteristics of ML algorithm for QPE modelof radar data can be briefly investigated in the study isresult can enhance our capacity to understanding ML al-gorithms in the QPE of radar data In addition the mostplausible candidate among the employedQPEmodels will beselected for ML-based QPE model of radar data in SouthKoreae selected QPEmodel can lead to improvements inaccuracy of QPE particularly in extreme rainfall events thatcause extreme hydrological events e improvement inaccuracy of QPE may help to mitigate impacts from extremehydrological events on the destruction of property andhuman life

is paper is organized as follows In Section 2 thecharacteristics of the radar and ground gauge rainfall dataare presented Section 3 presents a description of themethods employed eg ZR relationship andML algorithmse application methodology for the case studies is pre-sented in Section 4 In Section 5 the results of tested QPEmodels for all events and each event are presented Finallythe conclusions are presented in Section 6

2 Data

21 Radar and Ground Rainfall Gauge Gwangdeoksanweather radar station which has a dual-polarization weatherradar with an S-band is located on the border of Gyeonggi-do and Gangwon-do provinces close to Seoul (latitude38deg7prime25Prime longitude 127deg26prime1Prime and elevation 1064m) thecapital of the Republic of Koreae observation range of theGwangdeoksan radar is 240 km which is enough to coverthe northern part of South Korea Radar data within the

effective observation range 100 km are applied to QPEConsidering the high elevation of this radar station therelationship between the radar and ground rainfall gaugedata is increased with the application of PPI0 PPI (plan-position indicator) is an intensity-modulated display onwhich echo signals are shown in plan view with range andazimuth angle displayed in polar coordinates PPI0 is vol-ume scanned data when the azimuth angle is 0 whichrepresents a condition that can be observed by minimizingblocking in a flat state Data with a spatial resolution of1 kmtimes 1 km and stored at 10-minute interval are applied toestimate radar rainfall

e three main polarization parameters of the radar iereflectivity differential reflectivity and specific differentialphases are applied to QPE in this study Radar reflectivityrefers to the ratio between the transmitted and receivedenergies e differential reflectivity is the ratio betweenhorizontal and vertical radar reflectivity it can provideinformation on the sizes and shapes of raindrops Specificdifferential phases are the rate of change of the range in pulsephases because these are not affected by attenuation partialbeam blockage or radar miscalibration they are an at-tractive parameter to use in QPE [31]

Rainfall rate data from ground gauge stations in Seouland Gyeonggi-do province within the radar umbrella areanalyzed in this study All stations obtain rainfall data everyminute but the QPE in this study uses rainfall rate data at10-min intervals is is to compare the radar data and tominimize the fluctuation of ground gauge data Figure 1shows the ground gauges densely distributed across theKorean peninsula of these 20 gauges within the radar ef-fective range are selected e number of 20 stations islocated in near Seoul and had severe storm damage in thepast e location of Gwangdeoksan radar and the selectedrainfall gauges are described in the zoomed area in Figure 1and information on each station is given in Table 1e useddata can be downloaded from the data base of KoreaMinistration Administration (KMA) at datakmagokr

22 Rainfall Events Rainfall events for which the depth ofthe observed daily rainfall exceeded 30mm from August toNovember 2018 are used Four events are selected as casestudies Two events (the first and second) occurred fromAugust 28ndash29 (event 1) and on September 3 (event 2)ethird event occurred from October 5ndash6 (event 3) and thefourth event happened on November 8 (event 4) Heavyprecipitation was observed in Korea brought on by a rainyfront (the Changma front) for event 1 and by low pressurein the northern area for event 2 Event 3 occurred as partof Typhoon KONG-REY and event 4 was accompanied bythe collision of a cold and a warm front

Table 2 summarizes information on the total rainfallmean rainfall rate and standard deviation for each event atthe station in Seoul (108) e amount of rain that fellduring event 1 was the largest (larger than 100mm) in twodays Both events 1 and 4 show that the weather frontgreatly affected the increase in rainfall or cloud formatione average rainfall is very large in the rainy season and

2 Advances in Meteorology

typhoons such as in events 1 and 3 e largest variancewas observed in event 1is heavy rainfall during the rainyseason is representative of the summer monsoon climate inSouth Korea

3 Methods

31 ZR Relationship Radar rainfall can be dened by therelationship between radar parameters and rainfall gaugedata A variety of synthetic algorithms have been proposedto estimate quantitative radar rainfall based on the polari-zation parameters applied [32 33] e basic form of theequation which is well-known as ZR relationship is given asfollows

R θ0xθ11 x

θdd (d 1 2 n) (1)

where R is the ground gauge rainfall rate (mmh)x1 xd are radar polarization parameters such as

reectivity dierential reectivity and specic dierentialphase and θ0 θd are the parameters of the ZR re-lationship e main radar polarization parameters aredened as the following equations

Z 10 log ZH( )

DR 10 logZH

ZV( )

KD emptyDP r2( ) minus emptyDP r1( )

r2 minus r1

∣∣∣∣∣∣∣∣

∣∣∣∣∣∣∣∣

(2)

where Z is the radar reectivity (changed from mm6mminus 3 todBZ) ZH and ZV are horizontal and vertical reectivity DRis dierential reectivity (dB) KD is specic dierentialphase (deg kmminus 1) and emptyDP and r are phases of the radarbeam pulse and given range respectively Because the ZRrelationship stands on the physical phenomena the results

N

S

E

400 kilometers0 200

W

Height (m)1ndash164165ndash327328ndash490491ndash653

654ndash816817ndash979980ndash11421143ndash1305

1306ndash1468No data

Figure 1 Locations of the Korea Meteorological Administration ground gauge stations (red dots) and Gwangdeoksan (GDK) radar (bluetriangle) with its radar umbrella (eective radius 100 kmmaximum observation radius 240 km)e gure in the right-hand panel presentsthe locations of the ground gauge stations (red dots) used in the current study

Advances in Meteorology 3

of ZR relationship can be used to interpret characteristics ofprecipitation events unlike the ML algorithms e MLalgorithms used in this study are the predictive modelsough they can be used to predict rain rate extractingphysical meaning from the results is difficult For examplethe parameters of ZR relationship can be used to identifytype of cloud type of precipitation and type of storm eventsIn the case of the ML algorithms prediction models for eachvariable of interest such as type of cloud type of pre-cipitation and type of storm events have to be individuallybuilt

32 Machine Learning (ML) Algorithms

321 Random Forest RF has been widely applied in re-gression and forecasting problems [34ndash37] It was proposedby Breiman [38] and uses bagging (called bootstrapping instatistics) to build a number of decision trees with a con-trolled variance Each decision tree in the RF is grown usingrandomly selected samples Subsequently the nodes in eachtree use randomly selected features (called input variables)e RF has two major steps (1) randomness and (2) en-semble learning e randomness in the RF comes fromrandom sampling of the entire data set and the selection offeatures with which every classification and regression tree(CART) is built e data set is randomly sampled withreplacement to create a subset with which to train oneCART At each node optimal split rule is determined by

using the one of the randomly selected features from theemployed features

e ensemble learning method in the RF means that allindividual decision trees in a collection of decision trees(called an ensemble) contribute to a final prediction Atraining subset is created after the random selection step eCART without pruning is used to construct a single decisiontree To growK trees in the ensemble this process (resamplinga subset and training an individual tree) is repeated K timese final predicted value comes from averaging the results ofall the individual trees e ranger library in r package is usedto construct the RF model in the current study [39]

322 Stochastic Gradient Boosted Model GBM is a methodwidely used in classification and regression problems it wasproposed by Friedman [40] Decision stumps or regressiontrees are used widely as weak classifiers in the GBM [40ndash42]In the GBM weak learners are trained to decrease lossfunctions eg mean square errors Residuals in the formerweak learners are used to train the current weak learnerserefore the value of the loss function in the current weaklearners decreases e bagging method is employed toreduce correlation between weak learners and each weaklearner is trained with subsets sampled without replacementfrom the entire data set e final prediction is obtained bycombining predictions by a set of weak learners

e GBM and RF adapted ensemble learning with adecision tree model (the weak learner) Bothmodels produce

