Testing WRF capability in simulating the atmospheric water cycle over ...

Clim Dyn (2011) 37:1357–1379 DOI 10.1007/s00382-011-1024-2

Testing WRF capability in simulating the atmospheric water cycle over Equatorial East Africa Benjamin Pohl • Julien Cre´tat • Pierre Camberlin

Received: 13 September 2010 / Accepted: 7 February 2011 / Published online: 25 February 2011 Springer-Verlag 2011

Abstract Uncertainties in simulating the seasonal mean atmospheric water cycle in Equatorial East Africa are quantified using 58 one-year-long experiments performed with the Weather Research and Forecasting model (WRF). Tested parameters include physical parameterizations of atmospheric convection, cloud microphysics, planetary boundary layer, land-surface model and radiation schemes, as well as land-use categories (USGS vs. MODIS), lateral forcings (ERA-Interim and ERA40 reanalyses), and domain geometry (size and vertical resolution). Results show that (1) uncertainties, defined as the differences between the experiments, are larger than the biases; (2) the parameters exerting the largest influence on simulated rainfall are, in order of decreasing importance, the shortwave radiation scheme, the land-surface model, the domain size, followed by convective schemes and land-use categories; (3) cloud microphysics, lateral forcing reanalysis, the number of vertical levels and planetary boundary layer schemes appear to be of lesser importance at the seasonal scale. Though persisting biases (consisting of conditions that are too wet over the Indian Ocean and the Congo Basin and too dry over eastern Kenya) prevail in most experiments, several configurations simulate the regional climate with reasonable accuracy. Keywords WRF Regional climate modeling Water cycle East Africa Rainfall

B. Pohl (&) J. Cre´tat P. Camberlin Centre de Recherches de Climatologie, CNRS/Universite´ de Bourgogne, 6 boulevard Gabriel, 21000 Dijon, France e-mail: [email protected]

1 Introduction Rainfall variability is a key issue in East Africa, where several sectors of the economy depend acutely on water resources. However, the usefulness of seasonal forecasts or long-term climate projections is contingent upon our ability to resolve, at a sufficiently high resolution, the detailed patterns of the water cycle, especially rainfall distribution. This question can be addressed via downscaling methods. Downscaling is a particularly critical and demanding exercise in East Africa. This is due to the complex topography, which includes several mountains culminating above 4,000 m, elongated ridges and escarpments paralleling the faults of the East African Rift system, and large lakes embedded in contrasted topographical settings. Besides statistical methods, numerical simulations based on Regional Climate Models (RCM) are increasingly used to downscale rainfall patterns associated with large-scale climate forcings. Previous regional climate simulations dedicated to East Africa mostly relied on various versions of the ICTP (International Centre for Theoretical Physics) RegCM model. Using RegCM2, Sun et al. (1999a, b) carried out experiments aimed at defining an optimal model configuration to replicate precipitation amounts as observed over Equatorial Africa in October 1988, an average rainy season. They analyzed the part played by cumulus convection schemes, non-convective precipitation parameterization, radiative transfer formulation, surface processes, boundary layer formulation, and lateral boundary conditions. Using the same model, Anyah and Semazzi (2004) analyzed the response of East African rainfall to the surface temperature of Lake Victoria. Their model had a tendency to overestimate seasonal rainfall amounts over the eastern Congo Basin, Uganda and parts of Kenya. Anyah et al. (2006) and Anyah and Semazzi (2007)

123

1358

performed multi-year simulations on the RegCM3 model, which included changes in the precipitation and radiation schemes, but which resulted in a tendency to over-predict total precipitation. This prompted Davis et al. (2009) to carry out additional testing on RegCM3, and enlarge the model domain towards the Indian Ocean. They found that the Grell convection scheme underestimated convective rainfall rates over land, while the Massachusetts Institute of Technology–Emanuel scheme provided a realistic spatial distribution, despite a tendency to overestimate convective rainfall. The MIT-Emanuel scheme also produced a more realistic partitioning of stratiform and convective rainfall. Even more recently, Sylla et al. (2010) showed the capability of RegCM3 to simulate the African climate over a domain encompassing the whole continent, for the 1989–2005 period, at a 50 km resolution. The average annual cycle of rainfall and interannual climate variability were accurately reproduced. Using the Weather Research and Forecast (WRF) model, Zhang (2007) illustrated the sensitivity of simulated climate to radiation schemes. He also concluded that the average climate of East Africa and parts of the interannual variability were satisfactorily reproduced. Moore et al. (2010) evaluated the role played by more realistic land cover parameterization in simulating the East African climate. While a dramatic improvement is found for surface temperature, the effect on rainfall is generally small. Kaspar and Cubasch (2008) used the CLM regional model (Climate version of the Lokal-Modell, originating from Deutscher Wetterdienst). They performed simulations encompassing the whole year, and tested three configurations, differing in their convective parameterization schemes (Tiedtke and Kain-Fritsch, KF) and number of cloud ice categories. They found that, over Africa in general, the Tiedtke scheme performed better than KF, which strongly overestimated rainfall. However, the Tiedtke scheme strongly underestimated rainfall regionally, and detailed spatial patterns were not assessed. Most previous exercises in regional modeling for East Africa have concentrated on comparisons between convection schemes. As a consequence, the role of domain size, land-use datasets, and physical parameterizations of the planetary boundary layer (PBL) cloud microphysics (MP), land-surface models (LSM) and radiative budget, among others, remains insufficiently explored. A unified view and a precise quantification of the uncertainties associated with each of these settings remains to be established. The first aim of the present study is to participate in filling these gaps. The second aim is to extend previous analyses to the entire atmospheric water cycle (including rainfall, moisture transport and surface evapotranspiration) instead of precipitation alone. Here, we used the Weather Research and Forecasting model (WRF) because it allowed us to select a large panel

123

B. Pohl et al.: Testing WRF in East Africa

of physical schemes (Flaounas et al. 2010) to document the sensitivity to diverse settings (the physical package: PBL, MP, convective, radiation and LSM schemes; land-use surface categories and lateral forcing fields; domain size and vertical resolution). Due to the large number of parameters to be explored, a descriptive and synthetic approach was preferred to a detailed comprehensive analysis of each physical scheme. An extensive list of publications gives references for the parameterizations used and presented in this study. In this study, Sect. 2 describes the data, the model and the experimental setup. The following sections successively document the sensitivity to the various parameterizations tested. In Sect. 3, the planetary boundary layer, atmospheric convection and microphysics are considered. Section 4 is devoted to radiation schemes and land-surface models. Land-use conditions and lateral forcing fields are dealt with in Sect. 5. Section 6 examines the part played by domain size and vertical resolution. Section 7 summarizes the results by prioritizing all uncertainties. The main results are then discussed in Sect. 8.

2 Data, methods and experimental setup This paper focuses on rainfall over East Africa in 1999, at the seasonal timescale. All results presented herein concern the Long Rains (March through May: MAM). Similar analyses were applied to the Short Rains (October through December: OND) to assess the robustness and reproducibility of our results, but are not shown. Note that all our conclusions obtained for MAM hold for OND, despite the differences in the mean climate background between the two seasons. The year 1999 was chosen because (1) both rainy seasons show seasonal amounts fairly representative of the climatology; (2) sophisticated satellite estimates are available to document the spatial variability of the rainfall field; (3) reanalysis datasets are also more reliable during the satellite era, especially over Africa, where the amount of assimilated data is very low and inconsistent (Poccard et al. 2000). This issue is of non-negligible importance since reanalyses are used to drive the regional climate model. Rainfall data consist of the daily time series from 58 rain-gauge stations in Kenya and northern Tanzania, provided by the meteorological services of the two countries. For this paper we only retained the 44 stations that present no missing values in 1999. In order to complete this dataset, the GPCP-1dd estimations (Huffman et al. 2001) are used. Taken together, these two complementary datasets document well the spatial distribution of rainfall over the region (Fig. 1), showing larger seasonal rainfall amounts in the periphery of the Congo Basin and the region


1359

6N 3N 0 3S 6S 30E 100

200

35E 300

40E 400

45E 500

600

Fig. 1 Seasonal rainfall amounts (mm) in MAM 1999 estimated by GPCP (shading) and rain-gauge records (circles)

of Lake Victoria, and drier conditions in the eastern plains and lower slopes. GPCP is quite efficient at estimating the large-scale features of the seasonal rainfall field, but is logically less adequate to depict local-scale variability, especially over the areas characterized by complex topography. Atmospheric fields are taken from the ERA-Interim (Simmons et al. 2007) and ERA40 (Uppala et al. 2005) reanalyses. The same reanalysis data are also used to force the model laterally and provide its soil (temperature and moisture) and atmosphere initial conditions. All experiments were conducted with the Weather Research and Forecast/Advanced Research WRF (WRF hereafter) model, version 3.1 (Skamarock et al. 2008). This non-hydrostatic model, used in a large number of studies including several focusing on tropical Africa (e.g., Zhang 2007, Flaounas et al. 2010), presents a large number of different physical schemes, making it particularly adequate Table 1 Summary of the 58 one-year-long experiments, divided into four sets of runs Set #1 [27] Set #2 [24]

For each set, the cells in blue indicate the parameters to which sensitivity experiments are conducted. The number of experiments used in each set is indicated in the table. Settings in red correspond to the default configuration. See text for details

Set #3 [4]

Set #4 [6]

for the present work. Integrations are carried out between 1 January 1999 and 31 December 1999. A 2-month spin-up (January–February) was allowed to let the model adjust to soil and atmosphere initial conditions. We deliberately chose a short study period in order to multiply the number of experiments and quantify their sensitivity to a larger panel of physical schemes, boundary conditions and settings of the model geometry (three domain sizes and two vertical resolutions). Lateral forcing fields were prescribed every 6 h. Sea surface temperatures, provided by the reanalysis datasets, were interpolated at the daily timescale and updated every 24 h. Soil moisture and temperature provided by the forcing reanalysis were only used to initialize the land-surface models and were then calculated on-line. Table 1 summarizes the parameters whose influence was analyzed. For the physical package, cumulus (Cu), planetary boundary layer (PBL), microphysics, longwave (LW) and shortwave (SW) radiation schemes and landsurface model (LSM) parameterizations were considered. In each case, we retained several commonly used schemes (see Table 1). Sensitivity to land-use was estimated by prescribing two different sets of land-surface categories, respectively provided by the U.S. Geological Survey (USGS, 24 classes: Anderson et al. 1976, Hitt 1994) and more recently by the Moderate Resolution Imaging Spectroradiometer (MODIS, 20 classes: Friedl et al. 2002) databases. The influence of lateral forcings was quantified by successively driving WRF with ERA40 and ERAInterim reanalyses. In both cases, lateral forcings (sea surface temperatures) were similarly prescribed every 6 (24) h. The impact of domain size and location of the lateral boundaries was assessed by considering three distinct domains (E, L and XL, Fig. 2a), together with a fourth one (N, Fig. 2b) which was common to all experiments.

Physical Parameterizations Cu: KF, BMJ, Grell LW: RRTM PBL: ACM2, YSU, SW: Dudhia TKE MP: WSM6, Lin, LSM: NOAH Morrison LW: RRTM, RRTMG_LW Cu: KF SW: Dudhia, Goddard, PBL: ACM2 RRTMG_SW MP: WSM6 LSM: NOAH, Pleim(2), RUC Cu: KF LW: RRTM PBL: ACM2 SW: Dudhia MP: WSM6 LSM: NOAH

Cu: KF PBL: ACM2 MP: WSM6

LW: RRTM SW: Dudhia LSM: NOAH

Land Use

Forcing GCM

Domain Size

Vertical Res.

