Improved Seasonal Prediction of Rainfall over East Africa for Food Security

Forecasting: Statistical Downscaling of CFSv2 and GFDL-FLOR

O. Kipkogei1, A.M. Mwanthi1, J.B. Mwesigwa1,2, Z. K.K. Atheru1, and G. Artan1

1 IGAD Climate Prediction and Applications Centre, P. O. Box 10304 – 00100, Nairobi,

Kenya;

2Institute for Climate Change and Adaptation, University of Nairobi, P. O. Box 30197-

00100, Nairobi, Kenya

Corresponding Author:

Oliver Kipkogei

Climate Forecast Downscaling Assistant

IGAD Climate Prediction and Application Centre (ICPAC)

Email : oliver@icpac.net

Mobile: +254726018688

2

ABSTRACT

The evolution of 2015 El Nino event yielded excessive anxiety about the magnitudes of

rainfall expected over the East Africa. In this paper, we evaluate the utility of downscaled

forecasts from two GCM models. The Climate Predictability Tool (CPT) was employed to

statistically downscale GCM rainfall forecasts over Kenya and Tanzania for the month of

October 2015, which usually forms the onset period of the October-December season.

A blend of station and satellite estimates called Climate Hazards Infrared Precipitation with

Station (CHIRPS) at 5x5 Km resolution was used as benchmark data to downscale October

2015 model outputs from CFSv2 and GFDL-FLOR over Kenya and Tanzania. Canonical

Correlation Analysis (CCA) technique was used to make model correction by generating

Empirical Orthogonal Functions (EOFs) on the X (predictor) data then on the Y (observed)

data for the training period 1982-2011 for 5 modes. The 5 modes were then used to make

cross validated forecasts.

Various skill scores were used to validate the downscaled rainfall forecasts for the month of

October 2015. This study revealed that statistical downscaling approach using the CPT was a

worthwhile effort since it provided improved forecast skills in terms of Relative Operating

Characteristics (ROC), and Pearson Correlation for precipitation forecasts for the month of

October over Kenya and Tanzania. CPT is therefore a robust tool that can be applied for

skillful downscaling of weather and climate forecasts over East Africa to generate more

reliable products applicable for agricultural planning and decision making.

3

1. Introduction

East African economy is largely dependent on rain-fed agriculture, which is highly

vulnerable to the negative impacts of climate variability. Weather and climate shocks

frequently impact negatively on the livelihoods of people in the region. For instance, nearly a

third of African population dependent on rain-fed food production face chronic food

insecurity (Hail, 2005). The main driver of this perennial crisis is the close linkage between

weather and climate variability with agricultural production and food security. Improved

weather and climate forecasts can provide more reliable climate information that agricultural

decision makers can rely on to boost agricultural productivity and food security across the

region. A reliable forecasting system is thus of paramount importance for the East African

region.

This paper promotes the need for capacity building for the East Africa National

Meteorological and Hydrological Services (NMHS) to provide timely and high resolution

forecasts that can be used for decision making in vital socio-economic sectors. This will have

a strong effect on the sustainability and productivity of the highly susceptible regional

economies (Washington and Downing, 1999).

Over the years, there has been an increased need for high resolution climate forecasts

for target users in agriculture, hydrology, disaster management and health among others at

sufficient lead times. To generate high resolution local climate anomalies, downscaling

techniques are applied. These can either be statistical or dynamical. In the dynamical

technique, regional nests are used within global circulation models (GCMs) providing

boundary conditions to output climate forecasts of higher resolution (Castro et al., 2006).

However, this technique is quite costly in terms of computing resources and also requires

highly skilled manpower. It’s worthy to note that although dynamical downscaling works

4

under better representation of land features and atmospheric patterns, the model

configurations may introduce errors (Mironov and Raschendorfer, 2001). In statistical

downscaling, there is a target set of observations and the statistical model identifies statistical

relationships between the course resolution GCM and the observational data (Wilby et al.,

1998). This technique is quite effective as it demands minimal computing facilities.