Table 1 Ground precipitation gauge stations selected for this study

Name Code Latitude LongitudePaju 99 37885 126766Seoul 108 37571 126965Incheon 112 37477 126624Suwon 119 37272 126985Ganghwa 201 37707 126446Yangpyeong 202 37488 127494Gwanak 116 37445 126964Gangnam 400 37513 127046Gangseo 404 37573 126829Gangbuk 424 37639 127025Uijeongbu 532 37734 127073Namyangju 541 37634 127150Daeseongri 542 37684 127380Gwangju 546 37435 127259Yongin 549 37270 127221Osan 550 37187 127048Guri 569 37582 127156Hwaseong 571 37195 126820Yangju 598 37831 126990Bupyeong 649 37472 126750

Table 2 Precipitation events selected based on observed rainfall data from Seoul station

No Periods of precipitation events Total (mm) Mean (mmh) Standard deviation (mmh)1 20180828 11 50ndash20180829 22 20 1385 40 1192 20180903 08 50ndash20180903 21 30 345 27 583 20181005 08 20ndash20181006 12 20 920 33 334 20181108 01 30ndash20181108 23 40 640 28 35


one prediction based on a combination of predictions froma set of weak learners ough the methods seem to besimilar they are based on different concepts e majordifference between the GBM and RF is that the tree in theGBM is fit on the residual of a subset of the former treeswhile the RF trains a set of weak learners using a number ofsubsets erefore the GBM can reduce bias of predictionwhile the RF method can reduce variance of predictionerefore the RF can be trained in parallel computingwhereas the GBM cannot e gbm library in r package(httpsgithubcomgbm-developersgbm) is used to con-struct a GBM in the current study

323 Regularized Extreme Learning Machine ELM wasoriginally developed and then extended to generalizedsingle-hidden layer feed-forward networks in which thehidden layer need not to be neuron alike ELM is a single-layer network in which the weights and biases between inputand hidden layers are randomly generated [43] Unliketraditional iterative learning algorithms the randomly ini-tiated input weights and biases of ELM remain fixed withoutneed to iteratively tuned and the output weights are de-termined analytically Hence the model can be trained in asingle iteration which significantly reduces the training timeof ELM and makes ELM efficient for online and real-timeapplicationse ELM can be formulized using the followingequations

Y Hβ (3)

where Y β and H are the outputs weight matrix betweenhidden and output layers and the output vector of thehidden layer (called nonlinear feature mapping) re-spectively and

H fa(XW + B) (4)

where fa(middot)W B and x are the activation function weightmatrix between the input to hidden layer bias and inputsrespectively In the current study the sigmoid function(fa(x) 1(1 + exp(minus x))) is used as the activation func-tion in the ELM Since the weights (W) and bias (B) arerandomly generated and the activation function (fa(middot)) isknown in the ELMH represents the deterministic variablesfrom a data set us only β needs to be estimated in theELM

In the ELM finding an appropriate weight set is toavoid overfitting Tuning weights in the ELM can beconsidered a fitting linear regression model using theordinary least square method Ridge regression wasemployed to attenuate multicollinearity in the data set byadding the norm of the parameters to the parameter es-timations in the regression model [44] e ELM modelalso adapted this strategy for weight tuning e ELMattempts to perform better generalization by achieving thesmallest training error and the smallest output weightnorm is minimization problem can take the form ofridge regression or regularized least squares as follows [45]

min12β

2+

C

2Hβ minus Y

2 (5)

where the first term of the objective function is l2 the normregularization term that controls the complexity of themodel the second term is the training error associated withthe learnedmodel and Cgt 0 is a tuning parametere ELMgradient equation can be solved analytically and the closed-form solution can be written as follows

1113954β HTH +1CI1113874 1113875

minus 1HTY (6)

where I is an identity matrix e ELM models used in thisstudy are the regularized ELM model

4 Application Methodology

To examine the applicability of the three ML algorithmstheir input variables and hyperparameters should be definedZ DR and KD in the polarization radar data have beenwidely employed as input variables for the QPE modelerefore these variables are used as input variable can-didates in the ML algorithms e tested models with inputvariable combinations are presented in Table 3

ree ML algorithms use variables from both lag-zero(L0) and lag-one (L1) radar data for input variable while lag-zero and lag-one radar data respectively are used to con-struct ZR relationships Since radar data measure theamount of cloud in the air there is a short time differencebetween radar data and ground gauge observation e timedifference depends on the precipitation event conditionssuch as wind speed cloud movements and types of cloudAs the ZR relationship cannot account for time lag in itsformula QPE models based on the ZR relationship areconstructed using different time-lag data and their appro-priateness are investigated

e ML algorithms can use both lag-zero and lag-oneradar data simultaneously In addition the number ofvariables from the radar data (three) is much smaller thanthe number of data points (greater than thousands) A largernumber of input variables might improve the predictabilityof the ML algorithms employed Additionally since thisinput variable setting can take the time lag in modeling intoaccount automatically additional processes such as the ZRrelationship are unnecessary in ML-based models

To evaluate the performances of the models constructedthe data set should be grouped into training and test datasets e data from stations 112 201 400 546 and 571are used as randomly selected test data e data at the otherstations are used for the training data set For the case of allevents the numbers of training and test data are 3652 and1209 respectively e numbers of training data for event 1to 4 are 1079 319 1173 and 1081 respectivelye numbersof test data for event 1 to 4 are 318 107 441 and 343 re-spectively To build a regressionmodel usingML algorithmstheir hyperparameters should be tuned e number of thetrees is the most sensitive hyperparameter for the RF andGBM [30] hence the number of trees for the RF and GBMare optimized e tuning parameter and the number of


hidden nodes are the hyperparameters of the ELM erelationship between the tuning parameter and the numberof hidden nodes presents a trade-off relationship such asthe Pareto frontier us after one parameter is fixedanother will be optimized In this study the tuning pa-rameter is fixed (C 05) and the number of hidden nodesis optimized

In the current study leave-one-out cross-validation(LOOCV) is employed to optimize the hyperparameters ofthe three ML algorithms e root-mean-square error(RMSE) between the estimates and observations is cal-culated for the ML algorithms trained by the data set thatdoes not include any station among all those in thetraining set e expected numbers of train and test dataare 3227 (approximate 94) and 233 (approximate 6)respectively e fifteen models were trained and theirperformances were evaluated using the test data set eRMSEs without each station are calculated and averagevalue of these RMSEs is the criterion for measuring ap-propriateness of the hyperparameters e results of theLOOCV are presented in Figure 2 e optimal numbersof the tree for the RF and GBM are 380 and 4200 re-spectively any numbers greater than these do not lead tosignificant improvements in increasing the performanceof the RF and GBM e optimal number of hidden nodesfor the ELM is 950 ese numbers are used for thehyperparameters of ML algorithms

QPEmodels are built for five case studiese first casestudy uses all data including the four precipitation eventse other case studies built QPE models for each of theprecipitation events e first case study was carried out toevaluate the overall performances of the QPE modelsconstructed e results of the other case studies mayprovide detailed examinations of the performance of thedifferent rainfall events e RMSE Pearson correlationmean absolute error (MAE) mean bias (Mbias) andrelative root-mean-square error (RRMSE) are employedas evaluation criteria Equation (7) gives the equation ofthe RMSE

RMSE

1n

1113944

n

i1Ei minus Oi( 1113857

2

11139741113972

(7)

where Ei Oi and n are ith radar estimation ith observedprecipitation data point and the number of data pointsrespectively e correlation can be calculated using thefollowing equation

correlation

1113936

ni1 Ei minus E( 1113857 Oi minus O( 1113857

1113936ni1 Ei minus E( 11138571113936

ni1 Oi minus O( 1113857

1113971

(8)

where E and O are the means of the radar estimates andobserved precipitation data respectively MAE Mbias andRRMSE equations are given in equations (9)ndash(11)respectively

MAE 1n

1113944

n

i1Ei minus Oi

11138681113868111386811138681113868111386811138681113868 (9)

MBias 1n

1113944

n

i1Ei minus Oi( 1113857 (10)

RRMSE RMSE

otimes 100 (11)

5 Results

51 Overall Performance of the QPE Models e overallperformances of the constructed QPE models are evaluatedusing rainfall and radar data for all the events e trainingand test data sets are constructed from the data set thatincluded all the rainfall events Evaluation criteria for theconstructed QPE models are applied to the test data setResults of evaluation criteria are presented in Figure 3 eML-based models lead to lower RMSEs than ZR relation-ship-based models

