Using 3D-Var Data Assimilation for Improving the Accuracy of Initial Condition of Weather Research and Forecasting (WRF) Model in Java Region (Case Study : 23 January 2015)
ISSN: 0852-0682, EISSN: 2460-3945
, 30 2 20 6: 2- 9 20 6 -N -N 4 0
Using 3D-Var Data Assimilation for Improving the Accuracy of Initial
Condition of Weather Research and Forecasting (WRF) Model in Java
Region (Case Study : 23 January 2015)
,1 2 2 1 1 Novvria Sagita* , Rini Hidayati , Rahmat Hidayat , Indra , and Fatkhuroyan 1) Gustari 2) Indonesian Agency for Meteorology, Climathology, and Geophysics (BMKG) Department of Geophysics and Meteorology, Bogor Agricultural University (IPB) )- Corresponding author (e-mail: novvria.sagita.ut@gmail.com)
Abstract .
:HDWKHU 5HVHDUFK DQG )RUHFDVWLQJ :5) LV D QXPHULFDO ZHDWKHU SUHGLFWLRQ
PRGHO GHYHORSHG E\ YDULRXV SDUWLHV GXH WR LWV RSHQ VRXUFH EXW WKH :5) KDV WKH
GLVDGYDQWDJH RI ORZ DFFXUDF\ LQ ZHDWKHU SUHGLFWLRQ 2QH UHDVRQ RI ORZ DFFXUDF\ RI PRGHO
LV LQDFFXUDF\ LQLWLDO FRQGLWLRQ PRGHO WR WKH DFWXDO DWPRVSKHULF FRQGLWLRQV 7HFKQLTXHV
WR LPSURYH WKH LQLWLDO FRQGLWLRQ PRGHO LV WKH REVHUYDWLRQ GDWD DVVLPLODWLRQ ,Q WKLV
VWXG\ ZH XVHG WKUHH GLPHQVLRQDO YDULDWLRQDO ' 9DU WR SHUIRUP GDWD DVVLPLODWLRQ
RI VRPH REVHUYDWLRQ GDWD 2EVHUYDWLRQDO GDWD XVHG LQ GDWD DVVLPLODWLRQ DUH REVHUYDWLRQ
GDWD IURP EDVLF VWDWLRQV QRQ EDVLF VWDWLRQV UDGLRVRQGH GDWD DQG 7KH %LQDU\ 8QLYHUVDO
)RUP IRU WKH 5HSUHVHQWDWLRQ RI PHWHRURORJLFDO GDWD %8)5 data from the National
&HQWHUV IRU (QYLURQPHQWDO 3UHGLFWLRQ 1&(3 DQG DJJUHJDWH REVHUYDWLRQ GDWD IURP
DOO VWDWLRQV 7KH DLP RI WKLV VWXG\ FRPSDUHV WKH HIIHFW RI GDWD DVVLPLODWLRQ ZLWK GLIIHUHQW
GDWD REVHUYDWLRQ RQ -DQXDU\ DW 87& IRU -DYD LVODQG UHJLRQ 7KH UHVXOWV
VKRZHG WKDW FKDQJHV URRW PHDQ VTXDUH HUURU 506( RI VXUIDFH WHPSHUDWXUH IURP
o o o& WR & & GHZ SRLQW IURP & WR & & UHODWLYH KXPLGLW\ IURP
WR DIWHU WKH GDWD DVVLPLODWLRQ
Keywords: :5) LQLWLDO FRQGLWLRQ GDWD DVVLPLODWLRQ ' 9DU Abstrak.