Many studies have been undertaken to improve climate predictability over the Greater

Horn of Africa region. A recent contribution (Rowell et al., 2015) noted that the rains in the

region exhibited a downward trend which was in contrast with major global climate models

that projected an increasing trend for the coming decades, a phenomenon he termed as the

East African climate paradox. Kipkogei et al. (2016) combined global output products from

four centres using the multimodel superensemble technique and noted that it greatly reduced

the forecast errors from day 1 to day 10. Nicholson (2014) examined the predictability of

seasonal forecasts over the Greater Horn of Africa using multiple linear regression and cross

validation. She showed that surface variables like sea surface temperatures provided lower

forecast skills as compared to the atmospheric variables. Rowel et al (2015) further noted that

climate change projection has a high degree of uncertainty over Africa with global models

exhibiting a broad range of disproportion in the magnitude of their temperature and rainfall

change.

Rainfall predictability in the Eastern Africa region is usually derived from several

oceanic and atmospheric features related with the seasonal rainfall. Global ocean Sea Surface

Temperature (SST) patterns including El Nino Southern Oscillation (ENSO) and the Indian

Ocean Dipole (IOD) and SST gradients have been utilized extensively for seasonal rainfall

prediction (Mutai et al., 1998; Saji et al., 1999; Indeje et al., 2000; Marchant et al., 2007;

Gitau et al., 2015). The rainfall seasonality in the region is also modulated by atmospheric

circulations such as the pattern of the Inter-tropical Convergence Zone (Okoola, 1999). Other

5

systems include tropical cyclones over the western Indian Ocean, the East Africa low level jet

and the off-equatorial high pressure cells (Gitau et al., 2015). It is also worthy to note that the

ocean-atmosphere coupling is an important interaction that generate seasonal weather

patterns.

Over the Greater Horn of Africa (GHA) which is a consortium comprising 11

countries including East Africa, the Inter-Governmental Authority on Development (IGAD)

Climate Prediction and Applications Centre (ICPAC) has been spearheading the generation

of seasonal forecasts with a lead time of one month to three months over eleven countries.

The GHA Climate Outlook Forums (GHACOFs) provides a regional consensus of the

expected rainfall anomalies from which national meteorological agencies get a benchmark in

generation of downscaled forecasts.

This paper is part of an ongoing work on generation of improved downscaled monthly

and seasonal forecasts to promote food security initiatives at both national and farming

community levels over East Africa. It is expected that the downscaling technique presented in

this paper will be adopted and utilized for generation of high resolution seasonal forecasts

and derived agro-meteorological variables over the entire Eastern Africa region.

2.0 Area of Study

Kenya and Tanzania are part of the countries that make up the East African

community. Kenya lies within -50 to 50 latitude and 370 to 420 longitude whereas Tanzania

lies within -120 to 00 latitude and 280 to 410 longitude.

Rainfall patterns in the region exhibits a bimodal cycle with March, April and May

being the ‘long rains’ season while the October, November and December is the ‘short rains’

season. Systems that significantly control rainfall seasonality include the Inter-Tropical

6

Convergence Zone (Nicholson 2003; Donohoe et al., 2013) , Indian Ocean Dipole (Owiti et

al., 2008), El Nino Southern Oscillation and associated tele-connections (Terray and

Dominiak 2005; Shawn et al., 2007; Kumar et al., 1999), as well as secondary effects

resulting from tropical cyclone activities (Parker and Jury 1999), among other factors.

3.0 Data

The data used in this study included rainfall hind casts and forecasts from two global

models namely National Centre for Environmental Prediction (NCEP) Coupled Forecast

System version 2 (CFSv2), Geophysical Fluid Dynamics Laboratory - Forecast-Oriented Low

Ocean Resolution (GFDL- FLOR) and a high resolution (5km by 5km) Climate Hazards Infra

Red Precipitation with Station data (CHIRPS). CHIRPS is a blend of actual station data with

satelitte estimates from Climate Hazards Infra Red Precipitation (CHIRPs) (Funk et al.,

2015; Funk and Verdin 2010).

NCEP CFSv2 is a coupled global model with a resolution of 0.9370 (T126) with 24

ensemble members (Saha et al., 2006; 2010; and 2014). This model became operational in

April 2011 and continues to provide forecasts up to date. The scientists at NCEP used the

Climate Forecast System Reanalysis (CFSR) technique (running the model retrospectively) to

obtain the best reforecasts (past forecasts) estimates .