USGS

ERA-I

E:[10S 8N, 23E 50E]

28 lev.

USGS

ERA-I

E:[10S 8N, 23E 50E]

28 lev.

USGS ERA-I MODIS ERA40

E:[10S 8N, 23E 50E]

28 lev.

E:[10S 8N, 23E 50E] L:[15S 13N, 18E 55E] XL:[20S 18N, 13E 60E]

28 lev. 35 lev.

USGS

ERA-I

123

1360 Fig. 2 a Domains used in the study. b Domain E and location of the 44 rain-gauge stations used in this study. The inner red box corresponds to the Kenya rainfall index, for which a nested domain is used (see text for details)


(a)

(b)

3N

XL L

15N E

N

0

N

0

3S

15S

6S

30S 0

18E

36E

54E

This fourth domain was centered over Kenya and included Lake Victoria. It was two-way nested with the corresponding parent domain. In all cases, the spatial resolution was fixed at 36 km (12 km) for the parent (nested) domains. The three parent domains differed in that domain L (XL) contains 15 (30) more grid-points than domain E on all sides, corresponding to around 5 (10) degrees. Central grids are strictly similar, allowing direct comparisons. Lastly, the effect of the number of vertical levels was documented by comparing experiments using 28 and 35 sigma levels. Because of the huge number of possible combinations between all these settings, we chose to proceed iteratively, by designing four different sets of experiments. Set #1 deals for instance with Cu, PBL and MP schemes, while all other settings, relevant to Sets #2–4, were constant (see Table 1). A standard configuration must be defined for this protocol to set unchanging parameters for each step of the analysis. This configuration was obtained after preliminary analyses (not shown) in order to select a satisfactory combination. Corresponding settings appear in red in Table 1. We chose to work at the seasonal timescale, because our experimental setup was not designed to separate internal and external variability in the model (e.g., Giorgi and Bi 2000, Alexandru et al. 2007, Kgatuke et al. 2008). Hence, we did not study simulated wet and dry spells, nor did we analyze to what extent such spells occur in phase with observed data. We postulated that integrating model outputs from the daily to the seasonal timescale would greatly smooth their irreproducible component, in agreement with Kgatuke et al. (2008) and Vanvyve et al. (2008). We therefore interpreted the differences between our experiments as the consequences of the perturbations that we introduced, and considered internal variability as negligible. In order to validate this hypothesis, we duplicated an experiment by perturbing only its atmospheric initial conditions, and obtained fairly consistent and reproducible seasonal fields (not shown). Moreover, working at the seasonal scale enabled us to simplify the analysis of the

123

E

6N

30N

72E

30E

35E

40E

45E

atmospheric water budget (Meynadier et al. 2010a, b). The equation for water vapor closure (Trenberth and Guillemot 1998) is given by: MFD þ

oPW ¼EP ot

where P is cumulated precipitation, E is surface evapotranspiration, MFD is vertically integrated moisture divergence between the surface and the top of the atmosphere, and oPW=ot is the temporal variation of precipitable water within the air mass. At the seasonal scale, oPW=ot becomes negligible compared to the other three terms. We will therefore focus the capability of the model to simulate rainfall, surface evapotranspiration and vertically integrated moisture divergence fluxes.

3 Set #1: Atmospheric convection, planetary boundary layer, microphysics In this set, all possible combinations between 3 Cu, 3 PBL and 3 MP schemes were considered, leading to a first pool of 27 experiments. Sensitivity to these settings has often been documented in recent years by the regional modeling community (e.g., Bright and Mullen 2002; Jankov et al. 2005; Lim and Hong 2005; Im et al. 2008; Flaounas et al. 2010; Tchotchou and Kamga 2010, among many others). In the present paper, the schemes tested were: •

•

•

For atmospheric convection, Kain-Fritsch (KF, Kain 2004), Betts-Miller-Janjic (BMJ, Betts and Miller 1986, Janjic 1994) and the Grell-3D ensemble scheme (Grell and De´veńyi 2002). For PBL parameterizations, Yonsei University (YSU, Hong et al. 2006), the Mellor-Yamada-Janjic turbulent kinetic energy scheme (TKE, Mellor and Yamada 1982, Janjic 2002) and Asymmetrical Convective Model version 2 (ACM2, Pleim 2007a, b). For microphysics, the Lin et al. (1983) scheme, the WRF Single-Moment 6-Class scheme (WSM6, Hong and Lim 2006) and the Morrison et al. (2009) scheme.


1361

Levels (hPa)

Figure 3 shows the differences between the average of the 27 WRF simulations and ERA-Interim, in terms of vertical profiles of MAM seasonal mean temperature and air humidity. The air mass is drier over eastern Kenya and the East Africa Highlands in WRF, and slightly wetter over the nearby Indian Ocean. The difference is very small west of 30E. It is also too warm at 600 and 100 hPa and too cold between these two levels. The 100 hPa level, in particular, shows marked differences with the reanalyses, reaching up to 3 K, a feature remarkably constant from one experiment to another (not shown). The vertical resolution of the model (and particularly the number of sigma levels in the upper layers) could be at least partly responsible for this bias. Section 4 will be dedicated to this issue. The capability of these experiments to simulate the mean MAM rainfall field is documented in Figs. 4 and 5 for the overall domain, and in Table 2 for the rain-gauge station network. Moisture convergence biases are very similar to rainfall biases and are thus not shown. Evapotranspiration differences between WRF and ERA-Interim are shown in Fig. 6 and discussed below. Figure 4a shows seasonal rainfall amount biases against GPCP (Fig. 1) over domain E. Although the biases are very different from one experiment to another, common spatial structures suggest that parts of the biases are not directly dependent on the parameterization schemes tested in Set #1. Lake Victoria and the eastern slopes of Kenya and northern Tanzania are systematically too dry in WRF. In contrast, the nearby Indian Ocean and the eastern Congo Basin are too wet. In more detail, the BMJ Cu scheme clearly favors generalized dryness, while KF is associated

2.5

100 200 300 400 500 600 700

2 1.5 1 0.5

850 925 1000

0 −0.5 30E

35E

40E

45E

Longitudes Fig. 3 Vertical cross-sections of the atmosphere, averaged between the equator and 1S, showing the MAM seasonal mean differences in air temperature (K, shadings) and specific humidity (kg kg-1, contours) between the average of the 27 WRF experiments in Set #1 and ERA-Interim. Solid (dashed) contours correspond to positive (negative) differences. Contour interval is 5 9 10-4 kg kg-1 and the zero contour is omitted. Temperature (humidity) differences that are not significant at the 95% according to a t test applied to daily averages are shaded white (not contoured). The black patch corresponds to African surface topography such as it appears in WRF grids

with the largest rainfall amounts. MP and PBL parameterizations appear to be of less importance with regard to rainfall amounts. As a consequence, RMS statistics computed over a regional index encompassing Kenya (domain N in Fig. 2b) mainly vary according to the Cu scheme, lower (larger) values being found for the KF (BMJ) scheme. Grell convection produces results which fall between those of KF and BMJ. Consistent results are obtained at the more local scale of rain-gauge measurements (Table 2). Spatial correlations of seasonal rainfall amounts, computed between the raingauge stations and the nearest WRF grid-point, are clearly better for KF. These values, often reaching 0.6 (over 44 grid-points), suggest that the model can simulate the geography of the mean rainfall field with reasonable accuracy. The KF Cu scheme also produces the lowest RMS errors, especially when combined with the YSU or ACM2 PBL schemes. However, RMS values still remain quite large, given the moderate seasonal amounts (Fig. 1), and are indicative of either a methodological issue in the grid-to-point data comparison, or a systematic bias. This issue is further explored in Fig. 5, which shows for each experiment the rainfall amounts over the Kenya index. Because they present marked differences in terms of simulated seasonal rainfall, the continental part of this index and the Lake Victoria Basin are treated separately. For comparison purposes, GPCP and ERA-Interim rainfall amounts are shown, although their reliability over Lake Victoria is uncertain. Experiments using the KF scheme are indeed wetter than those using Grell. Experiments using BMJ are the driest. Although PBL parameterizations have little effect on the rainfall field, microphysics schemes appear to exert slight but still non-negligible influence on seasonal rainfall. The Morrison scheme is associated with the largest amounts, while WSM6 and Lin are drier. Analysis of the combinations between convective and microphysics schemes suggests that their respective influence is more or less additive (i.e., the overall largest amounts are obtained with KF cumulus and Morrison microphysics, the smallest ones with BMJ cumulus and Lin microphysics). This result is consistent with, e.g., Flaounas et al. (2010), as far as Cu schemes are concerned. Although most experiments show a larger bias over Lake Victoria than over land areas across Kenya (Fig. 4), as shown in Fig. 5b–c, the ranking of the experiments in terms of accuracy is virtually the same for the two sub-regions. Although Grell convection, using ACM2 PBL, performs quite well over land areas, it is deficient over Lake Victoria. KF convection thus offers the best compromise when land areas and water bodies are interspersed, as is the case over much of East Africa. In addition to seasonal rainfall amounts, Fig. 4b and c present the biases for the number of rainy days and their

123

1362


(a)

Cu KF

Cu BMJ

Cu Grell

PBL YSU

E1 − 162

E2 − 168

E3 − 189

E4 − 284

E5 − 282

E6 − 273

E7 − 210

E8 − 203

E9 − 195

PBL TKE PBL ACM2

E10 − 194

E11 − 192

E12 − 207

E13 − 279

E14 − 279

E15 − 274

E16 − 223

E17 − 220

E18 − 225

E19 − 184

E20 − 181

E21 − 240

E22 − 264

E23 − 257

E24 − 248

E25 − 197

E26 − 192

E27 − 187

MP Lin

MP WSM6 MP Morrison

MP Lin

−200 −150 −100

MP WSM6 MP Morrison −50

Cu KF

(b)

50

100

150

200

Cu BMJ E4 − 33.07

E5 − 32.12

Cu Grell

E1 − 16.19

PBL TKE PBL ACM2

E3 − 14.99

MP WSM6 MP Morrison

PBL YSU

E2 − 16.02

0

MP Lin

E6 − 27.47

E7 − 22.30

E8 − 22.04

E9 −21.41

E10 − 18.32 E11 − 18.16 E12 − 17.74 E13 − 32.69 E14 − 32.42 E15 − 29.04 E16 − 23.29 E17 − 22.98 E18 − 23.40

E19 − 16.56 E20 − 16.18 E21 − 14.41 E22 − 32.67 E23 − 32.09 E24 − 26.47 E25 − 20.03 E26 − 19.84 E27 − 18.64

MP Lin


(c)