:HDWKHU 5HVHDUFK DQG )RUHFDVWLQJ :5) PHUXSDNDQ VHEXDK PRGHO SUHGLNVL FXDFD
QXPHULN \DQJ VHGDQJ GLNHPEDQJNDQ ROHK EHUEDJDL SLKDN NDUHQD EHUVLIDW WHUEXND RSHQ VRXUFH
WHWDSL :5) PHPLOLNL NHNXUDQJDQ EHUXSD NHDNXUDVLDQ SUHGLNVL FXDFD \DQJ NXUDQJ EDLN 6DODK
VDWX SHQ\HEDEQ\D DGDODK NHWLGDNVHVXDLQ V\DUDW DZDO PRGHO LQLWLDO FRQGLWLRQ WHUKDGDS NRQGLVL
DWPRVIHU DNWXDO 6DODK VDWX PHWRGH XQWXN PHPSHUEDLNL V\DUDW DZDO PRGHO DGDODK GHQJDQ
DVLPLODVL GDWD REVHUYDVL PHQJJXQDNDQ PHWRGH WKUHH GLPHQVLRQDO YDULWLRQDO ' 9DU 'DWD
REVHUYDVL \DQJ GLJXQDNDQ DGDODK GDWD REVHUYDVL SHUPXNDDQ GDUL VWDVLXQ XWDPD GDWD REVHUYDVL
SHUPXNDDQ GDUL VWDVLXQ QRQ XWDPD GDWD REVHUYDVL UDVRQ GDWD REVHUYDVL IRUPDW %8)5 GDUL
1DWLRQDO &HQWHUV IRU (QYLURQPHQWDO 3UHGLFWLRQ 1&(3 GDQ GDWD JDEXQJDQ GDUL GDWD REVHUYDVL
GDUL VHOXUXK VWDVLXQ 7XMXDQ SHQHOLWLDQ LQL XQWXN PHPEDQGLQJNDQ SHQJDUXK DVLPLODVL GDWD
GHQJDQ GDWD \DQJ EHUEHGD WHUKDGDS V\DUDW DZDO PRGHO SDGD WDQJJDO -DQXDUL MDP
87& XQWXN ZLOD\DK SXODX -DZD +DVLO SHQHOLWLDQ PHQXQMXNNDQ SHUXEDKDQ QLODL URRW PHDQ
o o
VTXDUH HUURU 506( V\DUDW DZDO PRGHO SDGD HVWLPDVL VXKX GDUL & PHQMDGL NLVDUDQ &
o o o oKLQJJD & WLWLN HPEXQ GDUL & PHQMDGL NLVDUDQ & KLQJJD & GDQ NHOHPEDSDQ QLVEL
GDUL PHQMDGL KLQJJD VHWHODK DVLPLODVL
Kata kunci: :5) V\DUDW DZDO DVLPLODVL GDWD ' 9DU
: j 25 2
3 -
I
I S , et al.
1. Introduction
One of the problems of NWP system model is the inaccurate of initial conditions models. Initial condition of NWP system is one that is closer to the actual atmospheric FRQGLWLRQV 2QH WHFKQLTXH LQ IRUPLQJ WKH initial condition model that approaches the actual data of the atmosphere is by data DVVLPLODWLRQ 'DWD DVVLPLODWLRQ LV WKH WHFKQLTXH for combining observations data with an 1:3 SURGXFW WKH ÀUVW JXHVV RU EDFNJURXQG forecast) and their respective error statistics to provide an improved estimate (the analysis) of the atmospheric state (Skamarock et al., 2005; Talagrand, 1997).
Indonesia Agency for Meteorology, &OLPDWRORJ\ DQG *HRSK\VLFV %0.* GLÀQHV the types of stations into the basic and the non-basic based on the obligation insending the data. The basic station has the obligation to send data to the international and non- basic station only has the obligation of data into a national system (BMKG, 2014). Data coming from the basic station is the one used in building global forecasting models. Utilization of observational data from non-basic station has not been made to establish a system of global NWP models. Data assimilation method is used to incoporate data observations from non-basic station into establishing the system of NWP models. Observation data itself is divided into two types i.e. surface and upper air data.
Surface observation data are meteorological parameters up to a height of 10-meters, while the upper air observation data cover parameter up to a height of 10 kilometers (WMO, 2003).
The development of data assimilation method has improved weather prediction models. For instances, assimilation of observation data affects the predictions of planetary boundary layeras (PBL) heights as shown by Stauffer et al. (1991) who studied the effect of direct assimilation surface temperature by four-dimensional method of data assimilation (FDDA) resulting in reduced errors but the surface temperature can trigger high errors on the PBL height for changes that DUH QRW UHDOLVWLF IRU VXUIDFH EXR\DQF\ ÁX[ Alapaty et al. (2001) used assimilation method to increase the one-dimensional simulation RI 3%/ 7KH PHWKRG FDQ VLJQLÀFDQWO\ UHGXFH the error model of PBL height. One method that has been known is the three-dimensional variational methods of data assimilation
(3D-Var), which was introduced by Lorenc ' 9DU PHWKRG FDQ UHGXFH RYHUHVWLPDWH precipitation WRF models in Tanzania (Athumani, 2012). Using the method of 3D-Var assimilation from Automatic Weather Station (AWS) data produces little improvement on the prediction of meteorological parameters (Junaedhi, IG et al., 2008; Dash, SK et al., 2013; Hou, T et al., 2013; Sahu, DK et al, 2014). 3D-Var method has a disadvantage that is not sensitive to the uncertainty of the vertical limit (Gao et al, 2004). Development of data assimilation methods produce another method called the four-dimensional variational methods of GDWD DVVLPLODWLRQ ' 9DU WKH WHFKQLTXH RI advanced data assimilation which takes into account the observation data every hour (Yang et al., 2009; Banister, R.N., 2007).