GFDL-FLOR (Vecchi et al., 2014) is an improved version of the Climate Model

version 2.5 (CM2.5) model (Delworth et al., 2012) and Climate Model version 2.1 (CM2.1)

model (Delworth et al., 2006). It has been used extensively to simulate precipitation and

temperature over land (Jia et al., 2015) as well as drought patterns (Delworth et al., 2015). It

is an important contributor to the North American Multi-Model Ensemble (NMME) for

seasonal forecasting (Kirtman et al., 2014). Climate Model version 2.6 (CM2.6) and CM2.5

(Delworth 2012) each has an atmospheric resolution of 0.450 and 0.10 and 0.250 Oceanic

resolution respectively whereas the CM2.1 (Delworth 2006) has an Oceanic resolution of 10

7

and an atmospheric resolution of 1.80. GFDL-FLOR on the other hand has a high land and

atmospheric resolution of 0.450 as that in CM2.5 but with a coarser ocean resolution (10) as

that of CM2.5.

The CHIRPS data was used as benchmark analysis in this study. The methodology for

blending satellite estimaes and actual station data is explained in detail in Funk and Verdin

(2010).

4.0 Methodology

In this study, Climate Predictability Tool (CPT) software developed by the

International Research Institute was used to statistically downscale the GCM rainfall

forecasts. With statistical downscaling there are a target set of observations (CHIRPS) and

the statistical model identifies statistical relationships between the course resolution GCM

and the observational data.

Canonical Correlation Analysis (CCA) in CPT has the ability to detect and correct

bias in amplitude, mean and shape of anomally pattern when trained on hindcast data and

corresponding observations (Mason and Mimmack 2002). For CCA to make a model

correction, it starts with the generation of the Principal Component Analysis (PCA) or

Empirical Orthogonal Functions (EOF) on the X (predictor) data then on the Y (observed)

data for some few modes. CPT uses EOF modes to make cross validated forecasts, then

calculates a goodness index summarizing how good the forecasts are (Mason and

Chidzambwa 2009). Thereafter, the CCA is done on the amplitude of the PCA modes of the

X and Y data to detect the linear rules. The linear rules are then used to make real time

forecasts.

EOFs help in identifying the preferred patterns within many variables. EOF clusters

points or grids that are strongly related resulting to EOF modes that are used for

approximation (Lorenz, 1956). Each value is assigned a weight which can either be positive

8

or negative with the weights showing a coherent pattern in the spatial domain called EOF

loading pattern (eigenvector), (Weare and Nasstrom 1982). The sum of the product of values

at each grid point and their corresponding loading weights gives the temporal score

(amplitude) for that time. The first EOF mode explains the most variance (eigenvalues) with

the second and subsequent ones explaining the remaining variance. Two to six EOF modes

have been established to explain the total variability in a data set with the rest just working on

the noise (Kim et al., 2003; Kim and Wu 1999).

CPT uses the World Meteorological Organization (WMO) standard forecast

verification scores. Among them is the Relative Operating Characteristics (ROC) score which

is recognized as an equivalent to the Hanssen-Kuipers Skill score (KSS) when applied to the

deterministic forecasts (Hanssen and Kuipers 1965, Swets, 1973; Mason, 1982; Harvey, et

al., 1992; Mason and Graham, 1999). ROC is a measure of the quality of the forecasts that

relates the hit rates and the corresponding false alarm rates. A ROC score with larger hit rate

and smaller false alarm i.e., with the area under the curve of greater than 0.5 shows a skilful

forecast whereas one with more false alarms and less hit rate indicates a forecast with worse

skill (Mason and Graham 2002; Kharin and Zwiers 2003).

5.0 Results and Discussion

In this study, downscaled forecasts for the month of October 2015 relevant for

application in agricultural strategic planning were produced for Kenya and Tanzania.

Canonical Correlation Analysis (CCA) technique was used to make model correction by

generating Empirical Orthogonal Functions (EOFs) on the X (predictor) data then on the Y

(observed) data for the training period 1982-2011 for 5 modes. The 5 modes were then used

to make cross validated forecasts.

Figure 1 and 2 show downscaled CFSv2 and GFDL-FLOR model forecasts for

October 2015 over Kenya and Tanzania respectively, and their corresponding deviations.