MP Lin

−30

−20


Cu KF

0

10

20

MP Lin

30

MP WSM6 MP Morrison

40

Cu BMJ

Cu Grell

PBL YSU PBL TKE PBL ACM2

E1 − 2.7

E2 − 2.69

E3 − 3.08

E4 − 6.59

E5 − 7.59

E6 − 5.89

E7 − 2.57

E8 − 2.59

E10 − 2.9

E11 − 2.95

E12 − 3.17

E13 − 3.65

E14 − 3.82

E15 − 4.37

E16 − 3.34

E17 − 3.30 E18 − 3.28

E19 − 2.8

E20 − 2.75

E21 − 3.77

E22 − 9.59

E23 − 6.96

E24 − 6.13

E25 − 2.44

E26 − 2.43

MP Lin


MP Lin

MP WSM6 MP Morrison

−5

0

5

MP Lin

E9 − 2.48

E27 − 2.54

MP WSM6 MP Morrison

10

Fig. 4 a MAM seasonal rainfall amount biases (mm) for the 27 WRF experiments in Set #1. Experiments shown on the same row/ column/3 9 3 block use the same PBL/MP/Cu scheme, respectively. Shadings show the biases against GPCP, interpolated onto WRF grids. Biases that are not significant at the 95% according to a t test

applied to daily averages are shaded white. For each experiment, seasonal RMS errors computed over the Kenya rainfall index (domain N on Fig. 2b) are labeled on the figure. b As a but for the number of rainy days ([1 mm). c As a but for the average intensity of rainy days (mm)

average intensity against GPCP. The aim here is to investigate how the differences noted in Figs. 4a and 5 affect the statistical distribution of daily rainfall at the grid-

point scale. The definition of a rainy day (daily rainfall amount reaching at least 1 mm over the corresponding grid) is likely to be affected by the spatial resolution of the

123

B. Pohl et al.: Testing WRF in East Africa PBL YSU

(a)

PBL TKE

PBL ACM2

500 MP Lin MP WSM6 MP Morrison

400 300 200

Cu KF

Cu BMJ Cu Grell PBL YSU

Cu KF

Cu KF

E27

E24

E25 E26

E23

E21 E22

E18

Cu BMJ Cu Grell PBL TKE

E19 E20

E17

E15 E16

E13 E14

E11 E12

E9 E10

E7 E8

E5

E6

E4

E3

E1 E2

0

GPCP

100

ERA−I

Fig. 5 a MAM seasonal rainfall amounts averaged spatially over the Kenya rainfall index (mm) according to GPCP, ERA-Interim and the 27 WRF experiments in Set #1. Except for GPCP, the smallest bar indicates the convective rainfall seasonal amounts. b As a but for the continental part of the Kenya rainfall index (35E– 40E). c As a but for the Lake Victoria Basin (32E–35E)

1363

Cu BMJ Cu Grell PBL ACM2

(b) 300 MP Lin MP WSM6 MP Morrison

250 200 150 100

Cu KF

Cu BMJ Cu Grell PBL YSU

Cu KF

Cu BMJ Cu Grell PBL TKE

Cu KF

E27

E26

E24 E25

E23

E21 E22

E19 E20

E18

E17

E15 E16

E13

E14

E12

E9 E10 E11

E8

E6 E7

E5

E4

E3

E2

GPCP

0

ERA−I E1

50

Cu BMJ Cu Grell PBL ACM2

(c) 800 MP Lin MP WSM6 MP Morrison

600

400

Cu KF

Cu BMJ Cu Grell

respective datasets (Osborn and Hulme 1997, Sun et al. 2006, Ensor and Robeson 2008), potentially generating systematic biases in the comparison between WRF and GPCP. However, for the central USA, Gebremichael et al. (2005) showed that GPCP has a very high skill in discriminating rainy from non-rainy days. Besides, the focus of the present study is to compare and assess the respective effects of the various model settings. The KF and Grell Cu schemes show converging results, with a zonal tripole (Congo—eastern Kenya—Indian Ocean) resulting from too few and too dry rainy days compared to GPCP. BMJ confirms its singularity by producing too few rainy days

Cu KF

Cu BMJ Cu Grell

Cu KF

E27

E24 E25 E26

E23

E21 E22

E19 E20

E17 E18

E15 E16

E13 E14

E12

E9 E10 E11

E7 E8

E6

E4 E5

E3

E1 E2

GPCP

0

ERA−I

200

Cu BMJ Cu Grell

over the overall domain, except over the Indian Ocean when combined with the YSU or TKE PBL schemes. The rainy days are however associated with heavy rainfall. Consequently, RMS errors are larger for BMJ than for Grell and for KF. Another noteworthy result is that the contribution of convective rainfall varies strongly from one experiment to another (Fig. 5). It is largest for Grell and virtually null for BMJ. ERA-Interim is more in accordance with KF, with *70–80% of convective rainfall. Given the known features of the East African Long Rains (with the predominant fraction of convective rainfall over the Highlands and both stratiform and convective spells over

123

1364


Table 2 For each experiment in Set #1: spatial correlations between seasonal rainfall amounts recorded at the 44 rain-gauges and the closest WRF grid-points in MAM 1999, and RMS errors (mm) calculated for the same grid-points against the corresponding rain-gauge records Cu KF

Cu BMJ

Cu Grell

PBL

E1

E2

E3

E4

E5

E6

E7

E8

E9

YSU

0.69/219

0.68/232

0.67/222

0.51/325

0.53/306

0.66/256

0.40/362

0.39/361

0.38/345

PBL

E10

E11

E12

E13

E14

E15

E16

E17

E18

TKE

0.61/278

0.64/274

0.66/239

0.58/330

0.58/318

0.63/270

0.54/336

0.54/329

0.57/317

PBL

E19

E20

E21

E22

E23

E24

E25

E26

E27

ACM2

0.71/205

0.68/212

0.64/251

0.60/295

0.66/298

0.60/241

0.42/322

0.48/305

0.43/309

MP Lin

MP WSM6

MP Morrison

MP Lin

MP WSM6

MP Morrison

MP Lin

MP WSM6

MP Morrison

Correlation values reaching 0.6 are underlined Experiment names appear in bold Cu KF

Cu BMJ

Cu Grell

PBL YSU

E1

E2

E3

E4

E5

E6

E7

E8

E9

PBL TKE PBL ACM2

E10

E11

E12

E13

E14

E15

E16

E17

E18

E19

E20

E21

E22

E23

E24

E25

E26

E27

MP Lin

MP WSM6 MP Morrison

−100

MP Lin

MP WSM6 MP Morrison

−50

0

50

MP Lin

MP WSM6 MP Morrison

100

Fig. 6 As Fig. 4a but for cumulated seasonal differences in surface evapotranspiration between the 27 WRF experiments in Set #1 and ERAInterim

the coastal regions, e.g., Pohl and Camberlin 2006), the behavior of BMJ does not seem adequate. Categorized as a deep-layer control convective scheme (Stensrud 2007), BMJ is an adjustment-type scheme which assumes that the large-scale environment supplies Convective Available Potential Energy (CAPE). The first step in this scheme is a determination of cloud base (Lifting Condensation Level, LCL, of the parcel with maximum equivalent potential temperature) and cloud top (the highest model level where the parcel is still buoyant, usually just below Equilibrium Level). If the parcel is not buoyant at any level, convection will not be activated at that point. If the difference between the cloud base and cloud top is less than 200 hPa, the deep convection scheme aborts and instead a shallow convection scheme is used. The scheme structure thus favors activation in cases with significant amounts of moisture at low and mid levels and positive CAPE. As a consequence of this deep-layer control, large CAPE values in the lower layers are not a sufficient

123

condition for atmospheric convection to be activated in BMJ. In contrast, KF is considered as a low-level control convective scheme. Its trigger function is based on checking the parcel for buoyancy at the calculated LCL, starting with the lowest 50 hPa layer and repeating the same procedure up to 600–700 hPa. In the case of a nonbuoyant parcel, shallow convection is permitted, based on turbulent kinetic energy for mass flux rather than CAPE. Initiation of convection occurs when a parcel within a grid column overcomes negative buoyancy in order to rise. Rearrangement of mass then takes place by updraft, downdraft and entrainment calculations until 90% of the CAPE is saved. The Grell-3D scheme has a rather simple approach to convective clouds. Clouds constitute a onedimensional system with downdraft and updraft branches. Mixing between the two branches and between the convective system and the surrounding environment takes place only at the top and bottom of the cloud. A dynamic control and a trigger control are applied as a combination


of 144 member ensembles. Closure assumption is based on CAPE, low-level vertical velocity or moisture convergence for which a quasi-equilibrium is applied for the available buoyant energy (large-scale changes and changes due to convection are almost equal). Trigger mechanism and thresholds permitting convection vary for each member. Convective precipitation is proportional to the integral of the moisture advected by updraft. The total amount of cloud water due to condensation is removed by rainfall leaving no residual. This scheme allows subsidence to the nearby areas, or within the grid-point if resolution is greater than 10 km. Due to average deep subsiding motion in the air column over Equatorial East Africa (e.g., Pohl and Camberlin 2010), especially above 700 hPa, it is hypothesized that convection is inhibited in BMJ. Moisture convergence, which is mostly confined to the lower layers and results from advections from both the Indian Ocean and the Congo Basin, is however favorable to the activation of convective processes in the KF and Grell schemes. Figure 6 shows the difference in surface evapotranspiration between the 27 experiments in Set #1 and ERAInterim. The consistency with rainfall biases is quite good, as the same tripolar structure prevails in most experiments (except for those using the TKE PBL scheme). Evapotranspiration is recurrently too low over the lakes, and this systematic bias seems difficult to avoid (see Sect. 4 below). Over the landmass, the BMJ Cu scheme is associated with lower surface evapotranspiration, whatever the microphysics and PBL schemes. This result is probably the consequence of the too few but too wet rainy days in BMJ, favoring surface runoff instead of infiltration and evaporation. Over the oceanic domain, TKE PBL is clearly associated with lower evapotranspiration, reversing the sign of the differences with ERA-Interim. This particularity over the Indian Ocean explains why the recurrent tripolar structure is less clear in the experiments using this scheme. Unlike the other two PBL schemes, TKE is a local closure PBL scheme, that is, it estimates the unresolved turbulent fluxes at each grid-point from the mean atmospheric variables and/or their gradients at that point (Hu et al. 2010). In contrast, the YSU scheme considers nonlocal fluxes implicitly through a parameterized nonlocal term, while the ACM2 scheme considers them explicitly through a transilient term. Mellor and Yamada (1982) argue that the TKE scheme is appropriate for all stable and slightly unstable flows, but is less adequate as the flow approaches the freeconvection limit. Given the predominance of a strong convective regime over the Indian Ocean, this behavior is likely to affect the results, even at the seasonal scale (Fig. 6). The MP schemes are of less importance. The effects of the Cu and PBL schemes seem to be additive, as noted for rainfall between the cumulus and microphysics schemes.

1365

The largest uncertainties are thus related to the choice of convective scheme, an intuitive result in tropical regions, and already well documented in previous work. The nonnegligible influence of both the PBL and MP schemes is also highlighted, and the respective effects of the various parameterizations seem to be basically additive in nature. Despite large differences between one experiment and another, WRF tends to simulate too dry (wet) conditions over eastern Kenya and Lake Victoria (the eastern Congo Basin and western Indian Ocean). These results are partly in agreement with those obtained by Anyah and Semazzi (2004, 2007). In the following sections, we retain a standard configuration (corresponding to the E20 experiment) and perform sensitivity experiments on radiation and surface schemes.