WRF is a numerical weather forecast system designed for atmospheric research to weather forecasting. WRF has been developed by the National Center for Atmospheric Research (NCAR), the National Centers for Environmental Prediction (NCEP), Forecast Systems Laboratory (FSL), Air Force Weather Agency (AFWA), Naval Research Laboratory, University of Oklahoma and the Federation Aviation Administration (FAA) since late 1990 (Skamarock et al., 2005). WRF allows researchers to generate atmospheric simulations based on real data observation. Part of WRF is used for the assimillation of the data referred to WRFDA (WRF Data Assimiation) (Skamarock et al., 2005).
This study will examine the difference impacts of the data assimilation using observation from the basic stations (bas), non- basic stations (non-bas), upper air data (rason), all observation data (all) and observation data from NCEP in the format of BUFR by comparing between the value of root mean
VTXDUH HUURU 506( RI LQLWLDO FRQGLWLRQV without data assimilation and with data assimilation. This study also analyzes the spatial distribution of bias in the form of the difference of initial conditions between models with data assimilated initial condition and the model without data assimilation (control).
2. Research Method
We used WRF model data with input data from Global Forecasting System (GFS)
3 -
11 A.yani 110.42
42 Basic
8 Tegal 109.15
3 Basic
9 Cilacap -7.73 109.02 Basic
10 Tj.Emas 110.38
3 Basic
3 Basic
8 Basic Citeko 920 Basic
12 Bawean -5.85
3 Basic
13 Tj.Perak -7.13
3 Basic
14 Juanda -7.37 112.77
3 Basic
15 Tj.Perak 2 -7.20937 112.74
7 Jatiwangi 108.27
5 Cengkareng
I
Flowchart of data assimilation Table 1.
I S , et al.
with a resolution of 0.5 o x 0.5 o . Observation data is divided into a basic stations data, non- basic stations data, upper air observation data (rason) and observational data from NCEP in BUFR format.
NCEP data consists of a surface observation data, upper air observation data sent to the Global Telecommunication System (GTS), and data from satellites NESDIS (National Environmental Satellite Data and Information Service). This research will only be focused in java island region on January 23, 2015 at 00.00 UTC when. There 17 basic stations and 14 non-basic stations used in this study. Table 1 describes those stations. In addition, the only used rason data in Java is in Juanda meteorological stations.
7KH ZRUNÁRZ RI WKLV VWXG\ FDQ EH VHHQ in Figure 1. This study uses a background ( ) and background error (B) WRF models. Background ( ) is the initial conditions of the model obtained from the simulation results in the WRF-GFS data input. Background error (B) is the statistical error of the model. Background errors are divided into: the global and regional background.
Background global error has been provided by the WRFDA. This study used the data assimilation with the background of global error that has been provided in the application WRFDA. is an observation data used to data assimilation and error value of observation shown by the parameter R.
Figure 1.