When the number of input variables increases theRMSE becomes smaller For the ZR relationship-basedmodels RMSEs of ZR-L1-based models are smaller thanthose of ZR-L0-based modelse result means that usage oflag-data may provide more information onto QPE of theemployed radar data Models that include all the availableinput variables lead to lowest RMSE values e secondlowest RMSE is observed for models that use Z and DR asinput variables Models using DR and KD lead to the largestRMSE Based on RMSE the ELM5 (using Z DR and KD) isthe best model for QPE of the radar data Correlation resultsare similar to the results of the RMSEeML-based modelsgive larger correlations than ZR relationship-based modelsFor correlation the cases using all input variables providethe largest correlation values Based on MAE models usingZ and DR as input variables lead to the smallest MAE ebest model based on MAE is the ELM2 (using Z and DR)Based on MBias ZR-L0-based models are the best modelswith an MBias close to zero Estimations by ZR-L1-basedmodels have positive biases except for the ZR4-L1 whilethose of ML-based models have negative biases RRMSEs ofall employed QPEmodels are larger than 100 Based on theRRMSEs the ELM is the best model for QPE of radar datae second best model is the RF

Estimation-verse observation plots are presented inFigure 4 Rainfall rate estimations are underestimated forlarge amounts of rainfall rates (larger than 40mmh)eseunderestimations for large amount of rainfall rate areclearly observed in the results of ZR-L0-based models

Table 3 Tested models for quantitative precipitation estimationfrom radar data

ModelInput variables

Z Z DR Z KD DR KD Z DR KDZR-L1 ZR1-L1 ZR2-L1 ZR3-L1 ZR4-L1 ZR5-L1ZR-L0 ZR1-L0 ZR2-L0 ZR3-L0 ZR4-L0 ZR5-L0RF RF1 RF2 RF3 RF4 RF5GBM GBM1 GBM2 GBM3 GBM4 GBM5ELM ELM1 ELM2 ELM3 ELM4 ELM5


Number of trees

RMSE

(mm

hr)

0 100 300 400200 500970

975

980

985

990

(a)

Number of trees

RMSE

(mm

hr)

0 2000 6000 80004000

100

105

110

(b)

0 200 800600400 10001005

1010

1015

1020

1025

1030

1035

1040

Number of hidden modes

RMSE

(mm

hr)

(c)

Figure 2 Leave-one-out cross-validation results of hyperparameters for the random forest (the number of trees) stochastic gradientboosted model (the number of trees) and extreme learning machine (the number of hidden nodes) models e red circles indicate theselected optimal points of the employed hyperparameters based on the root-mean-square error LOOCV results of (a) RF (b) GBM and(c) ELM

RMSE

(mm

hr)

7

8

9

10

11

Z Z DR Z KD DR KD Z DR KD

ZR-L1ZR-L0

RFGBM

ELM

Input variable sets

(a)

Figure 3 Continued


Mbias of ZR-L0-based models is close to zero To meet thevalue ofMbias estimate rainfall rate estimations for small andmedium amounts of rainfall rates are overestimated Mbias ofZR-L1-based models have positive values and estimations areunderestimated for large amount of precipitation which

indicates that a large overestimation occurs for a smallamount of precipitation ese overestimations also are ob-served in Figure 4 e ELM5 leads to the best estimationperformance in Figure 4 Circles by the ELM5 are locatedcloser to the diagonal line than other models while ZR5-L1

COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM

Z Z DR Z KD DR KD Z DR KDInput variable sets

(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)

Figure 3 (a) Root-mean-square error (RMSE) (b) correlation (COR) (c) mean absolute error (MAE) (d) mean bias (Mbias) and(e) relative RMSE (RRMSE) of rainfall rate estimations by the tested quantitative precipitation estimation models for all rainfall eventsstudied


and GBM5 models seem to provide poor performances ecircle distribution for these models is L-shaped (orthogonalshape) in Figure 4

52 Performances of the Constructed QPE Models for Single-Rainfall Events Parameters of ZR relationship differdepending on rainfall events characteristics e perfor-mances of the QPE model differ from the rainfall events

Hence to obtain an accurate QPE the QPE model should bebuilt for every rainfall event To investigate the applicabilityand performance of QPE models all the tested QPE modelsare built using data from each precipitation event RMSEs ofthe tested QPE models for single-rainfall events are pre-sented in Table 4 ML-based models are selected for the bestmodels based on RMSEs For events 1 and 2 the ELM5and ELM3 lead to the lowest RMSEs respectively Based onRMSEs RF2 and RF5 lead to the best performance for events

Estim

atio

ns (m

mh

r)

0 60 140Observations (mmhr)

100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)

Figure 4 Plots of rainfall rate estimation versus observations for the models tested for all precipitation events studied (a) ZR1-L1 (RMSE875 R 058) (b) ZR2-L1 (RMSE 887 R 056) (c) ZR3-L1 (RMSE 874 R 058) (d) ZR5-L1 (RMSE 886 R 058) (e) ZR1-L0 (RMSE933 R 047) (f ) ZR2-L0 (RMSE 936 R 046) (g) ZR3-L0 (RMSE 935 R 047) (h) ZR5-L0 (RMSE 937 R 046) (i) RF1 (RMSE 856 R06) (j) RF2 (RMSE 826 R 063) (k) RF3 (RMSE 836 R 062) (l) RF5 (RMSE 818 R 063) (m) GBM1 (RMSE 838 R 061) (n) GBM2(RMSE 843 R 061) (o) GBM3 (RMSE 838 R 024) (p) GBM5 (RMSE 843 R 06) (q) ELM1 (RMSE 837 R 064) (r) ELM2 (RMSE799 R 066) (s) ELM3 (RMSE 833 R 064) (t) ELM5 (RMSE 791 R 067)


3 and 4 respectively Overall RMSEs of the models usingZ and KD data are lower than those in the models usingother input variable sets for event 2 For event 3 RMSEs ofthe models using Z and DR data are lower than the modelsusing other input variable sets Differences between RMSEsof ML and ZR relationship-based models are very small inevent 4 Although RF5 is selected as the best model in theevent 4 based on RMSEs the difference between RMSEs ofRF5 and ZR5-L1 is 001 Practically the performances of ZR-L1-based models are the best for event 4 based on RMSEs

Table 5 presents the correlations of the tested QPEmodels for single-rainfall eventse correlations of the QPEmodels for events 1 and 2 are much larger than those forevents 3 and 4 While the RMSEs of events 1 and 2 arelarger than events 3 and 4 their correlations are higherthan those of events 3 and 4 Results indicate that the QPEmodels lead to good estimation performance for heavyrainfall events e largest correlation values are observed inML-based models ELM2 and ELM3 lead to the largestcorrelations for events 1 and 2 respectively RF2 and RF5provide the largest correlations for events 3 and 4 re-spectively Results of MAE are similar to the results ofRMSE Based on Mbias ZR-L1- ZR-L0- GBM- and ELM-based models lead to the best performance for events 1 to4 respectively Detailed MAE and MBias results are notcontained in the current manuscript

Table 6 presents the RRMSEs of the tested QPE modelsfor single-rainfall eventse correlations of the QPEmodelsfor events 3 and 4 are smaller than those for events 1 and2 unlike the results of RMSE and correlation e smallest

RRMSEs for events 1 and 2 (heavy rainfall events) are718 and 681 respectively For events 3 and 4 (lightrainfall events) the smallest RRMSEs are 606 and 630

Table 4 Root-mean-square errors (RMSEs) of rainfall rates esti-mated by quantitative precipitation estimation models for selectedrainfall events

Event no ModelInput variables

Z Z DR Z KD DR KD Z DRKD

1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289

Italicized numbers indicate the smallest RMSEs among those calculatedduring the same rainfall events

Table 5 Correlations of rain rate estimations by quantitativeprecipitation estimation models for each selected rainfall event



1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415

Italicized numbers indicate the largest correlations among those calculatedduring the same rainfall events

Table 6 Relative root-mean-square errors (RRMSEs) of rain rateestimations by quantitative precipitation estimation models foreach selected rainfall event



1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634

Italicized numbers indicate the smallest RRMSEs among those calculatedduring the same rainfall events


respectively Overall difference between the smallest RRMSEsof heavy and light rainfall events are approximately 10 edifference between the smallest RMSE of heavy and lightrainfall events is approximately 43mmhr Because 43mmhr is larger than the smallest RMSE of event 3 the RRMSEdifference is relatively smaller than RMSE difference eresult indicates that the QPE models provide similar per-formances for heavy and light rainfall events based on RRMSEmeasures