4 Set #2: Radiation schemes and land-surface models By controlling radiation forcing and heat exchanges between the surface and the atmosphere, radiation schemes and land-surface models (LSM) are likely to have equal influence on the simulated water cycle. Three shortwave (SW) and 2 longwave (LW) radiation schemes, together with 4 LSM, were used in Set #2, giving 24 possible combinations. Sensitivity to these parameters is much less documented than sensitivity to the parameters explored in Set #1. Many studies use the Dudhia SW, RRTM LW and NOAH LSM schemes without documenting the implications of such choices. The schemes chosen for this set of experiments were: •

• •

for LW radiation, the Rapid Radiative Transfer Model (RRTM, Mlawer et al. 1997) and RRTM for global climate applications (RRTMG_LW, Barker et al. 2002). for SW radiation, Dudhia (1989), Goddard (Chou and Suarez 1994) and RRTMG_SW (Iacono et al. 2004). for LSM, the 4-layer NOAH model (Chen and Dudhia 2001), the 2-layer Pleim-Xiu (PX) model (Xiu and Pleim 2001) and the 6-layer Rapid Update Cycle (RUC) model (Smirnova et al. 2000). Sensitivity to the initialization of PX LSM (from analysis: PX #1 hereafter, or from surface moisture availability (MAVAIL): PX #2) is also documented.

Figure 7 shows rainfall biases against GPCP and raingauge data (Fig. 7a) and differences in surface evapotranspiration against ERA-Interim (Fig. 7b) for the 24 experiments in Set #2. Figure 8 shows seasonal rainfall amounts over East Africa. Clear over-estimations of the atmospheric water cycle prevail for most configurations (Figs. 7, 8), and even for all configurations over the edge of the Congo Basin and the Indian Ocean (Fig. 7a). In detail

123

1366

B. Pohl et al.: Testing WRF in East Africa LW RRTM

(a)

LW RRTMG_LW

LSM NOAH LSM PX #0 LSM PX #1 LSM RUC

E20 − 181

E28 − 378

E29 − 337

E30 − 187

E31 − 330

E32 − 303

E33 − 290

E34 − 605

E35 − 515

E36 − 273

E37 − 556

E38 − 491

E39 − 306

E40 − 602

E41 − 520

E42 − 281

E43 − 547

E44 − 483

E45 − 236

E46 − 517

E47 − 438

E48 − 223

E49 − 482

E50 − 420

SW Dudhia

SW Goddard

−400

−300

SW RRTMG_SW

−200

−100

SW Dudhia

0

100

SW Goddard

200

LW RRTM

(b)

300

SW RRTMG_SW

400

LW RRTMG_LW

LSM NOAH LSM PX #0 LSM PX #1

E20

E28

E29

E30

E31

E32

E33

E34

E35

E36

E37

E38

E39

E40

E41

E42

E43

E44

E45

E46

E47

E48

E49

E50

LSM RUC SW Dudhia

SW Goddard

−400

−300

SW RRTMG_SW

−200

−100

Fig. 7 a As Fig. 4(a) but for the 24 WRF experiments in Set #2. Experiments shown on the same row/column/3 9 4 block use the same LSM/SW/LW radiation scheme, respectively. Circles show the

123

SW Dudhia

0

100

SW Goddard

200

300

SW RRTMG_SW

400

biases against the closest rain-gauge station. b As a but for the cumulated seasonal differences in surface evapotranspiration between the 24 WRF experiments in Set #2 and ERA-Interim

B. Pohl et al.: Testing WRF in East Africa Fig. 8 As Fig. 5 but for the 24 WRF experiments in Set #2

1367 LSM NOAH

(a)

LSM PX #0

LSM PX #1

LSM RUC

800 SW Dudhia SW Goddard SW RRTMG_SW

600

400

E50

E48

E49

E47

E46

E45

E44

E43

E41

E42

E40

E39

E38

E37

E35

E36

E34

E33

E32

E31

E29

E30

E28

E20

GPCP

0

ERA−I

200

LW LW LW LW LW LW LW LW RRTM RRTMG_LW RRTM RRTMG_LW RRTM RRTMG_LW RRTM RRTMG_LW LSM NOAH

(b)

LSM PX #0

LSM PX #1

LSM RUC


600 500 400 300 200

E50

E49

E48

E47

E46

E45

E44

E43

E42

E41

E40

E39

E38

E37

E36

E35

E34

E33

E32

E31

E30

E29

E28

E20

GPCP

0

ERA−I

100

LW LW LW LW LW LW LW LW RRTM RRTMG_LW RRTM RRTMG_LW RRTM RRTMG_LW RRTM RRTMG_LW LSM NOAH

(c)

LSM PX #0

LSM PX #1

LSM RUC


1000 800 600 400

E50

E49

E48

E47

E46

E45

E44

E43

E42

E41

E40

E39

E38

E37

E36

E35

E34

E33

E32

E31

E30

E29

E28

E20

GPCP

0

ERA−I

200

LW LW LW LW LW LW LW LW RRTM RRTMG_LW RRTM RRTMG_LW RRTM RRTMG_LW RRTM RRTMG_LW

however, the SW Dudhia scheme and NOAH LSM provide the best (i.e., driest) results over Kenya (including both Lake Victoria and the eastern plains, Fig. 8), as confirmed by the RMS statistics calculated on regional rainfall biases (Fig. 7a). These schemes also act to increase the contribution of convective rainfall to the overall seasonal amount (Fig. 8). Both the Goddard and RRTMG_SW schemes simulate too much rainfall, especially of stratiform nature. PX and RUC LSM are clearly associated with strong rainfall over-estimations over the region. The impact of the LW scheme is much more limited, which could be due to

the strong similarities between the RRTM and RRTMG_LW schemes (RRTMG_LW includes a MonteCarlo Independent Column Approximation method of random cloud overlap, and simplified integration intervals to compute absorption over the various longwave band ranges). The combinations that show the largest rainfall amounts are also those which over-estimate surface evapotranspiration (Fig. 7b). PX LSM clearly simulates the strongest evapotranspiration from the surface. Its simplified vertical profiles of soil moisture (two layers which are respectively

123

1368

1 and 100 cm deep) may accelerate evaporation processes. RUC and more clearly NOAH provide less moisture to the air mass; they use more sophisticated vertical profiles of moisture and temperature, due to more realistic stratifications in their sub-surface layers. Combinations with the SW radiation schemes are also important, because all LSM tend to favor wetter conditions when combined with the Goddard or RRTMG_SW schemes. It is also noticeable that, although evaporation is over-estimated over the oceanic domain, it is strongly under-estimated over the lakes. For instance, cumulated evaporation spreads out between 121 and 208 mm in MAM over Lake Victoria (with an average of 180 mm for all experiments in Set #2), while Yin and Nicholson (1998) obtained values comprised between 399 and 446 mm based respectively on observations and remote sensing estimates. Additional experiments (not shown) revealed that similar results are obtained for all LSM, even the intermediate complexity thermal diffusion scheme. This bias is only corrected when no LSM is used, but to the detriment of the remaining land areas. These results are surprising, since lakes are considered as quasi-ocean points in all LSM, and present basic surface properties (albedo, surface roughness and emissivity) that are identical to the ocean grid-points (Sect. 5). Such systematic biases could be due to enhanced (reduced) day-time (night-time) divergence (convergence) over the lakes, in association with land/lake breezes, although further investigation of the diurnal cycle is needed to assess this issue. Recent advances in WRF include a coupling with a lake model, which could be useful to correct these systematic biases. Such a coupling was successfully implemented over the region by Moore et al. (2010), who used a one-dimensional lake model to improve the water budget over Lake Victoria in RAMS. The largest uncertainties in the simulation of the water cycle appear to be associated with the SW and LSM schemes. They seem to be even larger over Kenya than those related to convection, PBL and MP parameterizations (Figs. 5, 8), although the additivity between the parameters tested in Set #1 and Set #2 could not be documented here, because of the huge number of experiments required. This result is of importance, since most studies only focus on the effects of convective and/or PBL schemes and choose a priori their radiation and land-surface schemes without quantifying associated uncertainties. Those retained in most papers, namely Dudhia SW and NOAH LSM, constitute a combination that favors drier conditions and a weaker water cycle compared to other (and even more sophisticated) radiation and land-surface schemes. An extensive analysis of the different terms of the radiative budget is needed to fully document the reasons for such changes at the seasonal timescale, which is outside the scope of the present work.

123


5 Set #3: Land surface conditions and lateral forcing fields Sets #1 and 2 illustrated how the choice of the model physical package could influence simulated regional climate. In addition to these settings, the RCM solution is also known to strongly depend upon the lateral forcing fields (Denis et al. 2003, Diaconescu et al. 2007) and surface conditions (Liang et al. 2005). In Set #3, we performed additional experiments using the modified USGS Global Land Cover Characteristics dataset, compiled in 1992–1993 using the Advanced Very High Resolution Radiometer (AVHRR) sensor and containing 24 land-use categories (Anderson et al. 1976, Hitt 1994), and the newer IGBPMODIS land cover database, defined in 2001–2002 and partitioned into 20 classes (Friedl et al. 2002). We also explored the effects of lateral conditions by forcing experiments with ERA40 (Uppala et al. 2005) and ERA-Interim (Simmons et al. 2007) reanalyses. The main differences between the two generations of European reanalyses include improvements to the assimilation scheme, an update of the physical package of the IFS GCM and a better horizontal resolution (Simmons et al. 2007). Set #3 therefore comprises 4 experiments, obtained by combining the two lateral forcing datasets and the two land-use classifications. In experiment names, the letter M indicates the use of the MODIS land-use database (while experiments coded E correspond to the default configuration using USGS, already presented in previous sections) and the prefix 4 denotes experiments forced laterally by ERA40. The effects of land-use categories (M minus E experiments) on the seasonal mean climate are far from negligible (Fig. 9), as revealed by experiments driven laterally by ERA40 or ERA-Interim. MODIS land-use generally favors rainier conditions, especially over the northwestern part of the domain, while slightly drier conditions are found in the southeastern part (Fig. 9a–b, left panels). Locally, the differences reach up to 250 mm during the season (the observed seasonal amount being comprised between 200 and 600 mm according to GPCP, Fig. 1). MODIS also favors slightly wetter conditions over the East African lakes, especially Lake Victoria, correcting some of the dry biases noted in the previous sets of experiments. However, over the lakes and the ocean, surface evapotranspiration is barely modified (Fig. 9a–b, left panels). Over the landmass, it displays a dipole-like structure, with MODIS experiments providing larger (lower) evapotranspiration values in the northeastern (southwestern) parts of the domain. Figure 10 aims at determining the land-use categories responsible for such changes. Given that the 24 categories in the USGS database are different from the 20 categories derived from MODIS, a one-to-one comparison is not easy.