NO Station
4 Basic
ID station Latitude Longitude Elevation Kind of Station
1 Serang
50 Basic
2 Curug Basic
3 Tj.Priok
2 Basic
4 Kemayoran
3 Basic
3 -
(2)
h. Background error 9 km 45 s one 120 grid 120 grid 33 level Thompson scheme (Thompson, 2004) RUC %HWWV 0LOOHU -DQMLF VFKHPH %HWWV DQG 0LOOHU Global (Barker et al., 2005)
The method used for the data assimilation is varitional three-dimensional analysis (3D-Var). Formula of 3D-Var is (Kalnay, 2003):
^ `
1
1
1 (
2 T o o b b J x y H x R y H x x x B x x
ª º ª º ¬ ¼ ¬ ¼ (1)
1
1
2 ( ) T o o b b
J x y H x R y H x x x B x x
ª º ª º ¬ ¼ ¬ ¼
1
6XUIDFHB3K\VLFV
1
2 ( T o o b b
J x y H x R y H x x x B x x
ª º ª º ¬ ¼ ¬ ¼
(3)
1
1
2 ( T o o b b
J x y H x R y H x x x B x x
ª º ª º ¬ ¼ ¬ ¼
(4) 1 1
2 T b b o b o b J x x x B x x y H x H x x R y H x H x x
ª º ª º ¬ ¼ ¬ ¼ (5)
g. Kumulus
I
I
24 Karangploso -7.90139 112.597 Non-Basic
I S , et al.
NO Station
ID Station Latitude Longitude Elevation Kind of Station Kalianget -7.05 113.97
3 Basic
17 Banyuwangi -8.21 114.35
43 Basic
18 Pondok Betung Non-Basic
19 Tangerang
14 Non-Basic
20 Dramaga 207 Non-Basic
21 Bandung 829 Non-Basic
22 Banjarnegara -7.318 109.71 Non-Basic
23 Semarang 110.42 227 Non-Basic
25 Tretes -7.7 832 Non-Basic
Karangkates -8.15 112.45 325 Non-Basic0LNURÀVLN
27 Sawahan -7.74 111.79 835 Non-Basic
28 Adi Sumarmo (AURI) -7.5 110.75 127 Non-Basic
29 Abdurahman Saleh (AURI) -7.917 112.7 523 Non-Basic
30 Atang Sanjaya (AURI) Non-Basic
31 Tunggul Wulung 109.05
21 Non-Basic
The intial condition ( ), the background error (B), and observation and the error ( ) were used as input to the application
WRFDA to assimilate the data and generate the new initial conditions ( ) after the model RI DVVLPLODWLRQ :5) FRQÀJXUDWLRQ VHWWLQJV DUH performed in (Table 2) :
Table 2 :5) FRQÀJXUDWLRQ
a. Resolusi
b. Time step
c. Domain
d. Grid Utara-Selatan Timur-Barat Level vertikal e.
Divide (x-x b ) T to be :
3 -
1 T b T o
2 7 T T o
7KH LQÁXHQFH RI GDWD DVVLPLODWLRQ methods of 3D-Var to the initial conditions lead to a decrease of RMSE values of initial condition model for estimating surface weather parameter (surface temperature, dewpoint, and
4. Conclusion
On the data assimilation BUFR provides a broad impact on the parameters indicated surface temperature of the surface temperature distribution bias initial condition of data assimilation models BUFR the initial condition of assimilation model without reaching areas outside Java.
Figure 2 shows the estimation temperature bias at a height of 2 meters of the initial condition of the models with assimilation against initial condition of without assimilation. In general, all experiment of data assimilation resulted positive value of bias value for the eastern part of Java and negative bias for the region of western Java.
Table 2. E SE , ,
1.9 o C (rason), 1.4 o C (all), and 1.5 °C (BUFR). The RMSE value for RH have changed from :$ WR EDV QRQ EDV % (rason), 3.7 % (all), 3.5 % (BUFR). Almost all condition with data assimilation have RMSE value that are smaller than the initial condition without data assimilation. Data assimilation with rason generate increased RMSE for temperature parameters. This occurs because only one station used in data assimilation. While data assimilation with bass, all and BUFR have the smallest RMSE values for the third parameter (temperature, dewpoint, and relative humidity) weather among other data assimilation.This chapter contains the results of research.
1.7 o C (all), and 1.7 °C (BUFR). For dewpoint temperature RMSE values have changed from 2.1 °C (WA) to 1.4 ° C (bass), 1.5 °C (non-bass),
506( YDOXH LV XVHG WR LGHQWLI\ LQÁXHQFH data assimilation. Table 4 shows the changes of RMSE value for temperature from 2.0 o C (WA) to 1.7 °C (bass), 1.9 °C (non-bass), 2.4 o C (rason),
7KLV VWXG\ DQDO\]HV WKH LQÁXHQFH RI GDWD assimilation to the initial conditions (initial condition) models with several different observational data source. Analysis of the data DVVLPLODWLRQ LPSDFWV RQ WKH LQLWLDO UHTXLUHPHQW model is an important step for checking the LQÁXHQFH RI DVVLPLODWLRQ RI GDWD IRU ZHDWKHU prediction. Analysis of the impact of the initial condition the model is done by comparing the value of RMSE between initial condition without data assimilation against observational data (TA-O) and the initial conditions of models with data assimilation (bas, non-bas, rason, sma, and BUFR) against observation data (AO).