To evaluate the tested QPE models for single-rainfallevents rainfall rate estimation versus observation plots forevent 1 and 4 is presented in Figures 5 and 6 respectivelye tested models excluding ZR-L0-based models lead togood estimation performances for event 1 Of the testedmodels ELM5 gives the best estimation performance Somecircles are aligned to approximately 80mmh based onobservations in Figure 5 is is a recurrent issue in QPE ofrain radar data When the observed rainfall rates are thesame but the observed parameters of radar data are differentthis phenomenon occurs is result indicates that all testedQPEmodels cannot solve this issue For event 4 in Figure 6all the tested QPE models lead to poor performances ofrainfall rate estimation In the observed small magnitude ofrainfall rates the QPE models tend to overestimate rainfallrates On the other hand the QPE models provide an un-derestimation for the observed large magnitude of rainfallrates Five lines are observed at approximately 3mmh6mmh 9mmh 12mmh and 15mmh based on obser-vations in all the subfigures presented in Figure 6 eobserved rainfall depths for these small rainfall rates are05mm 1mm 15mm 2mm and 25mm and their du-ration is 10 minutes Because event 4 is light a largenumber of small rainfall rates are observed e phenomenawherein parameters of radar are different for the sameamount of observed rainfall rate occur frequently Due tothis phenomenon the tested QPE models show poor per-formances in event 4

Radar rainfall rate fields for events 1 and 4 are il-lustrated to investigate the difference between ZR relation-and ML-based models in Figures 7 and 8 Figure 7 presentsradar rainfall rate fields of event 1 at 20 10 28th August2018 e range of rainfall rates is from 0 to 100mmh forevent 1 e ML-based QPE models provide larger mag-nitudes of rainfall rates for very small magnitudes of rainfallrates based on estimates by ZR2-L1 For heavy magnitudesof rainfall rates estimates of all QPE models are similar eGBM leads to the largest magnitude of rainfall rate esti-mation in Figure 7 Rainfall rate estimates on ground gaugestations for the ZR2-L1 are larger than those of ML-basedmodels Due to the high magnitude of rainfall rate at thesepoints the ZR2-L1 overestimates rainfall rate in Figure 3 Inareas where there are no ground gauge stations the ZR2-L1estimates smaller rainfall rates than other models Figure 8presents radar rainfall rate fields of event 4 at 12 40 8thNovember 2018 Rainfall rates range from 0 to 15mmh forevent 4 Overall results of rainfall rate estimates by thetested models are similar to the results shown in Figure 7e ELM leads to the largest magnitude of rainfall rateestimation

6 Discussion

e comparison results of the ZR relationship- and ML-based models show that the application of ML algorithmscan lead to an improvement in the QPE of radar data in thetested rainfall eventsis result supports the notion that theML algorithm could be used in the development of QPEmodels of radar data in South Korea Increasing the numberof variables for the input variables of the ZR relationship-based models results in very small improvements In someevents this increment does not improve performances ofQPE models It can be inferred that Z is the most criticalvariable for the ZR relationship-based model Additionallythe application of other variables is often an inefficient wayto build the ZR relationship-based model

e performances of the ML-based models improvewhen Z and additional variables such as DR and KD areapplied as input variables In particular a combination of Zand DR for input variables of theML-basedmodels leads to agood QPE performance Studies have reported that thiscombination leads to the best performance among combi-nations of Z DR and KD for the ZR relationship-basedmodel [46 47] In many cases application of DR except forcombinations of DR and KD lead to a large improvement inthe QPE using the ML-based model unlike the ZR re-lationship-based modele results imply that theML-basedmodels could consider other variables in QPE Because theML-based model can extract a large amount of informationfrom the input variables and use this information in QPE ofrain radar data performances of the ML-based models maybe better than those of ZR relationship-based models Basedon results of RMSE for individual events the RF model withthree variables provided the smallest RMSEs in events 2 and4 Otherwise RMSEs of other RF models were smaller thanthose of the RFmodel with three variables In addition thereis a very small difference (006mmhr) between RSMEs ofRF models with three variables and with Z and KD e RFmodel with Z and KD is the best model when taking intoconsideration of parsimony for event 2 Hence the RFmodel with three variables can be considered for suboptimalin events 1 2 and 3

Computation times to build QPEmodels differ dependingon the ML algorithms employed RF has the shortest com-putation time and its computation time with the data sets ofall events is approximately 1 minute e computation timesof the GBM and ELM are approximately 7 minutes and 3minutes respectively As the measuring interval of rainfalldata is 10 minutes the computation time should be shorterthan 5 minutes e RF- and ELM-based models proposed inthe current study can be applied for QPE but the GBM has tobe modified before application

Based on the results of this study a comparison of theperformances of the employed algorithms can be carried oute ELM leads to the best performance for the case thatincludes all the events For single events the best algorithmsare different e ELM provides a good performance forheavy rainfall events while the RF is considered a goodalgorithm for light rainfall events e difference in per-formance between the RF and ELM is small in the light


rainfall events Hence the best ML algorithm for case studiesperformed in the current study is the ELM Each ML al-gorithm tested in this study uses popular setting ecomparison results of the ML algorithm for QPEmodels canbe altered by adopted setting and used data For example inthis study the CART is used for the decision tree in the RFOther decision tree models can be used in the RF such asinference dichotomiser 3 and chi-squared automatic itera-tion detection RF with other decision tree models can

outperform to the ELM model for QPE in South Koreaus the results in the current study should be restricted tothese data sets and the ML algorithms with adopted setting

Variation of neural network models like artificial neuralrecurrent neural and deep neural networks may have a highapplicability building QPE models of radar data in SouthKorea because the ELM is developed based on a neuralnetwork Additionally enhancing the precision of rainfallgauges may lead to improvements in the performance of

Rada

r (m

mh

r)

Ground gauge (mmhr)0 60 140

100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)

Figure 5 Plots of rainfall rate estimations versus observations of tested models for precipitation event 1 (a) ZR1-L1 (RMSE 1197 R 075)(b) ZR2-L1 (RMSE 1191 R 075) (c) ZR3-L1 (RMSE 1198 R 075) (d) ZR5-L1 (RMSE 1192 R 075) (e) ZR1-L0 (RMSE 1459 R 06)(f ) ZR2-L0 (RMSE 1462 R 06) (g) ZR3-L0 (RMSE 1462 R 06) (h) ZR5-L0 (RMSE 1466 R 06) (i) RF1 (RMSE 1129 R 079) (j) RF2(RMSE 113 R 079) (k) RF3 (RMSE 1101 R 08) (l) RF5 (RMSE 112 R 08) (m) GBM1 (RMSE 1105 R 08) (n) GBM2 (RMSE 1105R 08) (o) GBM3 (RMSE 1103 R 042) (p) GBM5 (RMSE 1106 R 08) (q) ELM1 (RMSE 103 R 083) (r) ELM2 (RMSE 1011 R 084)(s) ELM3 (RMSE 1021 R 083) (t) ELM5 (RMSE 1006 R 084)


QPE models for light rainfall events Precision for some ofthe employed rainfall gauges is 3mmhr Although thisprecision is good enough to measure rainfall rates for longduration or heavy rainfall events it should be higher forestimating rainfall rates of light rainfall events For examplewhen parameters of radar data for two points are differentbut their observed rainfall rates are the same the QPEmodelhas to fail estimations of rainfall rates at two points If the

precision of the rainfall gauge increases the observed rainfallrates may be different and could result in a more accurateconstructed QPE model As shown in Figure 6 three linescan be observed in all the subfigures Values of ground gaugefor first second and third lines are 3 6 and 9mmhr esethree lines indicate that the observed rainfall rates at groundgauge station are the same when the parameters of radar dataare different If precisions of these gauge stations become

Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh

r)Ground gauge (mmhr)

0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)