(a)

P [M20 − E20]

(f)

E [M20−ERA-I]

−200

−100

0

E [4M20−ERA40] 6N 3N 0 3S 6S

30E 35E 40E 45E

30E 35E 40E 45E

30E 35E 40E 45E

P [4M20−GPCP] − 237

6N 3N 0 3S 6S

6N 3N 0 3S 6S 30E 35E 40E 45E

E [4E20−ERA40] 6N 3N 0 3S 6S

30E 35E 40E 45E

30E 35E 40E 45E

P [M20−GPCP] − 260

6N 3N 0 3S 6S

30E 35E 40E 45E

P [4E20−GPCP] − 169

6N 3N 0 3S 6S

6N 3N 0 3S 6S 30E 35E 40E 45E

E [4M20 − 4E20] 6N 3N 0 3S 6S

30E 35E 40E 45E

(d)

E [E20−ERA-I]

P [E20−GPCP] − 181

6N 3N 0 3S 6S

P [4M20 − 4E20]

6N 3N 0 3S 6S 30E 35E 40E 45E

30E 35E 40E 45E

(e)

(b)

E [M20 − E20] 6N 3N 0 3S 6S

6N 3N 0 3S 6S

(c)

1369

100

30E 35E 40E 45E

200

Fig. 9 a Effect of MODIS land-use on MAM seasonal rainfall (mm, left panel) and surface evapotranspiration (mm, right panel) amounts (EM20–E20). See text for details. b As a but for (4EM20–4E20). c (Left panel) Shadings: seasonal rainfall amount biases (mm) of E20 against GPCP. Seasonal RMS error computed over the Kenya rainfall

index is labeled on the figure. Circles: biases against the closest raingauge station. (Right panel) Difference in surface evapotranspiration between E20 and ERA-Interim. d–e–f As c but for 4E20, EM20, 4EM20 experiments, respectively. For all panels, significance is tested and shown as for Fig. 4

Figure 10a and c therefore present the dominant categories for all grid-points and Fig. 10b and d their spatial representativity (i.e., the part of the grid-point area that is affiliated to the dominant categories). The most homogeneous grids are logically water bodies (lakes and ocean) but also the evergreen forest in the Congo Basin. Over East Africa, land-use is much more complex at the sub-grid scale but also at the resolution of the 36 km grid. Though both databases are available at the same resolution (30 arc s), MODIS tends to provide spatially noisier patterns than USGS, explaining the lower representativity of the dominant categories over large parts of East Africa. This could be due to the finer discrimination of vegetation types allowed by the 15 channels of the MODIS sensor that are dedicated to land-use and vegetation. USGS classification, for purposes of comparison, was based on the 5 channels of the more polyvalent AVHRR sensor. Figure 10e provides a contingency table of the grid-points that fall into each dominant land-use category. Water gridpoints are the easiest to classify, and USGS category #9 (U9) almost perfectly fits MODIS category #12 (M12). This is also the case for U4 and M5 (shrublands and open shrublands, respectively) but it is far less clear for all other categories. For instance U5 (savannas) is mainly distributed within M6 (woody savannas) and M7 (savannas), but also M1 and M8 (evergreen broadleaf forest and grasslands). Symmetrically,

M6 and M8 (M7) are dispatched into 5 (6) USGS categories, respectively. We chose to retain the {USGS—MODIS} couples that concern at least 50 grid-points. This limitation is not as drastic as it seems since the corresponding 14 couples regroup 85% of the grid-points of the domain. For each couple, the differences (MODIS minus USGS) in terms of seasonal rainfall amounts are shown in Fig. 10f–g. The same analysis was applied to surface evapotranspiration (not shown). USGS land-use is confirmed to favor generally drier conditions, whatever the reanalysis used to drive regional simulations. This is particularly true for U5, associated with sensibly lower rainfall amounts than the corresponding M1, M6, M7 and M8 categories. This remark can be generalized to all other couples, including those representing under 50 gridpoints (‘‘Other’’ category, but without distinction of all possible couples). A noticeable exception concerns water land-use (U9M12) for which differences of both signs are as probable. The reasons for such changes are assessed in Fig. 11, which presents surface albedo, green fraction and leaf area index (LAI) reconstructed from the USGS and MODIS land-use categories, as well as the differences between the two datasets. Similar analyses were also applied to surface emissivity and surface skin temperature calculated by the LSM (not shown). Over the northwestern part of the

123

1370


(b)

(a) 11 10 9 8 7 6 5 4 3 2 1

6N 3N 0 3S 6S 30E

(c)

35E

40E

45E 12 11 10 9 8 7 6 5 4 3 2 1

3N 0 3S 6S 40E

Water [1238] Barren/Sparsely Veget. [36] Cropland/Nat. Veget. Mos.[21] Croplands [48] Grasslands [277] Savannas [729] Woody Savannas [496] Open Shrubland [584] Closed Shrublands [30] Mixed Forests [2] Deciduous Broadleaf For. [8] Evergreen Broadleaf For. [458]

(f)

11 10 9 8 7

3S 6S 30E

35E

40E

45E

30E

35E

40E

45E

6N 3N 0 3S 6S

50

60

70

80

90

100

1000 500 0 −500

−1000

300

400

500

Seasonal Rainfall

U9M12 [1216]

200

9 10 11 12

U7M1 [274]

6 7 8 MODIS

U6M7 [70]

5

U6M6 [90]

100

4

U5M8 [67]

3

U5M7 [313]

2

U5M6 [259]

1

U5M1 [56]

500

1

U4M8 [103]

2

U4M5 [529]

(g) 1000

U2M7 [106]

3

U2M6 [54]

4

U1M7 [134]

5

U1M1 [53]

6

Other [603]

USGS

0

40

(e)

0

3N

45E

Seasonal Rainfall

35E

6N

(d)

6N

30E

Barren/Sparsely Veget. [45] Herbaceous Wetland [1] Water Bodies [1229] Mixed Forest [36] Evergreen Broadleaf [335] Deciduous Broadleaf For. [173] Savanna [727] Shrubland [738] Grassland [77] Cropland/Woodland Mos. [273] Dryland Cropland Pasture [293]

0 −500

−1000

Fig. 10 a Dominant USGS land-use category in each grid-point. b Spatial representativity (%) of the dominant USGS land-use category within each grid-point. c–d as a–b but for MODIS land-use categories. e Contingency between the dominant land-use categories according to USGS and MODIS. f Box-and-whisker diagram of the seasonal rainfall differences between experiments EM20 and E20 for

the main combinations of USGS and MODIS dominant land-use categories. Boxes extend from the 25th to the 75th percentile, with a horizontal red bar showing the median value. Whiskers extend to the most extreme datapoints. Red plus signs correspond to values that are considered as outliers. g As f but for experiments 4EM20 and 4E20

domain, predominantly covered by savanna, the green fraction is considerably larger in MODIS, which is consistent with the wetter conditions there (Fig. 9a–b). This contributes also to the wetter conditions found in MODIS in the savanna category (Fig. 10f–g). Surface albedo differences are positive in the northwest. They cancel the effects of the larger green fraction, resulting in slight differences in terms of surface evapotranspiration (Fig. 9a). Over western Tanzania and southeastern Congo, albedo differences are larger and explain the lower evapotranspiration found in MODIS experiments. The largest differences in terms of LAI are found at the periphery of the evergreen forest of the

Congo Basin (Fig. 10a–c), due to relative disagreement concerning its spatial extension in the two land-use classifications. Taken together, these three parameters explain most of the differences in the terms of the water budget equation (Fig. 9a–b). Does land-use categorization improve or worsen the simulated water cycle over the region? Figures 9c–f show rainfall biases against GPCP and evapotranspiration differences to the corresponding reanalyses. As far as rainfall is concerned, the familiar tripolar structure opposing the Congo Basin and the Indian Ocean on the one hand, to East Africa on the other hand, is prevalent in all four experiments.

123

B. Pohl et al.: Testing WRF in East Africa Fig. 11 Surface albedo (upper row), green fraction (central row) and leaf area index (lower row) derived from USGS (lefthand column), MODIS (central column), and difference between MODIS and USGS (right-hand column)

1371 ALBEDO USGS

ALBEDO [MODIS − USGS]

ALBEDO MODIS

6N

6N

6N

3N

3N

3N

0

0

0

3S

3S

3S

6S

6S

6S

30E

35E

40E

0

30E

45E 0.1

0.2

35E

0.3

GREEN FRACTION USGS

40E

30E

45E −0.2

0.4

GREEN FRACTION MODIS 6N

3N

3N

0

0

0

3S

3S

3S

6S

6S

6S

35E

40E

0

45E 0.25

30E 0.5

35E

0.75

40E

45E

30E −0.5

1

6N

6N

3N

3N

3N

0

0

0

3S

3S

3S

6S

6S

6S

0

40E 1

45E 2

In detail however, MODIS tends to produce more uniform wet biases over the domain. Due to this generalized rainfall over-estimation, RMS statistics computed over the Kenya rainfall index are 40% larger. Evapotranspiration is also affected, and the negative difference over the Horn of Africa is over-corrected in the experiments using MODIS. Over East Africa, prescribing MODIS land-use categories thus results in enhancing the atmospheric water cycle, which is not particularly adequate since it was already too intense in previous runs. Experiments driven by ERA40 are generally drier over Kenya (leading to slightly reduced RMS error values compared to experiments forced by ERA-Interim, Fig. 9e–f). Surface evapotranspiration is barely modified, which suggests the possible implication of regional-scale moisture fluxes and associated convergence. Such results are sensibly different from those obtained by Moore et al. (2010), who found a much more limited influence of LAI and vegetation fractional cover (both derived from MODIS) on East African rainfall using RAMS. Two major differences between our protocol and theirs could explain such disagreement: (1) Moore et al. prescribe an annual cycle in their LAI and vegetation fractional cover, while our fields are static, which is clearly less adequate to simulate realistic seasonality; (2) Moore et al. do not prescribe updated albedo fields, while in our study, the albedo fields differ between USGS and MODIS.

30E 3

0.1

0.2

4

35E 5

40E 6

35E

−0.25

40E 0

45E 0.25

0.5

LAI [MODIS − USGS]

LAI MODIS

6N

35E

0

45E

3N

LAI USGS

30E

−0.1

40E

GREEN FRACTION [MODIS − USGS] 6N

6N

30E

35E

45E

30E −2

−1

35E

40E 0

45E 1

2

Figures 12 and 13 show the differences in seasonal rainfall and moisture fluxes between ERA-Interim and ERA40, and the corresponding WRF experiments. ERAInterim is too wet over the Congo Basin and the Indian Ocean (Fig. 12b). In contrast, ERA40 is much drier over Africa (Fig. 12c), but far too wet over the Indian Ocean. The biases over the Indian Basin are partly (but not fully) corrected in ERA-Interim, while the biases over Africa are over-corrected (Fig. 12a). There are striking similarities between WRF (Fig. 9c) and ERA-Interim (Fig. 12b) in the spatial patterns of the rainfall biases. However, such similarities are not found between WRF (Fig. 9d) and ERA40 (Fig. 12c), in both the 4E20 and 4M20 experiments. Even when forced by ERA40, WRF simulates the tripolar bias structure, although it does not prevail in the forcing model. Consistently, WRF tends to be generally drier (wetter) over East Africa (the Indian Ocean) when forced by ERA40, than when forced by ERA-Interim (Figs. 9c–d, 12d–e). Figure 13 shows the vertically integrated seasonal mean moisture flux and convergence fields in the reanalyses, as well as the differences between ERA-Interim and ERA40 on the one hand, and the corresponding WRF experiments on the other hand. The dominant easterly fluxes (Fig. 13a– b) are weaker over tropical Africa in ERA-Interim (Fig. 13c). They are associated with stronger convergence over the Congo Basin and localized areas along the coast,

123

1372


(a)

(b)

[ERA−I − ERA40]

(c)

[ERA−I − GPCP]

24N

24N

24N

12N

12N

12N

0

0

0

12S

12S

12S

24S

24S

24S

12E

24E

36E

48E

60E

12E −400

(d)

24E

−200

36E 0

6N

6N

3N

3N

0

0

3S

3S

6S

6S 35E

40E

45E

60E

12E

200

(e)

[E20 − 4E20]

30E

48E

[ERA40 − GPCP]

24E

36E

48E

60E

400

[EM20 − 4M20]

30E

35E

40E

45E

Fig. 12 a Differences in MAM seasonal rainfall amounts (mm) between ERA-Interim and ERA40. b Seasonal biases of ERA-Interim against GPCP. c As b but for ERA40. d Difference between E20 and

4E20 experiments. e Difference between EM20 and 4EM20 experiments. See text for details. For all panels, significance is tested and shown as for Fig. 4

which may explain the higher precipitation when compared to ERA40. At the higher latitudes in both hemispheres and over ocean basins, divergence is however prevalent. In general agreement with the forcing fields, easterly fluxes are also weaker in the WRF experiments forced by ERAInterim. This is particularly true in the southern part of the domain, while the largest differences between the two generations of reanalyses are found in the northern part. Although the pattern is noisy, moisture convergence for downscaled ERA-Interim, when compared to ERA40, is also larger over the landmass as well as over parts of the nearby Indian Ocean. This is in general agreement with the slightly higher rainfall noted over many parts of East Africa (Fig. 12d–e) in the WRF experiments forced by ERA-Interim. Even if the differences between the two reanalyses and the differences between the corresponding WRF experiments show some spatial similarities (consisting of wetter conditions over the continent and drier conditions over the ocean for ERA-Interim), the amplitude of such differences is much larger for the reanalyses. This demonstrates (1) that the rainfall biases obtained by WRF are only marginally dependent on lateral boundary conditions, and primarily relate to the model physics (Sects. 3–4); (2) that a RCM used to downscale the outputs of two distinct GCM can be useful to reduce their differences and favors their

convergence towards similar solutions. This series of experiments also demonstrated that regional rainfall biases are not (only) related to surface conditions, as they persist whatever the land-use category.