(7) J (x) is a function that calculates the cost dissimilarity between models with observational data surface. The calculation of WKH FRVW IXQFWLRQ - [ GHVFULEHG LQ HTXDWLRQV where x and x b is an analysis of data expected and the data background. H is the observation operator, R is the observation error covariance operator, B is operator background error, y is an observation data and y o is observation data in the model grid.
x H R H H R y H x
1
I
1
1
B H R H x x H R y H x
1
1
so,
When
(6)
1 b T b T o J x B x x H R H x x H R y H x
1
1
I S , et al.
3. Results and Discussions
3 -
I
I S , et al.
relative humidity). The best data assimilation data assimilation are also spatially visible for used for decreasing RMSE value of the initial the surface temperature estimation which condition model are the ones that used data underestimated after data assimilation in the from the basic station, surface observation western part of Java and overestimated in the data from all stations, and data BUFR from eastern part of Java. 1&(3 ,Q WKH VWXG\ DUHD WKH LQÁXHQFHV RI Figure 2.
Bias temperature estimation model of assimilation 2m initial conditions (a) bass, (b) rason, (c) non, (d)
BUFR, and (e) all against without assimilation (control).
5. References
Alapaty, K., Seaman, NL., Niyogi, DS., Hanna, A.F.(2001) $VVLPLODWLQJ 6XUIDFH 'DWD WR ,PSURYH WKH
$FFXUDF\ RI $WPRVSKHULF %RXQGDU\ /D\HU 6LPXODWLRQV. Journal Of Applied Meteorology.40,
Yang, S.-C. et al., (2009)Comparison of Local Ensemble Transform Kalman Filter, 3DVAR, and
3 -
I
I S , et al.
4DVAR in a Quasigeostrophic Model.
0RQWKO\ :HDWKHU 5HYLHZ SS ²
Bannister, R.N.,(2007) (OHPHQWDU\ 'YDU '$5& 7HFKQLFDO 5HSRUW 1R . Data Assimilation Research Centre, University of Reading, UK
Barker, D., Guo, Y.R., Lin, H.C.(2005) 6SHFLÀFDWLRQ DQG 8VH RI )RUHFDVW (UURU &RYDULDQFHV LQ :5) 9DU.
NCAR, Taiwan. %HWWV $ . DQG 0LOOHU 0 - $ QHZ FRQYHFWLYH DGMXVWPHQW VFKHPH 3DUW ,, 6LQJOH FROXPQ WHVWV
XVLQJ *$7( ZDYH %20(; $7(; DQG $UFWLF $LUPDVV GDWD VHWV. Q4 - 5 0HWHRURO 6RF , 112,
BMKG, (2014)Regulation of Chief Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG) No 12. Jakarta
WMO, (2003) Manual on the Global Observing System. World Meteorological Organization (WMO). Geneva. &KLQ + 1 6 *ODVFRH / /XQGTXLVW - :KDUWRQ 6 ,PSDFW RI :5) 3K\VLFV DQG *ULG 5HVROXWLRQ
RQ /RZ /HYHO :LQG 3UHGLFWLRQ 7RZDUG WKH $VVHVPHQW RI &OLPDWH &KDQJH ,PSDFW RQ )XWXUH :LQG
3RZHU.Dash, S.K., Sahu, D.K., Sahu, S.C.(2013) ,PSDFW RI $:6 REVHUYDWLRQ LQ :5) '9$5 GDWD DVVLPLODWLRQ
V\VWHP D FDVH VWXG\ RQ DEQRUPDO ZDUPLQJ FRQGLWLRQ LQ 2GLVKD 1HW +D]DUG
Gao, J.D. et al., (2004) $ WKUHH GLPHQVLRQDO YDULDWLRQDO GDWD DQDO\VLV PHWKRG ZLWK UHFXUVLYH ÀOWHU IRU
'RSSOHU UDGDUV -RXUQDO RI $WPRVSKHULF DQG 2FHDQLF 7HFKQRORJ\ SS ² Hou, T., Kong, F., Chen, X. and Lei, H.,(2013) ,PSDFW RI '9$5 GDWD DVVLPLODWLRQ RQ WKH SUHGLFWLRQ RI KHDY\ UDLQIDOO RYHU 6RXWKHUQ &KLQD.$GYDQFHV LQ 0HWHRURORJ\, , 1-17.