9

Kenya and the northern part of Tanzania basically lie within the equatorial belt and is

associated with a wet signal during the October, November and December season. For the

month of October 2015, the rainfall forecast for both the CFSv2 and GFDL-FLOR over

Kenya (Figure 1 b, c) indicated that most parts could receive amounts of between 11-100

mm. Generally, the western and central regions were forecasted to be wetter with regions of

the northeastern being driest, with about 101-435 mm and less than 10 mm of rainfall

respectively. The downscaled forecasts from both models had a close spatial resemblance

with the observed rainfall amounts for October 2015 (Figure 1 a, b and c). The computed

difference between the observation and forecasts (Figure 1 d, e) shows that both models had

tendency to over-forecast in the western parts of Kenya around the Lake Vicoria basin.

GFDL was better than the CFSv2 model in simulating the precipitation patterns since

majority of the regions had deviations in the range of -19 to 20 mm (shaded cyan) as

compared to the CFSv2 forecast that had larger coverage of the less than 20mm classes, in

the majority of central and eastern regions of Kenya. However, this is with the exception of

the coastal region where the GFDL forecasts had a tendency to over-estimate.

Tanzania was forecasted to experience dry conditions in the central and southern parts

and wet conditions in the northern regions around the Lake Victoria basin as well as the

western and parts of the northen coast (Figure 2 b, c). A similar spatial pattern is depicted by

the observation over the same period (Figure 2 a). The computed difference between the

observed and downscaled forecasts (Figure 2 d, e) shows that most areas received a fairly

good score of between -19 mm and 20 mm. Regions close to water bodies had extreme

patterns. Both models overforecasted in areas around the lake Victoria region and

underforecasted in the central coastal areas.

Figure 3 shows the forecasts series and cross validated hindcasts (Fig 3a and 3c) for

the location 1.02o S latitude and 32o E longitude (a location in Kenya) for the month of

10

October 2015 for CFSv2 and GFDL-FLOR respectively. The forecasted amounts captured

the direction of the observation both in the training and forecast phases though with some few

cases of false alarm and underforecasting. Both models gave a forecast within the normal

category in the location considered.

Skill scores in terms of ROC and correlation were also computed to gauge on the skill

of the downscaled forecasts with a cross-validation window of five years (2011 to 2014).

Figure 3b and 3d show the ROC score for the same location for the CFSv2 and GFDL-FLOR

respectively. ROC scores for the location 1.02o S latitude and 32o E longitude indicates that

CFSv2 had scores of 0.750 and 0.809 whereas the GFDL-FLOR had scores of 0.755 and

0.675 for the forecast categories of above normal and below normal respectively. This

implies that the models presented are skilful in capturing categorical forecasts in the given

location.

Figure 4 and 5 show the spatial distribution of the ROC scores over Tanzania and

Kenya respectively. The figures show that most of the areas over the two countries had ROC

values greater than 0.5 with some few patches scoring less than 0.5. One striking area is a

small patch over the central Tanzania that scored the lowest ROC score of 0.1 and the

western parts of Kenya for both the forecast categories.

Figure 6 shows the Spearman correlation scores for the month of October 2015 for

both the CFSv2 (Figure 6a and 6b) and GFDL-FLOR (Figure 6c and 6d) for Kenya and

Tanzania. The Spearman correlation gauges the strength of association between two ranked

variables. Correlation scores over Kenya for CFSv2 model showed good skills for the south

eastern parts with a score of between 0.45 and 0.9 for both CFSv2 and GFDL-FLOR. Over

western Kenya, a correlation score of between -0.45 and -0.9 was attained. Most of Tanzania

areas had a score of between 0.15 and 0.3. These are possible forecasting gaps that need to be

addressed for improved climate forecasts over the East African region.

11

6.0 Conclusion and Future work

Weather and climate downscaling is an important area for both research and

applications. Currently, many techniques are being tested and proposed for use in translating

climate forecasts into finer resolutions for improved decision making. Statistical downscaling

using the climate predictability tool is one of the techniques.

This study utilized forecast products from two global models namely the National

Centre for Environmental Prediction (NCEP) Coupled Forecast System (CFSv2),

Geophysical Fluid Dynamics Laboratory -Forecast-Oriented Low Ocean Resolution (GFDL-

FLOR) and a high resolution blended CHIRPS rainfall dataset. Error metrics such as Relative

Operating Characteristics (ROC) and Spearman Correlation were used for the evaluation of

the downscaled forecasts over the two countries. Forecasts for the month of October 2015

were covered in this study.