123

6 Set #4: Domain size and vertical resolution Though the model physics strongly influences the simulated water cycle over East Africa, rainfall biases show recurrent spatial patterns of unidentified origin. In Set #4, we propose to explore the effects of domain size and vertical resolution. The reduced size of the domain enabled us to design sensitivity experiments for a broad variety of settings. It is nevertheless likely to induce numerical artifacts if it is smaller in size than the so-called ‘‘spatial spin-up’’, corresponding to the characteristic distance that the large-scale flow needs to travel before developing small-scale features (Leduc and Laprise 2009). The spatial spin-up depends on the modulus of the fluxes that penetrate the regional domain, which is far lower in equatorial latitudes than, e.g., in mid-latitudes. This justified the choice of a spatially limited standard domain. However, in order to discuss this point and quantify associated uncertainties, we conducted a series of experiments using enlarged domains (Fig. 2a)

B. Pohl et al.: Testing WRF in East Africa Fig. 13 a Vectors: MAM seasonal mean moisture fluxes (m kg s-1 kg-1) vertically integrated between the surface and 200 hPa according to ERAInterim. Shadings: associated moisture convergence (mm). The black box represents the standard domain. b As a but for ERA40. c Differences between ERA-Interim and ERA40. Differences in the moisture fluxes (convergence) that are not significant at the 95% level according to a t test applied to daily averages are not represented (shaded white). d As c but differences between E20 and 4E20 experiments. e As c but differences between EM20 and 4EM20 experiments

1373

(a) ERA−I

(b) ERA40

24N

24N

12N

12N

0

0

12S

12S

24S

24S 12E

24E

36E

48E

−2000

60E

−1000

12E 0

24E 1000

36E

48E

60E

2000

(d) [E20 − 4E20]

(c) [ERA−I − ERA40] 6N 24N

3N

12N

3S

0 6S 30E

0

35E

40E

45E

(e) [EM20 − 4M20] 6N 3N

12S

0 3S 24S

6S 12E

24E −500

consisting of adding 15 and 30 grid-points (approximately 5 and 10) at all boundaries to the standard domain. They are referred to as domains L and XL, respectively; only central grid-points, which are common (i.e., strictly identical) to the three domains, are compared. The number of vertical levels is also known to be a key parameter which potentially impacts the simulated climate (e.g., Kimoto et al. 2005, Wakazuki et al. 2007). We addressed this issue by performing experiments with 28 or 35 vertical levels (using the default distributions provided with the WRF package). The runs using 35 levels are given the suffix V. In Set #4, we thus consider 6 experiments (obtained from the combinations between the three domain sizes and the two vertical resolutions). Figure 14 presents the effects of the domain size on the atmospheric water cycle. Larger domains simulate drier conditions in the western part of the domain and wetter ones in the centre and the eastern parts (Fig. 14a). Though the wet (dry) biases over the Congo Basin (East Africa) are

36E

48E

60E 0

30E

35E

40E

45E

500

partly corrected, leading to more realistic spatial patterns, larger domains generally favor larger rainfall amounts over East Africa. They are thus accompanied by a slight increase in RMS statistics over the Kenya index (not shown). Over the Indian Ocean, the wet biases increase more strongly. All these changes in the seasonal rainfall at the regional scale affect both the number of rainy days (Fig. 14b) and their average intensity (Fig. 14c), thus reducing the biases noted for E20 experiment on Fig. 4 (except for the Indian Ocean). Importantly, the spatial structure of the differences between the largest domains and the standard domain is qualitatively similar, but the differences are generally of larger amplitude for domain XL. This can be interpreted as weaker control by large-scale variability, imposed through the lateral forcing field, on the simulated regional climate (Leduc and Laprise 2009). In other words, the solution of experiment E20 is strongly controlled by ERA-Interim lateral forcing. This is less true for experiment L20 and,

123

1374 Fig. 14 Effect of domain size on MAM seasonal rainfall amounts. a Difference in seasonal rainfall amount (mm) between L20 and E20 (left panel) and XL20 and E20 (right panel). b–c–d As a but for the number of rainy days ([1 mm), the average intensity of rainy days (mm), and surface evapotranspiration (mm), respectively. e Vectors: difference between the same experiments in the vertically integrated moisture fluxes (m kg s-1 kg-1). Shadings: difference in associated moisture convergence (mm). For all panels, significance is tested and shown as for Fig. 13


(a)

[XL20−E20]

[L20−E20]

6N

6N

3N

3N

0

0

3S

3S

6S

6S

100 0 −100

30E

(b)

200

35E

40E

45E

30E 35E [XL20−E20]

[L20−E20]

40E

45E

−200 15

6N

6N

3N

3N

0

0

3S

3S

−5

6S

−10

6S 30E

(c)

35E

40E

45E

10 5 0

30E

35E

40E

45E

[XL20−E20]

[L20−E20]

6N

6N

3N

3N

0

0

3S

3S

6S

6S

−15 10 5 0 −5

30E

(d)

35E

40E

−10

45E

30E

35E

40E

45E

[XL20−E20]

[L20−E20]

6N

6N

3N

3N

0

0

3S

3S

6S

6S

200 100 0 −100

30E [L20−E20]

(e)

35E

40E

−200

45E

30E 35E [XL20−E20]

6N

6N

3N

3N

0

0

3S

3S

6S

6S

40E

45E 200 100 0 −100

30E

35E

even more clearly, for XL20. The two experiments using the largest domains thus show how WRF simulates its own climate, depending on the prescribed surface conditions (Set #3) and physical package (Sets #1 and 2), but without strong interference from the lateral forcing. Changing the domain size also induces modifications in surface evapotranspiration (Fig. 14d). The eastern Congo Basin is barely affected, but changes are more pronounced

123

40E

45E

−200 30E

35E

40E

45E

over East Africa (E is enhanced in L20 and XL20) and the Indian Ocean (E generally decreases). Over these regions, the differences noted between E20 and ERA-Interim (Fig. 6) are thus reduced. The fact that changes in P and E do not show similar spatial patterns implies that moisture convergence is also modified, as confirmed by Fig. 14e. Larger domains favor moisture convergence over the Indian Ocean and neighboring East Africa, at least during


1375 [L20V − L20]

[E20V − E20]

(a)

6N

6N

3N

3N

3N

0

0

0

3S

3S

3S

6S

6S

6S

35E

40E

30E

45E

−200

(b)

Levels (hPa) 30E

35E

40E

45E

40E

45E

−100 0 [L20V − L20]

[E20V − E20]

100 200 300 400 500 600 700 850 1000

35E

30E

100

100 200 300 400 500 600 700 850 1000 30E

Longitudes

35E

40E

−0.5

0

40E

45E

40E

45E

100 200 300 400 500 600 700 850 1000

45E

30E

Longitudes −1

35E

200 [XL20V − XL20] Levels (hPa)

30E

Levels (hPa)

[XL20V − XL20]

6N

35E

Longitudes 0.5

1

Fig. 15 a Difference in the MAM seasonal rainfall amounts (mm) between E20V and E20 (left panel), L20V and L20 (middle panel), XL20V and XL20 (right panel). See text for details. Significance is tested and shown as for Fig. 4. b Vertical cross-sections of the

differences between the same pairs of experiments in air temperature (K, shadings) and specific humidity (kg kg-1, contours) averaged between the equator and 1S. Differences tested and represented as for Fig. 3

our period of analysis. Further inland, the effects are smaller, although moisture convergence is generally reduced in the western part of the domain. The main differences in the moisture fluxes affect once again the Indian Ocean, where easterly flows are weakened in L20 and XL20 (see Fig. 13d). This corresponds to an increase in the differences between WRF and ERA-Interim. Beside possible improvements in the model biases, Fig. 14 clearly demonstrates that the capability of a RCM to simulate the atmospheric water cycle is strongly ruled by domain size and the location of its boundaries. Similar analyses applied during another year, under different background climate conditions (e.g., El Ninõ or La Ninã years), might however provide different results. The effects of vertical resolution are noisier spatially and more difficult to interpret (Fig. 15a). They do not present similar structures for the three domain sizes. Rainfall amounts are mainly modified over the ocean but without clear and coherent spatial structures. Over the continent, the amplitude of the changes is smaller, particularly north of the equator. Changes in the other terms of the water budget are barely affected and equally difficult to interpret (not shown). Hence, it is not clear whether a larger number of vertical levels offers significant added value for the simulation of the atmospheric water cycle in the region. Increasing vertical resolution does significantly improve the vertical profiles of air temperature (Fig. 15b). Experiments using 35 levels provide vertical profiles that are more in accordance with ERA-Interim. Humidity is not

dramatically modified and, consistently with Fig. 3, the air mass thus remains too dry over East Africa below 600 hPa. The strong temperature biases found in the lower, middle and upper troposphere are all reduced. The strong positive bias at 100 hPa is reduced by 30% and the bias at 600 hPa is fully corrected. Experiments in Set #4 illustrated the strong sensitivity to domain size and the importance of the ‘‘spatial spin-up’’ to let the model adjust to lateral boundary conditions and generate a physically coherent climate. The role of vertical resolution is less clear. The horizontal resolution of the forcing model and, perhaps more importantly, that of the RCM itself could also influence the results.

7 Prioritization of uncertainties Figure 16 summarizes the uncertainties associated with all sensitivity experiments considered in this work, for a regional rainfall index corresponding to northern Tanzania, eastern Uganda, Lake Victoria and most of Kenya (31E– 41E, 4S–3N: domain N in Fig. 2), independently for the four sets of experiments. Box-and-whisker plots show the inter-experiment spread in terms of seasonal rainfall, when a given parameter remains constant. Figure 16 seeks to estimate the specific effect of a given scheme or parameter, and documents the effects of its combinations with other parameters.