Figure 6 Plots of rainfall rate estimations versus observations of tested models for precipitation event 4 (a) ZR1-L1 (RMSE 288 R 042)(b) ZR2-L1 (RMSE 286 R 043) (c) ZR3-L1 (RMSE 288 R 042) (d) ZR5-L1 (RMSE 286 R 043) (e) ZR1-L0 (RMSE 305 R 027) (f )ZR2-L0 (RMSE 302 R 03) (g) ZR3-L0 (RMSE 305 R 028) (h) ZR5-L0 (RMSE 302 R 03) (i) RF1 (RMSE 293 R 043) (j) RF2(RMSE 297 R 042) (k) RF3 (RMSE 289 R 045) (l) RF5 (RMSE 285 R 046) (m) GBM1 (RMSE 289 R 042) (n) GBM2 (RMSE 286R 042) (o) GBM3 (RMSE 288 R 042) (p) GBM5 (RMSE 286 R 043) (q) ELM1 (RMSE 291 R 039) (r) ELM2 (RMSE 29 R 041)(s) ELM3 (RMSE 29 R 04) (t) ELM5 (RMSE 289 R 041)


better these lines may be disappeared and the data points inthe lines are dissipated is dispersion of data pointscaused by the high precision of measuring instrument maylead to improvement of the performance of QPE models forlight rainfall event

e tested QPE models lead to good performances forheavy rainfall events but not for light rainfall events is

characteristic of QPE with rain radar data is also observed inthis study ML-based QPE models outperform ZR re-lationship-based models for events 3 and 4 albeit in-significantly As mentioned above the ML-based QPEmodels show good performances by efficiently extractinginformation from given radar data If additional variablescan be applied in QPE models the performance of the QPE

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)

Figure 7 Radar rainfall rate fields for four selected quantitative precipitation estimationmodels (ZR2ndashL1 RF3 GBM3 and ELM5) for event1 (August 28 2018 20 10) (a) ZR (b) RF (c) GB (d) EL


model may improve particularly for light rainfall eventsus various sets of input variables that are frequently usedin conventional QPE algorithms should be tested for ML-based QPE models to improve the performance of QPEmodels for light rainfall events

7 Conclusions

e applicability of three ML algorithms in QPE models isinvestigated using case studies of polarization radar data of fourrainfall events fromGwangdeoksan radar station Gyeonggi-do

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)

Figure 8 Radar rainfall rate fields of four selected quantitative precipitation estimation models (ZR2ndashL1 RF5 GBM2 and ELM5) for event4 (November 8 2018 12 40) (a) ZR2-L1 (b) RF5 (c) GBM2 (d) ELM5


South Korea Various combinations of input variable sets arealso tested for QPE models Conventional ZR relation-basedmodels are also constructed and compared to ML-basedmodels In the current study we reach the followingconclusions

(1) ML algorithms can be applied to build a QPE modelof polarization radar data Overall the ML-basedQPE models outperform or are equal to ZR re-lationship-based models ML algorithms can extractinformation from radar data more efficiently thanthe ZR relationship which leads to an improvementin QPE of the radar data

(2) Application of the ML algorithms for QPE modelsimproves rainfall rate estimations for heavy events inSouth Korea by far e performances of the ML-based QPE model are significantly improved basedon performances of ZR relationship-based modelsfor heavy rainfall events is improvement will behelpful in modeling floods and forecasting flashfloods

(3) ELM algorithm may be the best ML algorithmamong the tested ML algorithm with the adoptedsetting for QPE models of radar data in South KoreaOverall the ELM outperforms other tested QPEmodels in QPE of radar data employed in the currentstudy Based on evaluation results of single-rainfallevents the ELM also leads to the best performance intwo heavy rainfall events Although the ELM is notthe best QPE model for the two light rainfall eventsthe performances of QPE models using ELM arecomparable to other QPE models

In the current study four rainfall events in 2018 wereemployed to evaluate the applicability of ML algorithms forthe QPE model of polarization radar data as the radarinstrument in Gwangdeoksan radar was updated Futurerainfall events should be included in data sets to furtherinvestigate the applicability and characteristics of ML al-gorithms in the QPE of polarization radar data in SouthKorea In addition the applicability of ML algorithms forQPF should be examined Because ML algorithms showhigh applicability in QPE they make good candidates formodeling functions between radar data and QPF

Data Availability

e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was funded by the Korea Meteorological Admin-istration Research and Development Program ldquoDevelopmentof Application Technology onAtmospheric ResearchAircraftrdquounder Grant (1365003069)

References

[1] B Martens P Cabus I De Jongh and N E C VerhoestldquoMerging weather radar observations with ground-basedmeasurements of rainfall using an adaptive multiquadricsurface fitting algorithmrdquo Journal of Hydrology vol 500pp 84ndash96 2013

[2] S Yang and S W Nesbitt ldquoStatistical properties of pre-cipitation as observed by the TRMM precipitation radarrdquoGeophysical Research Letters vol 41 no 15 pp 5636ndash56432014

[3] W F Krajewski and J A Smith ldquoRadar hydrology rainfallestimationrdquo Advances in Water Resources vol 25 no 8ndash12pp 1387ndash1394 2002

[4] J Kim J Lee D Kim and B Kang ldquoe role of rainfall spatialvariability in estimating areal reduction factorsrdquo Journal ofHydrology vol 568 pp 416ndash426 2019

[5] W Cho J Lee J Park and D Kim ldquoRadar polygon methodan areal rainfall estimation based on radar rainfall imageriesrdquoStochastic Environmental Research and Risk Assessmentvol 31 no 1 pp 275ndash289 2017

[6] T Park T Lee D Ko J Shin and D Lee ldquoAssessing spatiallydependent errors in radar rainfall estimates for rainfall-runoffsimulationrdquo Stochastic Environmental Research and RiskAssessment vol 31 no 7 pp 1823ndash1838 2016

[7] J Simpson R F Adler and G R North ldquoA proposed tropicalrainfall measuring mission (TRMM) satelliterdquo Bulletin of theAmerican Meteorological Society vol 69 no 3 pp 278ndash2951988

[8] R J Joyce J E Janowiak P A Arkin and P Xie ldquoCMORPHa method that produces global precipitation estimates frompassive microwave and infrared data at high spatial andtemporal resolutionrdquo Journal of Hydrometeorology vol 5no 3 pp 487ndash503 2004

[9] Y Wang and V Chandrasekar ldquoQuantitative precipitationestimation in the CASA X-band dual-polarization radarnetworkrdquo Journal of Atmospheric and Oceanic Technologyvol 27 no 10 pp 1665ndash1676 2010

[10] G Delrieu I Braud A Berne et al ldquoWeather radar andhydrologyrdquo Advances in Water Resources vol 32 no 7pp 969ndash974 2009

[11] T Lee J Shin T Park and D Lee ldquoBasin rotation method foranalyzing the directional influence of moving storms on basinresponserdquo Stochastic Environmental Research and Risk As-sessment vol 29 no 1 pp 251ndash263 2015

[12] N Balakrishnan D S Zrnic J Goldhirsh and J RowlandldquoComparison of simulated rain rates from disdrometer dataemploying polarimetric radar algorithmsrdquo Journal of Atmo-spheric and Oceanic Technology vol 6 no 3 pp 476ndash4861989

[13] A V Ryzhkov and D S Zrnic ldquoComparison of dual-po-larization radar estimators of rainrdquo Journal of Atmosphericand Oceanic Technology vol 12 no 2 pp 249ndash256 1995

[14] S Verrier L Barthes and C Mallet ldquoeoretical and em-pirical scale dependency of Z-R relationships evidenceimpacts and correctionrdquo Journal of Geophysical ResearchAtmospheres vol 118 no 14 pp 7435ndash7449 2013

[15] J A Smith andW F Krajewski ldquoEstimation of the mean fieldbias of radar rainfall estimatesrdquo Journal of Applied Meteo-rology vol 30 no 4 pp 397ndash412 1991

[16] M-K Suk K-H Chang J-W Cha and K-E Kim ldquoOper-ational real-time adjustment of radar rainfall estimation overthe South Korea regionrdquo Journal of the Meteorological Societyof Japan Ser II vol 91 no 4 pp 545ndash554 2013


[17] C Yoo C Park J Yoon and J Kim ldquoInterpretation of mean-field bias correction of radar rain rate using the concept oflinear regressionrdquo Hydrological Processes vol 28 no 19pp 5081ndash5092 2014

[18] L Yang Y Yang P Liu and L Wang ldquoRadar-Derivedquantitative precipitation estimation based on precipitationclassificationrdquo Advances in Meteorology vol 2016 Article ID2457489 16 pages 2016

[19] D-J Seo and J A Smith ldquoRainfall estimation using raingagesand radarmdashA Bayesian approach 1 Derivation of estima-torsrdquo Stochastic Hydrology and Hydraulics vol 5 no 1pp 17ndash29 1991