The main findings of this study shows that downscaled forecasts from both models

had a good skill in estimating the October rainfall amounts indicating quality and value

addition to the forecast products. There was a good spatial resemblance between the forecasts

and the observations. ROC scores for the location 1.02o S latitude and 32o E longitude

indicates that CFSv2 had scores of 0.750 and 0.809 whereas the GFDL- FLOR had scores of

0.755 and 0.675 for the forecast categories of above normal and below normal respectively.

With an exception of some few patches over Kenya and Tanzania that had low skill, the

general results seem to bear the conclusive high skills for the downscaled forecasts.

This paper promotes generation of improved downscaled forecasts to promote food

security initiatives at both national and farming community levels. It is expected that the

downscaling technique proposed here will be adopted in other climate centers within the

Greater Horn of Africa region and beyond and the forecast products utilized for informed

policy making and improved decision making. Further research and work is however needed

12

to improve the existing downscaling techniques and propose new ones. In a future Part II of

this work, the results for November, December and the entire OND season will be presented.

Acknowledgements

This is a publication of research findings by IGAD Climate Prediction and Applications

Centre (ICPAC). The work was partly supported by CGIAR’s Research Program on Climate

Change, Agriculture and Food Security (CCAFS) under a partnership project called

“Integrated Food Production and Food Security Forecasting System for East Africa

(INAPFS)” in 2015. This paper is part of ICPAC’s contribution to the project under activity 2

[P40A235] which is aimed at developing a robust seasonal climate forecasting system with

high spatial and temporal resolution for the East Africa region.

13

References

Castro C.L., R.S. Pielke, and G. Leoncini, 2006: Dynamical downscaling: assessment of

value retained and added using the regional atmospheric modeling system (RAMS). J

Geophysical Res 110(D6):1-21. Doi: 10.1029/2004JD004721

Delworth T.L, A.J. Broccoli, A. Rosati, R.J. Stouffer, V. Balaji, J.A. Beesley, W.F. Cooke,

K.W. Dixon, J. Dunne, K.A. Dunne, J.W. Durachta, K.L. Findell, P. Ginoux, A.

Gnanadesikan, C.T. Gordon, S.M. Griffies, R. Gudgel, M.J. Harrison, I.M. Held, R.S.

Hemler, L.W. Horowitz, S.A. Klein, T.R. Knutson, P.J. Kushner, A.R. Langenhorst,

H.C. Lee, S.J. Lin, J. Lu, L. Sergey, S.L. Malyshev, P.C.D. Milly, V. Ramaswamy, J.

Russell, M.D. Schwarzkopf, E. Shevliakova, J.J. Sirutis, M.J. Spelman, W.F. Stern,

M. Winton, A.T. Wittenberg, B. Wyman, F. Zeng, and R. Zhang, 2006: GFDL's CM2

Global Coupled Climate Models. Part I: Formulation and Simulation Characteristics.

J. Climate, 19, 643–674. doi: http://dx.doi.org/10.1176/JCLI3629.1

Delworth T.L., A. Rosati, W. Anderson, A.J. Adcroft, V. Balaji, R. Benson, K. Dixon, S.M.

Griffies, H.C. Lee, R.C. Pacanowski, G.A. Vecchi, A.T. Wittenberg, F. Zeng, and R.

Zhang, 2012: Simulated Climate and Climate Change in the GFDL CM2.6 High-

Resolution Coupled Climate Model. J. Climate, 26, 2766–2781. doi:

http://dx.doi.org/10.1176/JCLI-D-11-00316.1

Delworth T.L., F. Zeng, A. Rosati, G.A. Vecchi, and A.T. Wittenberg, 2015: A Link between

the Hiatus in Global Warming and North American Drought. J. Climate, 28, 3834–

3845.

doi: http://dx.doi.org/10.1175/JCLI-D-14-00616.1

Donohoe A, J. Marshall, D. Ferreira, D. McGee 2013: The relationship between ITCZ

location and cross equatorial atmospheric heat transport; from the seasonal cycle to

14

the last glacial maximum. J. Climate, 26, 3597–3618. Doi:

http://dx.doi.org/10.1175/JCLI-D-12-00467.1

Funk, C., A. Verdin, J. Michaelsen, P. Peterson, D. Pedreros, and G. Husak, 2015: "A global

satellite assisted precipitation climatology". Earth Syst. Sci. Data Discuss., 8, 401-

426, doi:10.6194/essdd-8-401-2016. 2016.