123

1376


400 200

−200

ALL [58] YSU [9] TKE [9] ACM2 [9] K-F [9] BMJ [9] Grell [9] Lin [9] WSM6 [9] Morrison [9] RRTM [12] RRTMG_LW [12] Dudhia [8] Goddard [8] RRTMG_SW [8] NOAH [6] PX #0 [6] PX #1 [6] RUC [6] MODIS [2] USGS [2] ERA-I [2] ERA40 [2] E [2] L [2] XL [2] 28 lev [3] 35 lev [3]

[291.9] 0

PBL

Cu

MP

LW

Set #1

SW Set #2

LSM

Land GCM Set #3

Dom

Lev

Set #4

Fig. 16 Summary of the MAM seasonal rainfall amount biases and uncertainties averaged spatially over the Kenya rainfall index, against GPCP, and for the four sets of WRF experiments. GPCP seasonal rainfall amount is labeled on the figure and considered as the origin of the y-axis. The left-hand box-and-whisker shows the spread of the

overall 58 experiments. For each set of experiments, the box-andwhiskers indicate the spread of the experiments using the corresponding model setting (i.e., physical parameterizations, land-use, forcing fields, domain size and vertical resolution). Standard settings for other sets are highlighted in red

The first plot from the left corresponds to the overall spread of the 58 experiments. Over the region, the 1999 Long Rains recorded a seasonal rainfall amount of 292 mm. WRF simulates amounts between 117 and 777 mm, i.e., 40 to 266% of the observed amount. The remaining box-and-whisker plots attribute these huge uncertainties to the relevant scheme or setting. The second plot corresponds to the spread of the experiments in Set #1 that all use the YSU PBL scheme. These nine experiments correspond to all possible combinations between the 3 cumulus and 3 microphysics schemes. All acronyms were defined in previous sections. As far as physical parameterizations are concerned (Sets #1 and 2), SW radiation schemes and LSM clearly contribute to the largest spread. The Dudhia SW scheme unambiguously favors drier conditions than Goddard and RRTMG_SW. Similarly, the NOAH LSM is associated with the weakest water cycle, while the RUC and PX schemes favor much heavier rainfall. These results are of great importance, because most studies neglect to test the sensitivity of their results to these parameterizations. Longwave schemes, in contrast, are not responsible for large changes in the seasonal rainfall amounts. Although they induce non-negligible modifications in simulated rainfall, cumulus schemes contribute less strongly to overall uncertainties. Among the tested schemes, KF produces the largest rainfall amounts. BMJ and Grell are clearly drier. These seasonal values conceal the singular behavior of the deep-layer controlled BMJ scheme, consisting of virtually no convective rainfall, and too few but too wet rainy days. Grell and KF, on the contrary, are

mostly controlled by lower-layer thermodynamics and produce too many but too dry rainy days. MP schemes account for lower uncertainties. Morrison is wetter than WSM6 which is wetter than the Lin et al. scheme. Planetary boundary layer (PBL) schemes provide more convergent results. The effects of boundary conditions (land-use surface categories, lateral forcing reanalysis) are quantified in Set #3. MODIS land-use generally enhances the atmospheric water cycle, leading to larger rainfall amounts. Spatially averaged over Kenya, the differences between the two databases are relatively modest (120 mm on average). However, larger differences were found in other parts of East Africa. They were mostly attributed to strong differences in terms of leaf area index and green fraction over southern Sudan and the eastern Central African Republic, as well as non-negligible modifications of surface albedo derived from MODIS over northern Tanzania. Due to stronger moisture convergence, ERA-Interim provides slightly wetter conditions than ERA40. Associated uncertainties are two to three times lower than those generated by surface conditions (40–50 mm). The effects of domain geometry (size and number of vertical levels) are detailed in Set #4. Enlarging the domain was found to correct parts of the wet (dry) biases over the Congo Basin (central and eastern Kenya). Over East Africa, the dominant effects are an intensification of the water cycle, inducing larger rainfall amounts (Fig. 16). This is particularly true for the largest domain (XL), suggesting that domains E and L are more strongly controlled by the forcing GCM. Domain XL is thus hypothesized to

123


reflect the natural WRF solution, in the absence of strong interference from the lateral forcings allowed by a sufficient ‘‘spatial spin-up’’ (Leduc and Laprise 2009). The role of vertical resolution is more difficult to analyze, because it is spatially very noisy. Averaging this effect over a regional index confirms that the amplitude of the changes is moderate.

8 Discussion In this study we document the capability of a current stateof-the-art regional climate model (RCM) to simulate the three main terms of the atmospheric water budget (i.e., precipitation, evapotranspiration, moisture convergence), over a domain encompassing Equatorial East Africa (Kenya and Uganda, northern Tanzania, eastern D. R. Congo, southern Sudan, Ethiopia and Somalia) for 1999, as it is representative of the climatology. These choices were made because (1) much of the region experiences semi-arid conditions; as water is the main limiting factor for agricultural yields, the analysis of rainfall is fundamental for the economy and the well-being of the local populations; (2) surface topography is complex, suggesting that RCM may improve greatly upon low-resolution GCM outputs; (3) the capability of current state-of-the-art RCM to simulate the regional climate has attracted only a limited number of studies to date. All results presented in this study concern the Long Rains season alone. Qualitatively similar results were obtained for the Short Rains (not shown), suggesting that our results are, to some extent, reproducible. Separating all sources of uncertainties in a RCM and quantifying their respective effects on simulated climate is not an easy task, given the large number of parameters to analyze. In the literature, convective and planetary boundary layer schemes have often been tested. In this study, we also quantified the influence of other physical parameterization schemes (such as land-surface schemes and landsurface models), the forcing GCM, land-use categories, as well as domain size and the number of vertical levels. Some of these settings are also of primary importance (especially shortwave radiation and the land-surface model), although they have been the object of far fewer studies. We found that the atmospheric water cycle showed realistic behavior in some of the tested configurations. Most of the parameters tested nonetheless modified the simulated climate in depth, and the uncertainties associated with the model physics or domain geometry (size and vertical resolution) appeared much larger than average biases. Hence, performing sensitivity experiments using a large variety of settings and parameterization schemes seems highly advisable. Among the configurations tested, the default settings (marked in red on Fig. 16) provide rather satisfactory

1377

results, with a spatially averaged simulated rainfall amount of 299 mm over central East Africa (i.e., an error of only 2%). However, biases vary strongly across the region (Fig. 4a). The most recurrent biases, only slightly modified by the physical package, include wet biases over the eastern Congo Basin and the western equatorial Indian Ocean, separated by contrasting dry biases over the eastern slopes and lower plains of Kenya. Roughly consistent results were obtained by Anyah and Semazzi (2004). These biases were however partly corrected when enlarging the domain, thus suggesting that the natural model solution, in the absence of strong interference with lateral boundary conditions, is less clearly affected. In spite of the predominant importance of the model physics and domain size, other key parameters are known to strongly affect the simulated climate. The forcing GCM and RCM resolutions are probably among the most important (e.g., Denis et al. 2003, Dimitrijevic and Laprise 2005). The size of the buffer zone used to relax the RCM towards its GCM-based forcing fields (Zhong et al. 2010) has also been shown to have considerable influence on results. Once these settings have been established, it should become possible to analyze whether the simulated water cycle is reproducible, by computing ensemble simulations initialized with perturbed atmospheric conditions. Intraseasonal variability could then be documented, which was outside the scope of the present study. Acknowledgments This work is part of the French ANR PICREVAT project. WRF was provided by the University Corporation for Atmospheric Research website (for more information see http://www.mmm.ucar.edu/wrf/users/download/get_source.html). ERA-Interim and ERA40 data were provided by the ECMWF Meteorological Archival and Retrieval System (MARS). Rain-gauge records were provided by the Kenyan and Tanzanian meteorological services. The authors are grateful to Sivarajan Sijikumar, Thierry Castel and Pascal Roucou for helpful discussions on WRF, two anonymous reviewers for their constructive comments and suggestions, and Carmela Chateau for editing the manuscript. Calculations were performed using HPC resources from DSI-CCUB (universite´ de Bourgogne).

References Alexandru A, De Elia R, Laprise R (2007) Internal variability in regional climate downscaling at the seasonal scale. Mon Weather Rev 135:3221–3238 Anderson JR, Hardy EE, Roach JT, Witmer RE (1976) A land-use and land cover classification system for use with remote sensor data. US geological survey professional paper 964, 28 pp Anyah RO, Semazzi FHM (2004) Simulation of the sensitivity of Lake Victoria Basin climate to lake surface temperature. Theor Appl Climatol 79:55–69. doi:10.1007/s00704-004-0057-4 Anyah RO, Semazzi FHM (2007) Variability of East African rainfall based on multiyear RegCM3 simulations. Int J Climatol 27:357–371

123

1378 Anyah RO, Semazzi FHM, Xie L (2006) Simulated physical mechanisms associated with multiscale climate variability over Lake Victoria Basin in East Africa. Mon Weather Rev 134:3588–3609 Barker HW, Pincus R, Morcrette J-J (2002) The Monte-Carlo independent column approximation: application within largescale models. Proceedings of the GCSS/ARM workshop on the representation of cloud systems in large-scale models, May 2002, Kananaskis, Alberta, Canada, 10 pp Betts AK, Miller MJ (1986) A new convective adjustment scheme. Part II: single column tests using GATE wave, BOMEX, ATEX and arctic air-mass data sets. QJR Meteorol Soc 112:693–709 Bright D, Mullen S (2002) The sensitivity of the numerical simulation of the southwest monsoon boundary layer to the choice of PBL turbulence parameterization in MM5. Weather Forecast 17:99–114 Chen F, Dudhia J (2001) Coupling an advanced land-surface/ hydrology model with the Penn State/NCAR MM5 modeling system. Part I: model description and implementation. Mon Weather Rev 129:569–585 Chou MD, Suarez MJ (1994) An efficient thermal infrared radiation parameterization for use in general circulation models. NASA Tech. Memo. 104606, vol 3, 85 pp ¨ nol B (2009) Customization Davis N, Bowden J, Semazzi F, Xie L, O of RegCM3 regional climate model for Eastern Africa and a Tropical Indian Ocean Domain. J Clim 22:3595–3616 Denis B, Laprise R, Caya D (2003) Sensitivity of a regional climate model to the resolution of the lateral boundary conditions. Clim Dyn 20:107–126. doi:10.1007/s00382-002-0264-6 Diaconescu EP, Laprise R, Sushama L (2007) The impact of lateral boundary data errors on the simulated climate of a nested regional climate model. Clim Dyn 28:333–350. doi:10.1007/ s00382-006-0189-6 Dimitrijevic M, Laprise R (2005) Validation of the nesting technique in a RCM and sensitivity tests to the resolution of the lateral boundary conditions during summer. Clim Dyn 25:550–580 Dudhia J (1989) Numerical study of convection observed during the winter experiment using a mesoscale two-dimensional model. J Atmos Sci 46:3077–3107 Ensor LA, Robeson SM (2008) Statistical characteristics of daily precipitation: comparisons of gridded and point datasets. J Appl Meteorol Climatol 47:2468–2476. doi:10.1175/2008JAMC 1757.1 Flaounas E, Bastin S, Janicot S (2010) Regional climate modelling of the 2006 West African monsoon: sensitivity to convection and planetary boundary layer parameterisation using WRF. Clim Dyn. doi:10.1007/s00382-010-0785-3 Friedl MA, McIver DK, Hodges JCF, Zhang XY, Muchoney D, Strahler AH, Woodcock CE, Gopal S, Schneider A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover mapping from MODIS: algorithms and early results. Remote Sens Environ 83:287–302 Gebremichael M, Krajewski WF, Morrissey ML, Huffman GJ, Adler RF (2005) A detailed evaluation of GPCP 1 daily rainfall estimates over the Mississippi River Basin. J Appl Meteorol 44:665–681. doi:10.1175/JAM2233.1 Giorgi F, Bi X (2000) A study of internal variability of a regional climate model. J Geophys Res 105:29, 503–529, 521 Grell GA, De´veńyi D (2002) A generalized approach to parameterizing convection combining ensemble and data assimilation techniques. Geophys Res Lett. doi:10.1029/2002GL015311 Hitt LJ (1994) Refining 1970s land-use data with 1990 population data to indicate new residential development. US geological survey, water-resources investigations report 94-4250 Hong SY, Lim JOJ (2006) The WRF single-moment 6-class microphysics scheme (WSM6). J Kor Meteorol Soc 42:129–151