[20] D-J Seo and J A Smith ldquoRainfall estimation using raingagesand radarmdashA Bayesian approach 2 An applicationrdquo Sto-chastic Hydrology and Hydraulics vol 5 no 1 pp 31ndash441991

[21] Q Dai M A Rico-Ramirez D Han T Islam and S LiguorildquoProbabilistic radar rainfall nowcasts using empirical andtheoretical uncertainty modelsrdquo Hydrological Processesvol 29 no 1 pp 66ndash79 2013

[22] A Mosavi P Ozturk and K-w Chau ldquoFlood predictionusing machine learning models literature reviewrdquo Watervol 10 no 11 p 1536 2018

[23] P Ganguli and M J Reddy ldquoEnsemble prediction of regionaldroughts using climate inputs and SVM-copula approachrdquoHydrological Processes vol 28 no 19 pp 4989ndash5009 2013

[24] D J Gagne II A McGovern and M Xue ldquoMachine learningenhancement of storm-scale ensemble probabilistic quanti-tative precipitation forecastsrdquo Weather and Forecastingvol 29 no 4 pp 1024ndash1043 2014

[25] Y Tao X Gao A Ihler S Sorooshian and K Hsu ldquoPre-cipitation identification with bispectral satellite informationusing deep learning approachesrdquo Journal of Hydrometeorol-ogy vol 18 no 5 pp 1271ndash1283 2017

[26] X Lu Y Ju L Wu J Fan F Zhang and Z Li ldquoDaily panevaporation modeling from local and cross-station data usingthree tree-based machine learning modelsrdquo Journal of Hy-drology vol 566 pp 668ndash684 2018

[27] B T Nolan C T Green P F Juckem L Liao and J E ReddyldquoMetamodeling and mapping of nitrate flux in the un-saturated zone and groundwater Wisconsin USArdquo Journal ofHydrology vol 559 pp 428ndash441 2018

[28] H I Erdal and O Karakurt ldquoAdvancing monthly streamflowprediction accuracy of CARTmodels using ensemble learningparadigmsrdquo Journal of Hydrology vol 477 pp 119ndash128 2013

[29] Y-M Chiang F-J Chang B J-D Jou and P-F Lin ldquoDy-namic ANN for precipitation estimation and forecasting fromradar observationsrdquo Journal of Hydrology vol 334 no 1-2pp 250ndash261 2007

[30] P-S Yu T-C Yang S-Y Chen C-M Kuo andH-W Tseng ldquoComparison of random forests and supportvector machine for real-time radar-derived rainfall fore-castingrdquo Journal of Hydrology vol 552 pp 92ndash104 2017

[31] M R Kumjian ldquoPrinciples and applications of dual-polari-zation weather radar Part I description of the polarimetricradar variablesrdquo Journal of Operational Meteorology vol 1no 19 pp 226ndash242 2013

[32] A V Ryzhkov S E Giangrande and T J Schuur ldquoRainfallestimation with a polarimetric prototype of WSR-88drdquoJournal of Applied Meteorology vol 44 no 4 pp 502ndash5152005

[33] R Cifelli V Chandrasekar S Lim P C Kennedy Y Wangand S A Rutledge ldquoA new dual-polarization radar rainfallalgorithm application in Colorado precipitation eventsrdquo

Journal of Atmospheric and Oceanic Technology vol 28 no 3pp 352ndash364 2011

[34] T Yang A A Asanjan E Welles X Gao S Sorooshian andX Liu ldquoDeveloping reservoir monthly inflow forecasts usingartificial intelligence and climate phenomenon informationrdquoWater Resources Research vol 53 no 4 pp 2786ndash2812 2017

[35] J Fan W Yue L Wu et al ldquoEvaluation of SVM ELM andfour tree-based ensemble models for predicting daily refer-ence evapotranspiration using limited meteorological data indifferent climates of Chinardquo Agricultural and Forest Meteo-rology vol 263 pp 225ndash241 2018

[36] B Pang J Yue G Zhao and Z Xu ldquoStatistical downscaling oftemperature with the random forest modelrdquo Advances inMeteorology vol 2017 Article ID 7265178 11 pages 2017

[37] C Choi J Kim J Kim D Kim Y Bae and H S KimldquoDevelopment of heavy rain damage prediction model usingmachine learning based on big datardquo Advances in Meteo-rology vol 2018 Article ID 5024930 11 pages 2018

[38] L Breiman ldquoRandom forestsrdquo Machine Learning vol 45no 1 pp 5ndash32 2001

[39] M N Wright and A Ziegler ldquoRanger a fast implementationof random forests for high dimensional data in C++ and RrdquoSurvival Analysis vol 77 no 1 p 17 2017

[40] J H Friedman ldquoStochastic gradient boostingrdquo Computa-tional Statistics amp Data Analysis vol 38 no 4 pp 367ndash3782002

[41] A Torres-Barran A Alonso and J R Dorronsoro ldquoRe-gression tree ensembles for wind energy and solar radiationpredictionrdquo Neurocomputing vol 326-327 pp 151ndash160 2019

[42] S A Naghibi H R Pourghasemi and B Dixon ldquoGIS-basedgroundwater potential mapping using boosted regression treeclassification and regression tree and random forest machinelearning models in Iranrdquo Environmental Monitoring andAssessment vol 188 no 1 p 44 2016

[43] G-B Huang Q-Y Zhu and C-K Siew ldquoExtreme learningmachine theory and applicationsrdquo Neurocomputing vol 70no 1ndash3 pp 489ndash501 2006

[44] M Pentildea and H van den Dool ldquoConsolidation of multimodelforecasts by ridge regression application to pacific sea surfacetemperaturerdquo Journal of Climate vol 21 no 24pp 6521ndash6538 2008

[45] G B Huang H Zhou X Ding and R Zhang ldquoExtremelearning machine for regression and multiclass classificationrdquoIEEE Transactions on Systems Man and Cybernetics Part B(Cybernetics) vol 42 no 2 pp 513ndash529 2012

[46] M J Simpson and N I Fox ldquoDual-polarized quantitativeprecipitation estimation as a function of rangerdquo Hydrologyand Earth System Sciences vol 22 no 6 pp 3375ndash3389 2018

[47] B-C Seo B Dolan W F Krajewski S A Rutledge andW Petersen ldquoComparison of single- and dual-polarization-based rainfall estimates using NEXRAD data for the NASAIowa flood studies projectrdquo Journal of Hydrometeorologyvol 16 no 4 pp 1658ndash1675 2015


Hindawiwwwhindawicom Volume 2018

Journal of

ChemistryArchaeaHindawiwwwhindawicom Volume 2018

Marine BiologyJournal of


BiodiversityInternational Journal of


EcologyInternational Journal of


Hindawiwwwhindawicom

Applied ampEnvironmentalSoil Science

Volume 2018

Forestry ResearchInternational Journal of



International Journal of

Geophysics

Environmental and Public Health

Journal of




Microbiology


Public Health Advances in

AgricultureAdvances in


Agronomy




MeteorologyAdvances in

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018Hindawiwwwhindawicom Volume 2018

ChemistryAdvances in

ScienticaHindawiwwwhindawicom Volume 2018


Geological ResearchJournal of

Analytical ChemistryInternational Journal of


Submit your manuscripts atwwwhindawicom

QPE of rain radar data Recently random forest (RF)stochastic gradient boosted model (GBM) and extremelearning machine (ELM) have been actively employed as MLalgorithms [26ndash28] ese advanced ML algorithms whichhave been tested recently would increase our capacity tobuild QPE model Chiang et al [29] proposed a QPE modelusing a recurrent neural network and three-dimensionalradar data ey reported that the ML-based model pro-duced more accurate estimations than the ZR relation-basedmodel Yu et al [30] attempted to develop quantitativeprecipitation forecast (QPF) models of rain radar data usingRF and support vector regression eir proposed meth-odology focused on QPFmodels for typhoons in Taiwan andperformed well