Funk, C., and J.P. Verdin, 2010: Real-time decision support systems - The famine early

warning system network, in Gebremichael, M., and Hossain, F., eds., 2010, Satellite

rainfall applications for surface hydrology: Springer, Netherlands, Springer, p. 296–

320.

Gitau, W., P. Camberlin, L. Ogallo and R. Okoola. 2015: Oceanic and atmospheric linkages

with short rainfall season intraseasonal statistics over Equatorial Eastern Africa and

their predictive potential. International J. Climatol. 35 (5): 2382–2399.

Haile, M. 2006: Weather patterns, food security and humanitarian response in sub-Saharan

Africa. Philosophical Transactions of the Royal Society B: Biological Sciences, 360

(1463), 2169-2182.

Hansen a.w., and W.J.A. Kuipers, 1965: On the relationship between the frequency of rain

and various meteorological parameters. Koninklijk Nederlands Meteorologisch

institu, meded. Verhand., 81: 2-15

Harvey, L. O., K. R. Hammond, C. M. Lusk, and E. F. Mross, 1992: The application of signal

detection theory to weather forecasting behavior. Monthly Weather Review,120, 863-

883.

Indeje M., F.H. Semazzi, and L.A. Ogallo, 2000: ENSO signals in East African rainfall

seasons. Int. J. Climatol. 20: 19-46

Jia J, X. Yang, G.A. Vecchi, R.G. Gudgel, T.L. Delworth, A. Rosati, W.F. Stern, A.T.

Wittenberg, L. Krishnamurthy, S. Zhang, R. Msadek, S. Kapnick, S. Underwood, F.

15

Zeng, W.G. Anderson, V. Balaji, and K. Dixon, 2015: Improved Seasonal Prediction

of Temperature and Precipitation over Land in a High-Resolution GFDL Climate

Model. J. Climate, 28, 2044–2062. doi: http://dx.doi.org/10.1176/JCLI-D-14-00112.1

Kharin, V., and F. Zwiers. 2003: On the ROC score of Probability Forecasts. Journal of

Climate, 16, 4145-4150.

Kim K.Y, J.J. O’Brien, and A.I. Barcilon, 2003: The principal physical modes of variability

over the tropical pacific, Earth interactions, 7(3), 1-32

Kim K.Y. and Q. Wu 1999: A comparison study of EOF techniques: Analysis of

nonstationary data with periodic with periodic statistics, Clim J. , 12, 186-199

Kipkogei O., A. Bhardwaj, V. Kumar, L.A. Ogallo, F.J. Opijah, J.N. Mutemi, and T.N.

Krishnamurti 2016: Improving multimodel medium range forecasts over the Greater

Horn of Africa using the FSU superensemble. Meteorology and Atmospheric Physics,

1-11.

Kirtman B.P, D. Min, J.M. Infanti, J.L. Kinter, D.A. Paolino, Q. Zhang, H.V.D. Dool, S.

Saha, M.P. Mendez, E. Becker, P. Peng, P. Tripp, J. Huang, D.G. DeWitt, M.K.

Tippett, A.G. Barnston, S. Li, A. Rosati, S.D. Schubert, M. Rienecker, M. Suarez,

Z.E. Li, J. Marshak, Y.K. Lim, J. Tribbia, K. Pegion, W.J. Merryfield, B. Denis, and

E.F. Wood, 2014: The North American Multimodel Ensemble: Phase-1 Seasonal-to-

Interannual Prediction; Phase-2 toward Developing Intraseasonal Prediction. Bull.

Amer. Meteor. Soc., 96, 686–601. doi: http://dx.doi.org/10.1176/BAMS-D-12-00060.1

Kumar K., B. Rajagopalan, M. Cane, 1999: On the Weakening Relationship between the

Indian Monsoon and ENSO. Science 284(2); 2156-2159; American Association for

the Advancement of Science; Washington DC

Lorenz E.N 1956: Empirical orthogonal functions and statistical weather prediction, Sci. Rep.

1, 49p., Mass. Inst. Of Technol., Cambridge.