123

B. Pohl et al.: Testing WRF in East Africa Hong SY, Noh Y, Dudhia J (2006) A new vertical diffusion package with an explicit treatment of entrainment processes. Mon Weather Rev 134:2318–2341 Hu XM, Nielsen-Gammon JW, Zhang F (2010) Evaluation of three planetary boundary layer schemes in the WRF model. J Appl Meteorol Climatol 49:1831–1844. doi:10.1175/2010JAMC 2432.1 Huffman GJ, Adler RF, Morrissey M, Bolvin DT, Curtis S, Joyce R, McGavock B, Susskind J (2001) Global precipitation at onedegree daily resolution from multi-satellite observations. J Hydrometeor 2:36–50 Iacono MJ, Delamere JS, Mlawer EJ, Clough SA, Morcrette JJ, Hou YT (2004) Development and evaluation of RRTMG_SW, a shortwave radiative transfer model for general circulation model applications. Proceedings of the 14th atmos radiation measurement (ARM) science team meeting, Albuquerque, New Mexico, March 22–26 Im E, Ahn J, Remedio A, Kwon W (2008) Sensitivity of the regional climate of East/Southeast Asia to convective parameterizations in the RegCM3 modelling system. Part 1: focus on the Korean peninsula. Int J Climatol 28:1861–1877 Janjic ZI (1994) The step-mountain eta coordinate model: further developments of the convection, viscous sublayer, and turbulence closure schemes. Mon Weather Rev 122:927–945 Janjic ZI (2002) Nonsingular implementation of the Mellor–Yamada level 2.5 scheme in the NCEP meso model. NCEP Office Note N 437, 61 pp Jankov I, Gallus W Jr, Segal M, Shaw B, Koch S (2005) The impact of different WRF model physical parameterizations and their interactions on warm season MCS rainfall. Weather Forecast 20:1048–1060 Kain JS (2004) The Kain–Fritsch convective parameterization: an update. J Appl Meteorol 43:170–181 Kaspar F, Cubasch U (2008) Simulation of East African precipitation patterns with the regional climate model CLM. Meteorol Z 17(4):511–517. doi:10.1127/0941-2948/2008/0299 Kgatuke MM, Landman WA, Beraki A, Mbedzi MP (2008) The internal variability of the RegCM3 over South Africa. Int J Climatol 28:505–520 Kimoto M, Yasutomi N, Yokoyama C, Emori S (2005) Projected changes in precipitation characteristics near Japan under the global warming. SOLA 1:85–88 Leduc M, Laprise R (2009) Regional climate model sensitivity to domain size. Clim Dyn 32:833–854. doi:10.1007/s00382-0080400-z Liang XZ, Choi HI, Kunkel KE, Dai Y, Joseph E, Wang JXL, Kumar P (2005) Surface boundary conditions for mesoscale regional climate models. Earth interactions 09-018 Lim J, Hong S (2005) Effects of bulk ice microphysics on the simulated monsoonal precipitation over East Asia. J Geophys Res. doi:10.1029/2005JD006166 Lin YL, Rarley RD, Orville HD (1983) Bulk parameterization of the snow field in a cloud model. J Appl Meteorol 22:1065–1092 Mellor GL, Yamada T (1982) Development of a turbulence closure model for geophysical fluid problems. Rev Geophys Space Phys 20:851–875 Meynadier R, Bock O, Guichard F, Boone A, Roucou P, Redelsperger JL (2010a) The West African monsoon water cycle. Part I: a hybrid water budget dataset. J Geophys Res 115, D19106. doi: 10.1029/2010JD013917 Meynadier R, Bock O, Gervois S, Guichard F, Redelsperger JL, AgustiPanareda A, Beljaars A (2010b) The West African Monsoon water cycle. Part II: assessment of NWP water budgets. J Geophys Res 115, D19107. doi:10.1029/2010JD013919 Mlawer E, Taubman S, Brown P, Iacono M, Clough S (1997) Radiative transfer for inhomogeneous atmosphere: RRTM, a

B. Pohl et al.: Testing WRF in East Africa validated correlated-k model for the long-wave. J Geophys Res 102:16663–16682 Moore N, Torbick N, Lofgren B, Wang J, Pijanowski B, Andresen J, Kim DY, Olson J (2010) Adapting MODIS-derived LAI and fractional cover into the RAMS in East Africa. Int J Climatol 30:1954–1969. doi:10.1002/joc.2011 Morrison H, Thompson G, Tatarskii V (2009) Impact of cloud microphysics on the development of trailing stratiform precipitation in a simulated squall line: comparison of one-and twomoment schemes. Mon Weather Rev 137:991–1007 Osborn TJ, Hulme M (1997) Development of a relationship between station and grid-box rainday frequencies for climate model evaluation. J Climate 10:1885–1908 Pleim JE (2007a) A combined local and nonlocal closure model for the atmospheric boundary layer. Part I: model description and testing. J Appl Meteorol Climatol 46:1383–1395 Pleim JE (2007b) A combined local and nonlocal closure model for the atmospheric boundary layer. Part II: application and evaluation in a mesoscale meteorological model. J Appl Meteorol Climatol 46:1396–1409 Poccard I, Janicot S, Camberlin P (2000) Comparison of rainfall structures between NCEP/NCAR reanalyses and observed data over tropical Africa. Clim Dyn 16:897–915 Pohl B, Camberlin P (2006) Influence of the Madden-Julian Oscillation on East African rainfall. Part I: intraseasonal variability and regional dependency. QJR Meteorol Soc 132:2521–2539. doi:10.1256/qj.05.104 Pohl B, Camberlin P (2010) Intraseasonal and interannual zonal circulations over the Equatorial Indian Ocean. Theor Appl Climatol, published on line. doi:10.1007/s00704-010-0336-1 Simmons A, Uppala S, Dee D, Kobayashi S (2007) ERA-interim: new ECMWF reanalysis products from 1989 onwards. ECMWF Newslett 110:25–35 Skamarock WC, Klemp JB, Dudhia J, Gill DO, Barker DM, Duda M, Huang XY, Wang W, Powers JG (2008) A description of the advanced research WRF version 3. NCAR technical note, NCAR/TN\u2013475?STR, 123 pp Smirnova TG, Brown JM, Benjamin SG, Kim D (2000) Parameterization of cold-season processes in the MAPS land-surface scheme. J Geophys Res 105(D3):4077–4086 Stensrud DJ (2007) Parameterization schemes. Keys to understanding numerical weather prediction models. Cambridge University Press, Cambridge. ISBN: 9780521865401, 478 pp

1379 Sun LQ, Semazzi F, Giorgi F, Ogallo L (1999a) Application of the NCAR regional climate model to eastern Africa, 1. Simulation of the short rains of 1988. J Geophys Res 104:6529–6548 Sun LQ, Semazzi F, Giorgi F, Ogallo L (1999b) Application of the NCAR regional climate model to eastern Africa, 2. Simulation of interannual variability of short rains. J Geophys Res 104:6549–6562 Sun Y, Solomon S, Dai A, Portmann RW (2006) How often does it rain? J Climate 19:916–934. doi:10.1175/JCLI3672.1 Sylla MB, Coppola E, Mariotti L, Giorgi F, Ruti PM, Dell’Aquila A, Bi X (2010) Multiyear simulation of the African climate using a regional climate model (RegCM3) with the high resolution ERA-interim reanalysis. Clim Dyn 35:231–247. doi:10.1007/ s00382-009-0613-9 Tchotchou L, Kamga M (2010) Sensitivity of the simulated African monsoon of summers 1993 and 1999 to convective parameterization schemes in RegCM3. Theor Appl Climatol 100:207–220 Trenberth KE, Guillemot CJ (1998) Evaluation of the atmospheric moisture and hydrologic cycle in the NCEP/NCAR reanalyses. Clim Dyn 14:213–231 Uppala SM et al (2005) The ERA40 re-analysis. QJR Meteorol Soc 131:2961–3012. doi:10.1256/qj.04.176 Vanvyve E, Hall N, Messager C, Leroux S, Van Ypersele JP (2008) Internal variability in a regional climate model over West Africa. Clim Dyn 30:191–202 Wakazuki Y, Kanada S, Muroi C, Hashimoto A, Kato T, Nakamura M, Noda A, Yoshizaki M, Yasunaga K (2007) Regional climate projection experiments on the Baiu frontal activity around the Japan islands using a non-hydrostatic cloud-system-resolving model. J Earth Sim 8:13–25 Xiu A, Pleim JE (2001) Development of a land-surface model part I: application in a mesoscale meteorology model. J Appl Meteorol 40:192–209 Yin XG, Nicholson SE (1998) The water balance of Lake Victoria. Hydrol Sci J 43:789–811 Zhang X (2007) Adapting the weather research and forecasting model for the simulation of regional climate in East Africa. PhD thesis, North Carolina State University, Raleigh, 227 pp Zhong Z, Wang X, Lu W, Hu Y (2010) Further study on the effect of buffer zone size on regional climate modeling. Clim Dyn 35:1027–1038. doi:10.1007/s00382-009-0662-0

123

Testing WRF capability in simulating the atmospheric water cycle over ...

Testing WRF capability in simulating the atmospheric water cycle over ...

Suggest Documents

The Atmospheric Hydrologic Cycle over the Arctic Basin from

Atmospheric water budget over the western Himalayas in a regional ...

WRF - University Corporation for Atmospheric Research

Stress Testing Over the Financial Cycle: General ...

Simulating atmospheric tracer concentrations for

Atmospheric hydrogen cycle - CiteSeerX

WRF-CHEM simulations over India

Simulations over South Asia using WRF-Chem

development and testing of polar wrf

Simulating Ocean Water

Simulating thick atmospheric turbulence in the lab with application to

Approach for Simulating Acidification and the Carbon Cycle in the ...

Comparison of Cloud Liquid Water Paths Over Atmospheric Radiation ...

Atmospheric water balance over oceanic regions as estimated from ...

Monitoring Atmospheric Water Vapour over Paranal to Optimise VISIR ...

Monitoring Atmospheric Water Vapour over Paranal to Optimise ... - ESO

Simulating atmosphere flow for wind energy applications with WRF

WRF

The International Atmospheric Circulation Reconstructions over the ...

Simulating Atmospheric Turbulence by Synthetic ... - Science Direct

Performance of WRF (ARW) over River Basins in Odisha ... - IJIRST

Modelling the long-term carbon cycle, atmospheric

Adrienne Macartney Simulating atmospheric loss and ...

The Atmospheric Branch of the Hydrological Cycle over the Negro and ...