Junnaedhi, I.D.G.(2008)
3HQJDUXK DVLPLODVL GDWD GHQJDQ PHWRGH '9$5 WHUKDGDS KDVLO SUHGLNVL FXDFD QXPHULN GL ,QGRQHVLD Tesis Institut Teknologi Bandung. Bandung. 73 hlm.
Kalnay, E.(2003) $WPRVSKHULF PRGHOLQJ GDWD DVVLPLODWLRQ DQG SUHGLFWDELOLW\. New York. Cambridge University Press. Kurniawan, R., et al.(2014)
3HQJJXQDDQ 6NHPD .RQYHNWLI 0RGHO &XDFD :UI %HWWV 0LOOHU -DQMLF .DLQ
)ULWVFK 'DQ *UHOO G (QVHPEOH 6WXGL .DVXV 6XUDED\D 'DQ -DNDUWD . Jurnal meteorologi dan
JHRÀVLND 112,
/RUHQF $ & $QDO\VLV PHWKRGV IRU QXPHULFDO ZHDWKHU SUHGLFWLRQ 4 - 5 0HWHRURO 6RF 1177–1194. Rajeevan, M., Kesarkar, A., Thampi, S.B., Rao, T.N., Radhakrishna, B. and Rajasekhar, M.(2010)
6HQVLWLYLW\ RI :5) FORXG PLFURSK\VLFV WR VLPXODWLRQV RI D VHYHUH WKXQGHUVWRUP HYHQW RYHU 6RXWKHDVW
,QGLD. In $QQDOHV JHRSK\VLFDH DWPRVSKHUHV K\GURVSKHUHV DQG VSDFH VFLHQFHVSahu, D.K., Dash, S.K., Bhan, S.C.(2014)
,PSDFW RI 6XUIDFH REVHUYDWLRQV RQ VLPXODWLRQ RI UDLQIDOO RYHU
1&5 'HOKL XVLQJ UHJLRQDO EDFNJURXQG HUURU VWDWLVWLF LQ :5) '9$5 PRGHO. Meteorol Atmos
Phys.125:17-42.
Skamarock, W.C., Klemp, J.B., Dudhia, J., Gill, D.O., Barker, D.M., Wang, W. and Powers, J.G.(2005) $ GHVFULSWLRQ RI WKH DGYDQFHG UHVHDUFK :5) YHUVLRQ
1R 1&$5 71
STR). National Center For Atmospheric Research Boulder Co Mesoscale and Microscale
Meteorology Div.Stauffer, D.R., Seaman, N.L.,and Binkowski, F. S.(1991)
8VH RI IRXU GLPHQVLRQDO GDWD DVVLPLODWLRQ LQ
D OLPLWHG DUHD PH VRVFDOH PRGHO 3DUW ,, (IIHFWV RI GDWD DVVLPLODWLRQ ZLWKLQ WKH SODQHWDU\ ERXQGDU\
OD\HU.Mon. Wea. Rev., 119, 734–754.Talagrand, O.(1997)
$VVLPLODWLRQ RI REVHUYDWLRQ DQ LQWURGXFWLRQ. J. Met.Soc. Japan Special Issue 75, 1B, 191-209.
3 -
I
I S , et al.
Thompson, G., Rasmussen, R.M. & Manning, K.(2004)
([SOLFLW )RUHFDVWV RI :LQWHU 3UHFLSLWDWLRQ 8VLQJ DQ ,PSURYHG %XON 0LFURSK\VLFV 6FKHPH 3DUW , 'HVFULSWLRQ DQG 6HQVLWLYLW\ $QDO\VLV. Monthly
Weather Review, 132(2), 519–542. Yang, S.-C. et al., (2009)
&RPSDULVRQ RI /RFDO (QVHPEOH 7UDQVIRUP .DOPDQ )LOWHU '9$5 DQG '9$5 LQ D 4XDVLJHRVWURSKLF 0RGHO 0RQWKO\ :HDWKHU 5HYLHZ SS ²