To the best of our knowledge advanced ML algorithmseg RF GBM and ELM have not been employed for QPE ofrain radar data in South Korea is should be resolvedbecause applying ML algorithms may provide more accuraterainfall estimation of rain radar data than the conventionalZR relation-based model erefore this study investigatedthe applicability of the ML algorithms for QPE usingGwangdeoksan radar station South Korea as a case study inorder to enhance performance of QPE in radar data RFGBM and ELM are the ML algorithms used their appli-cability is investigated using four rainfall events and theirperformances for the QPE model are compared is studycan provide fundamental information on the developmentof QPE model using ML algorithms in South Korea Par-ticularly the characteristics of ML algorithm for QPE modelof radar data can be briefly investigated in the study isresult can enhance our capacity to understanding ML al-gorithms in the QPE of radar data In addition the mostplausible candidate among the employedQPEmodels will beselected for ML-based QPE model of radar data in SouthKoreae selected QPEmodel can lead to improvements inaccuracy of QPE particularly in extreme rainfall events thatcause extreme hydrological events e improvement inaccuracy of QPE may help to mitigate impacts from extremehydrological events on the destruction of property andhuman life

is paper is organized as follows In Section 2 thecharacteristics of the radar and ground gauge rainfall dataare presented Section 3 presents a description of themethods employed eg ZR relationship andML algorithmse application methodology for the case studies is pre-sented in Section 4 In Section 5 the results of tested QPEmodels for all events and each event are presented Finallythe conclusions are presented in Section 6

2 Data

21 Radar and Ground Rainfall Gauge Gwangdeoksanweather radar station which has a dual-polarization weatherradar with an S-band is located on the border of Gyeonggi-do and Gangwon-do provinces close to Seoul (latitude38deg7prime25Prime longitude 127deg26prime1Prime and elevation 1064m) thecapital of the Republic of Koreae observation range of theGwangdeoksan radar is 240 km which is enough to coverthe northern part of South Korea Radar data within the

effective observation range 100 km are applied to QPEConsidering the high elevation of this radar station therelationship between the radar and ground rainfall gaugedata is increased with the application of PPI0 PPI (plan-position indicator) is an intensity-modulated display onwhich echo signals are shown in plan view with range andazimuth angle displayed in polar coordinates PPI0 is vol-ume scanned data when the azimuth angle is 0 whichrepresents a condition that can be observed by minimizingblocking in a flat state Data with a spatial resolution of1 kmtimes 1 km and stored at 10-minute interval are applied toestimate radar rainfall

e three main polarization parameters of the radar iereflectivity differential reflectivity and specific differentialphases are applied to QPE in this study Radar reflectivityrefers to the ratio between the transmitted and receivedenergies e differential reflectivity is the ratio betweenhorizontal and vertical radar reflectivity it can provideinformation on the sizes and shapes of raindrops Specificdifferential phases are the rate of change of the range in pulsephases because these are not affected by attenuation partialbeam blockage or radar miscalibration they are an at-tractive parameter to use in QPE [31]

Rainfall rate data from ground gauge stations in Seouland Gyeonggi-do province within the radar umbrella areanalyzed in this study All stations obtain rainfall data everyminute but the QPE in this study uses rainfall rate data at10-min intervals is is to compare the radar data and tominimize the fluctuation of ground gauge data Figure 1shows the ground gauges densely distributed across theKorean peninsula of these 20 gauges within the radar ef-fective range are selected e number of 20 stations islocated in near Seoul and had severe storm damage in thepast e location of Gwangdeoksan radar and the selectedrainfall gauges are described in the zoomed area in Figure 1and information on each station is given in Table 1e useddata can be downloaded from the data base of KoreaMinistration Administration (KMA) at datakmagokr

22 Rainfall Events Rainfall events for which the depth ofthe observed daily rainfall exceeded 30mm from August toNovember 2018 are used Four events are selected as casestudies Two events (the first and second) occurred fromAugust 28ndash29 (event 1) and on September 3 (event 2)ethird event occurred from October 5ndash6 (event 3) and thefourth event happened on November 8 (event 4) Heavyprecipitation was observed in Korea brought on by a rainyfront (the Changma front) for event 1 and by low pressurein the northern area for event 2 Event 3 occurred as partof Typhoon KONG-REY and event 4 was accompanied bythe collision of a cold and a warm front

Table 2 summarizes information on the total rainfallmean rainfall rate and standard deviation for each event atthe station in Seoul (108) e amount of rain that fellduring event 1 was the largest (larger than 100mm) in twodays Both events 1 and 4 show that the weather frontgreatly affected the increase in rainfall or cloud formatione average rainfall is very large in the rainy season and



3 Methods


R θ0xθ11 x

θdd (d 1 2 n) (1)



Z 10 log ZH( )

DR 10 logZH

ZV( )


r2 minus r1

∣∣∣∣∣∣∣∣

∣∣∣∣∣∣∣∣

(2)


N

S

E

400 kilometers0 200

W




















Y Hβ (3)


H fa(XW + B) (4)



min12β

2+

C

2Hβ minus Y

2 (5)


1113954β HTH +1CI1113874 1113875

minus 1HTY (6)











RMSE

1n

1113944

n


2

11139741113972

(7)


correlation

1113936


1113936ni1 Ei minus E( 11138571113936


1113971

(8)


MAE 1n

1113944

n

i1Ei minus Oi

11138681113868111386811138681113868111386811138681113868 (9)

MBias 1n

1113944

n

i1Ei minus Oi( 1113857 (10)

RRMSE RMSE

otimes 100 (11)

5 Results








Number of trees

RMSE

(mm

hr)

0 100 300 400200 500970

975

980

985

990

(a)

Number of trees

RMSE

(mm

hr)

0 2000 6000 80004000

100

105

110

(b)

0 200 800600400 10001005

1010

1015

1020

1025

1030

1035

1040


RMSE

(mm

hr)

(c)


RMSE

(mm

hr)

7

8

9

10

11


ZR-L1ZR-L0

RFGBM

ELM

Input variable sets

(a)

Figure 3 Continued




COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM


(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)






Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy
















3 Methods


R θ0xθ11 x

θdd (d 1 2 n) (1)



Z 10 log ZH( )

DR 10 logZH

ZV( )


r2 minus r1

∣∣∣∣∣∣∣∣

∣∣∣∣∣∣∣∣

(2)


N

S

E

400 kilometers0 200

W




















Y Hβ (3)


H fa(XW + B) (4)



min12β

2+

C

2Hβ minus Y

2 (5)


1113954β HTH +1CI1113874 1113875

minus 1HTY (6)











RMSE

1n

1113944

n


2

11139741113972

(7)


correlation

1113936


1113936ni1 Ei minus E( 11138571113936


1113971

(8)


MAE 1n

1113944

n

i1Ei minus Oi

11138681113868111386811138681113868111386811138681113868 (9)

MBias 1n

1113944

n

i1Ei minus Oi( 1113857 (10)

RRMSE RMSE

otimes 100 (11)

5 Results








Number of trees

RMSE

(mm

hr)

0 100 300 400200 500970

975

980

985

990

(a)

Number of trees

RMSE

(mm

hr)

0 2000 6000 80004000

100

105

110

(b)

0 200 800600400 10001005

1010

1015

1020

1025

1030

1035

1040


RMSE

(mm

hr)

(c)


RMSE

(mm

hr)

7

8

9

10

11


ZR-L1ZR-L0

RFGBM

ELM

Input variable sets

(a)

Figure 3 Continued




COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM


(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)






Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy





























Y Hβ (3)


H fa(XW + B) (4)



min12β

2+

C

2Hβ minus Y

2 (5)


1113954β HTH +1CI1113874 1113875

minus 1HTY (6)











RMSE

1n

1113944

n


2

11139741113972

(7)


correlation

1113936


1113936ni1 Ei minus E( 11138571113936


1113971

(8)


MAE 1n

1113944

n

i1Ei minus Oi

11138681113868111386811138681113868111386811138681113868 (9)

MBias 1n

1113944

n

i1Ei minus Oi( 1113857 (10)

RRMSE RMSE

otimes 100 (11)

5 Results








Number of trees

RMSE

(mm

hr)

0 100 300 400200 500970

975

980

985

990

(a)

Number of trees

RMSE

(mm

hr)

0 2000 6000 80004000

100

105

110

(b)

0 200 800600400 10001005

1010

1015

1020

1025

1030

1035

1040


RMSE

(mm

hr)

(c)


RMSE

(mm

hr)

7

8

9

10

11


ZR-L1ZR-L0

RFGBM

ELM

Input variable sets

(a)

Figure 3 Continued




COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM


(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)






Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy


















RMSE

1n

1113944

n


2

11139741113972

(7)


correlation

1113936


1113936ni1 Ei minus E( 11138571113936


1113971

(8)


MAE 1n

1113944

n

i1Ei minus Oi

11138681113868111386811138681113868111386811138681113868 (9)

MBias 1n

1113944

n

i1Ei minus Oi( 1113857 (10)