16

Marchant, R., C. Mumbi, S. Behera, and T. Yamagata, 2007: The Indian Ocean dipole–the

unsung driver of climatic variability in East Africa. African Journal of Ecology, 46

(1), 4-16.

Mason, I., 1982: A model for assessment of weather forecasts. Australian Meteorological

Magazine, 30, 291-303.

Mason, S.J., and G.M. Mimmack, 2002: Comparison of some statistical methods of

probabilistic forecasting of ENSO. Journal of Climate, 16, 8–29.

Mason, S.J., and N.E. Graham, 1999: Conditional Probabilities, Relative Operating

Characteristics, and Relative Operating Levels., Weather and Forecasting, 14, 713-

725.

Mason, S.J., and N.E. Graham, 2002: Areas beneath the relative operating characteristics

(ROC), and relative operating levels (ROL) curves: Statistical significance and

interpretation. Q.J.R. Meteorol. Soc., 128, 2145-2166.

Mason, S.J., and S. Chidzambwa, 2009: Verification of RCOF Forecasts. WMO RCOF

Review 2008 Position Paper, 26 pp.

Mironov D., and M. Raschendorfer, 2001: Evaluation of empirical parameters of the new LM

surface layer parameterization scheme: results from numerical experiments including

soil moisture analysis. COSMO technical report 1, DWD, Offenbach, Germany

Mutai, C. C., M.N. Ward, and A.W. Colman, 1998: Towards the prediction of the East Africa

short rains based on sea-surface temperature–atmosphere coupling. International

Journal of Climatology, 18 (9), 976-997.

Nicholson S.E., 2003: Comments on “The South Indian Convergence Zone and interannual

rainfall variability over southern Africa” and the question of ENSO’s influence on

southern Africa. Journal of Climate 16: 555 – 562

17

Nicholson S.E. 2014: The predictability of rainfall over the Greater Horn of Africa. Part I:

prediction of seasonal rainfall. J Hydrometeor 15(3):1011–1027

Okoola R.E., 1999: Midtropospheric Circulation Patterns Associated with Extreme Dry and

Wet Episodes over Equatorial Eastern Africa during the Northern Hemisphere Spring.

J. Appl. Meteor., 38, 1161–1169. doi: http://dx.doi.org/10.1175/1520-

0450(1999)038<1161:MCPAWE>2.0.CO;2

Owiti, Z., L. A. Ogallo, and J. Mutemi, 2008: Linkages between the Indian Ocean dipole and

East African seasonal rainfall anomalies. J. Kenya Meteor. Soc., 2 (1), 3–17.

Parker B.A., M.R. Jury, 1999: Synoptic environment of composite tropical cyclones in the

South-West Indian Ocean. South African Journal of Marine Science 21: 99 – 115.

Rowell D.P, B.B.B. Booth, S.E. Nicholson, and P. Good, 2015: Reconciling Past and Future

Rainfall Trends over East Africa. J. Climate, 28, 9768–9788.

doi: http://dx.doi.org/10.1175/JCLI-D-15-0140.1

Rowell D. P., C. A. Senior, M. Vellinga, and R.J. Graham 2015: Can climate projection

uncertainty be constrained over Africa using metrics of contemporary performance?

Climatic Change, 1-13.

Saha S, S. Moorthi, H.L. Pan, X. Wu, J. Wang, S. Nadiga, P. Tripp, R. Kistler, J. Woollen,

D. Behringer, H. Liu, D. Stokes, R. Grumbine, G. Gayno, J. Wang, Y.T. Hou, H.Y.

Chuang, H.M.H. Juang, J. Sela, M. Iredell, R. Treadon, D. Kleist, P.V. Delst, D.

Keyser, J. Derber, M. Ek, J. Meng, H. Wei, R. Yang, S. Lord, H.V.D. Dool, A.

Kumar, W. Wang, C. Long, M. Chelliah, Y. Xue, B. Huang, J.K. Schemm, W.

Ebisuzaki, R. Lin, P. Xie, M. Chen, S. Zhou, W. Higgins, C.Z. Zou, Q. Liu, Y. Chen,

Y. Han, L. Cucurull, R.W. Reynolds, G. Rutledge, and M. Goldberg, 2010: The

NCEP Climate Forecast System Reanalysis. Bull. Amer. Meteor. Soc., 91, 1016–1067.