RRMSE RMSE

otimes 100 (11)

5 Results








Number of trees

RMSE

(mm

hr)

0 100 300 400200 500970

975

980

985

990

(a)

Number of trees

RMSE

(mm

hr)

0 2000 6000 80004000

100

105

110

(b)

0 200 800600400 10001005

1010

1015

1020

1025

1030

1035

1040


RMSE

(mm

hr)

(c)


RMSE

(mm

hr)

7

8

9

10

11


ZR-L1ZR-L0

RFGBM

ELM

Input variable sets

(a)

Figure 3 Continued




COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM


(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)






Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy















Number of trees

RMSE

(mm

hr)

0 100 300 400200 500970

975

980

985

990

(a)

Number of trees

RMSE

(mm

hr)

0 2000 6000 80004000

100

105

110

(b)

0 200 800600400 10001005

1010

1015

1020

1025

1030

1035

1040


RMSE

(mm

hr)

(c)


RMSE

(mm

hr)

7

8

9

10

11


ZR-L1ZR-L0

RFGBM

ELM

Input variable sets

(a)

Figure 3 Continued




COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM


(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)






Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy

















COR

00

02

04

06

08

ZR-L1ZR-L0

RFGBM

ELM


(b)

MA

E (m

mh

r)

40

50

60

ZR-L1ZR-L0

RFGBM

ELM


(c)

MBi

as (m

mh

r)

ndash15

ndash05

05

15

ZRminusL1ZRminusL0

RFGBM

ELM


(d)

ZRminusL1ZRminusL0

RFGBM

ELM


RRM

SE (

)

100

120

140

(e)






Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy


















Estim

atio

ns (m

mh

r)


100

40

0

(a)

0 60 140Es

timat

ions

(mm

hr)

Observations (mmhr)

100

40

0

(b)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(c)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(d)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(e)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(f )

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(g)

0 60 140

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

40

0

(h)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(i)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(j)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(k)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(l)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(m)

0 60 140

40

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

0

(n)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(o)

0 60 140

0

40

100

Estim

atio

ns (m

mh

r)

Observations (mmhr)

(p)


Estim

atio

ns (m

mh

r)

100

40

0

(q)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(r)

0 60 140

Estim

atio

ns (m

mh

r)

Observations (mmhr)

100

40

0

(s)

0 60 140

0

40

100Es

timat

ions

(mm

hr)

Observations (mmhr)

(t)










1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy






















1

ZR-L1 1197 1191 1198 1812 1192ZR-L0 1459 1462 1462 1786 1466RF 1129 1130 1101 1552 1120

GBM 1105 1105 1103 1633 1106ELM 1030 1011 1021 1312 1006

2

ZR-L1 539 526 540 700 530ZR-L0 613 599 615 751 600RF 525 516 514 634 510

GBM 508 516 507 653 515ELM 493 517 471 599 505

3

ZR-L1 333 333 333 357 333ZR-L0 334 333 334 357 333RF 347 317 343 354 320

GBM 335 331 335 348 332ELM 364 391 375 385 399

4

ZR-L1 288 286 288 324 286ZR-L0 305 302 305 326 302RF 293 297 289 339 285

GBM 289 286 288 322 286ELM 291 290 290 308 289





1

ZR-L1 0750 0753 0749 0154 0752ZR-L0 0603 0600 0600 0201 0597RF 0785 0785 0800 0513 0796

GBM 0799 0799 0800 0425 0799ELM 0829 0837 0830 0687 0836

2

ZR-L1 0721 0736 0721 0455 0732ZR-L0 0615 0639 0611 0332 0637RF 0745 0750 0749 0593 0756

GBM 0758 0749 0759 0559 0750ELM 0793 0785 0805 0649 0788

3

ZR-L1 0308 0309 0308 0007 0308ZR-L0 0287 0292 0286 minus 0056 0292RF 0228 0423 0230 0208 0404

GBM 0298 0362 0298 0207 0358ELM 0364 0417 0355 0371 0389

4

ZR-L1 0418 0431 0419 0047 0429ZR-L0 0275 0299 0279 minus 0065 0298RF 0433 0418 0449 0018 0459

GBM 0422 0435 0423 minus 0004 0433ELM 0395 0409 0400 0258 0415





1

ZR-L1 855 851 856 1295 852ZR-L0 1043 1045 1045 1276 1047RF 811 802 785 1122 798

GBM 788 790 791 1168 792ELM 735 721 729 935 718

2

ZR-L1 778 759 779 1010 764ZR-L0 885 864 888 1084 866RF 764 740 745 916 736

GBM 731 746 732 941 745ELM 712 750 681 865 726

3

ZR-L1 636 637 636 683 637ZR-L0 638 638 638 683 638RF 663 606 656 674 608

GBM 641 634 641 666 634ELM 696 758 719 733 769

4

ZR-L1 635 630 635 714 630ZR-L0 674 667 673 720 667RF 647 653 641 750 630

GBM 636 630 635 714 630ELM 642 638 638 680 634






6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy


















6 Discussion









Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy


















Rada

r (m

mh

r)


100

40

0

(a)

Rada

r (m

mh

r)


100

40

0

(b)

Rada

r (m

mh

r)


100

40

0

(c)

Rada

r (m

mh

r)


100

40

0

(d)

Rada

r (m

mh

r)


100

40

0

(e)

Rada

r (m

mh

r)


100

40

0

(f )

Rada

r (m

mh

r)


100

40

0

(g)

Rada

r (m

mh

r)


100

40

0

(h)

Rada

r (m

mh

r)


100

40

0

(i)

Rada

r (m

mh

r)


100

40

0

(j)

Rada

r (m

mh

r)


100

40

0

(k)

Rada

r (m

mh

r)


100

40

0

(l)

Rada

r (m

mh

r)


100

40

0

(m)

Rada

r (m

mh

r)


100

40

0

(n)

Rada

r (m

mh

r)


100

40

0

(o)

Rada

r (m

mh

r)


100

40

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 60 140

100

40

0

(q)

Rada

r (m

mh

r)


100

40

0

(r)

Rada

r (m

mh

r)


100

40

0

(s)

Rada

r (m

mh

r)


100

40

0

(t)





Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy

















Rada

r (m

mh

r)

Ground gauge (mmhr)

15

5

0

0 10 20

(a)

Rada

r (m

mh


0 10 20

15

5

0

(b)

Rada

r (m

mh

r)


15

5

0

(c)

Rada

r (m

mh

r)


15

5

0

(d)

Rada

r (m

mh

r)


15

5

0

(e)

Rada

r (m

mh

r)


15

5

0

(f )

Rada

r (m

mh

r)


15

5

0

(g)

Rada

r (m

mh

r)


15

5

0

(h)

Rada

r (m

mh

r)


15

5

0

(i)

Rada

r (m

mh

r)


15

5

0

(j)

Rada

r (m

mh

r)


15

5

0

(k)

Rada

r (m

mh

r)


15

5

0

(l)

Rada

r (m

mh

r)


15

5

0

(m)

Rada

r (m

mh

r)


15

5

0

(n)

Rada

r (m

mh

r)


15

5

0

(o)

Rada

r (m

mh

r)


15

5

0

(p)

Ground gauge (mmhr)

Rada

r (m

mh

r)

0 10 20

15

5

0

(q)

Rada

r (m

mh

r)


15

5

0

(r)

Rada

r (m

mh

r)


15

5

0

(s)

Rada

r (m

mh

r)


15

5

0

(t)






AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy


















AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(a)

AWS

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

GDK radar

50ndash60

60ndash80

80ndash100

No data

mmhr

(b)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(c)

AWS GDK radar

0ndash2

2ndash4

4ndash6

6ndash8

8ndash10

10ndash15

15ndash20

20ndash30

30ndash40

40ndash50

50ndash60

60ndash80

80ndash100

No data

mmhr

(d)




7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy
















7 Conclusions


GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(a)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(b)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(c)

GDK radarAWSmmhr

0ndash1

1ndash2

2ndash3

3ndash4

4ndash5

5ndash6

6ndash7

7ndash8

8ndash9

9ndash10

10ndash11

11ndash12

12ndash13

13ndash14

14ndash15

No data

(d)








Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy




















Data Availability




Acknowledgments


References




















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy

















































Journal of










Volume 2018





Geophysics


Journal of




Microbiology





Agronomy















assessingtheapplicabilityofrandomforest,stochasticgradient...

Documents