18

Saha S, S. Moorthi, W. Xingren, J. Wang, S. Nadiga, P. Tripp, D. Behringer, Y. Hou, H.

Chuang, M. Iredell, M. Ek, J. Meng, R. Yang, M.P. Mendez, H. Dool, Q. Zhang, W.

Wang, M. Chen, and E. Becker, 2014: The NCEP Climate Forecast System Version 2.

J. Climate, 27, 2186–2208. doi: http://dx.doi.org/10.1176/JCLI-D-12-00823.1.

Saha S., S. Nadiga, C. Thiaw, J. Wang, W. Wang, Q. Zhang, H. M. van den Dool, H.-

L. Pan, S. Moorthi, D. Behringer, D. Stokes, M. Peña, S. Lord, G. White, W.

Ebisuzaki, P. Peng, P. Xie, 2006: The NCEP Climate Forecast System. J.

Climate. 19, 3483-3617.

Saji N.H., B.N. Goswanmi, P.N. Vanayachandran T. Yamagata 1999: A dipole mode in the

tropical Indian Ocean. Nature, 401: 360-363

Shawn R.S., J. Brolley, J.J. O’Brien, A.T. Carissa, 2007: ENSO’s Impact on Regional U.S.

Hurricane Activity. J. Climate, 20, 1404–1414.

Swets, J. A., 1973: The relative operating characteristic in psychology. Science, 182, 990-

1000.

Terray P, and S. Dominiak, 2005: Indian Ocean Sea Surface Temperature and El Nino-

Southern Oscillation: A new perspective. Journal of Climate 18: 1351 – 1368,

DOI:10.1175/JCLI3338.1

Vecchi G. A., T. Delworth, R. Gudgel, S. Kapnick, A. Rosati, A. T. Wittenberg, F. Zeng, W.

Anderson, V. Balaji, K. Dixon, L. Jia, H.-S. Kim, L. Krishnamurthy, R. Msadek, W.

F. Stern, S. D. Underwood, G. Villarini, X. Yang, and S. Zhang, 2014: On the

Seasonal Forecasting of Regional Tropical Cyclone Activity. J. Climate, 27, 7994–

8016. doi: http://dx.doi.org/10.1175/JCLI-D-14-00158.1

Washington, R., and T.E. Downing, 1999: Seasonal forecasting of African rainfall:

prediction, responses and household food security. Geographical Journal, 266-274.

19

Weare, B.C., and J.S. Nasstrom, 1982: Examples of extended empirical orthogonal function

analyses, Mon. Weather Rev., 110, 481-486

Wilby R.L., H. Hassan, and K. Hanaki, 1998: Statistical downscaling of hydrometeorological

variables using general circulation model output. J. hydrol., 206, 1-19

20

Figure 1: Downscaled CFSv2 and GFDL-FLOR model forecasts for October 2015 over

Kenya (Fig. 1, b and c respectively). Figure 1, d and e shows the difference between the

model forecast and October observation. Figure 1a shows the observation for October.

21

Figure 2: Downscaled CFSv2 and GFDL-FLOR model forecasts for October 2015 over

Tanzania (Fig. 2, b and c respectively). Figure 2, d and e shows the difference between

the model forecast and October observation. Figure 2a shows the observation for

October.

22

Figure 3: Forecasts and cross validated hindcasts (Fig 3a and 3c) for the location 1.02oS

latitude and 32oE longitude (a location in Kenya) for the month of October 2015 for

CFSv2 and GFDL-FLOR respectively. Fig 3b and 3d shows the ROC score for the same

location for the CFSv2 and GFDL-FLOR respectively

23

Figure 4: Relative Operating Characteristics area for the month of October 2015 for the

above normal and below normal rainfall forecast categories for Tanzania. Fig 4a and 4b

is scores for the CFSv2 model while fig. 4c and 4d shows score for the GFDL-FLOR

model.

24

Figure 5: Relative Operating Characteristics area for the month of October 2015 for the

above normal and below normal rainfall forecast categories for Kenya. Fig 5a and 5b is

scores from the CFSv2 model while Fig. 5c and 5d shows score for the GFDL-FLOR

model.

25

Figure 6: Spearman correlation scores for the month of October 2015 for CFSv2 (Fig 6a

and 6b) and GFDL-FLOR (Fig 6c and 6d) for Kenya and Tanzania