Proposed research strategy for short lag predictions in Eastern and Southern Africa
White Paper

7/21/2003

Based on a teleconference between Maxx Dilley, Chris Funk, Matayo Indeje and Neil Ward, building on previous discussions including the expanded RVF group (Paul Rwambo, Laban Ogallo, Jerry Stuth, Hussein Gadain) as well as prior efforts and discussions carried out by hydrologic and macroscale crop modeling researchers (Gideon Galu, Tamuka Magadzire, Elijah Mukhala, Kennedy Masamvu, Emmanuel Dlamini, Emma Archer, James Hansen, Brad Lyon, Jim Verdin, Jim Rowland, Gabriel Senay, Kwabena Asante, Eric Wood and Guleid Artan).

Introduction

This document outlines research that combines the experience and efforts of both the early warning (EW) and forecast communities to provide short lag predictions of several key environmental variables:   Normalized Difference Vegetation Index (NDVI), streamflow and soil moisture from the USGS Stream Flow Model (SFM), and water requirement satisfaction index (WRSI) values. The goal is to integrate early warning and forecast information into a single set of products. At long lag times1 forecasts can be based solely on climate forecast information. At short lags (just before or within a season) a combination of observed and forecast information will likely provide the best estimates. At the end-of-season values based solely on observations can be used (Figure 1).   Many institutions currently produce long lag (100% forecast) and observational (100% monitoring) products. Combining these products (Figure 2) can improve skill levels and make products much easier to use, since both monitoring and forecasts are made in the same ‘units’ (i.e. NDVI, WRSI, ...).

This collaborative research could result in three integrated monthly forecast/monitoring products: i) maximum AVHRR Normalized Difference Vegetation Index imagery, ii) extended-to-end-of-season WRSI, iii) Stream Flow Model (SFM) percent of historic max stream flow, and iv) SFM soil moisture. These products help fulfill the specified needs of the user community, with specific links to the Rift Valley Fever project (Dilley, 2003), tailored agricultural products (Archer, 2003), and flood forecasting products in the GHA and Southern Africa.

Background

Early warning systems, based primarily on satellite observing systems operating at high temporal and spatial resolutions, offer useful information about current conditions, including floods (Artan et al., 2001; Dvorsky et al., 1999) and agricultural drought (Senay and Verdin, 2001; Verdin and Klaver, 2002). Forecast systems, such as those at the IRI, CPC and DMCs, offer information regarding future conditions in the Greater Horn (Mutai and Ward, 2000; Mutai et al., 1998a; Mutai et al., 1998b; Ward, 1998) and Southern Africa (Barnston et al., 1996; Nicholson and Kim, 1997; Makarau and Jury, 1997; Mason, 1998; Jury et al., 1999; Landman and Mason, 1999; Tennant, 1999). Furthermore, recent downscaling techniques (Indeje et al., 2001) such as the Model Output Statistics (MOS) approach (Landman and Goddard, 2001) offer the exciting new possibility of providing rainfall forecasts on the same scale as satellite-based rainfall estimates (RFE) observations.

The existence of RFE-analog precipitation time series with longer periods of records (Funk et al., 2003b) can assist in the creation of tailored forecasts. Validations of this data set (ibid) showed some systematic bias, but good signal-to-noise ratios with mean absolute errors of about ½ the standard deviation for 0.2 o/dekadal accumulations. A validation of the CPC’s satellite RFE time-series for March in western Kenya revealed a strong similarity to high density gridded station data at regional scales with a correlation of 0.9 and a mean bias of only 5 mm/dekad (Funk and Verdin, 2003). Products derived from the CHARM and CPC datasets (WRSI, SFM streamflow and soil moisture) are natural candidates for current tailored forecast activities of the IRI and collaborators, as is the NDVI time series.

Preliminary efforts have been made forecasting NDVI (Verdin et al., 1999; Indeje, 2003)   (see Figure 3 for an example produced by Matayo Indeje) and WRSI (Funk et al., 2002b) with quite promising results. While these efforts are notable, short lag forecasts of these products could probably be improved by integrating observed values of precipitation and/or NDVI. An example of a successful application of such an integrated system (operating on a scale of days) is the USGS Streamflow model. The SFM uses satellite RFE to estimate current soil moisture and stream flow, and daily quantitative precipitation forecasts to make projections of flood conditions. The SFM, guided by longer period climate products, has provided the basis for several successful flood forecasts2 in Somalia.  Forecast SFM values combine observed and forecast information in a meaningful, easy to interpret format that is updated in a timely manner. Collaboration between the IRI, USGS, DMCN, DMCH and RRSU could extend the SFM forecast envelope and apply a similar approach to NDVI and WRSI time series. Each of these categories is outlined briefly, and possible points of collaboration highlighted.

NDVI Forecasting

Hypothesis: Monthly NDVI values in East Africa are well correlated with precipitation from the previous two months, plus the current month (Nicholson et al., 1990)3. Knowing the current NDVI values, and/or the two most recent months of precipitation, should give a good indication of future NDVI values. This skill, based on time-lagged precipitation-vegetation response, should increase skill levels beyond the (already good) levels of skills found for forecasts based on climate fields alone (c.f. Figure 3). The addition of observed mesoscale information will also assist in producing forecasts with higher spatial resolutions. A joint EW/forecast-based NDVI prediction will satisfy the user needs of the RVF community (possibly a 1 month ahead, 11 km 2 forecasts with cross-validated correlation values on the order of 0.8?).

Data: The recent deployment of an enhanced 8 km 20 year (1982-present) enhanced NDVI AVHRR time series for Africa (NDVIe) by NASA GIMMS for FEWS NET partners. This data set is the result of several years of work, and has been corrected for satellite drift and cloud contamination (Pinzón et al., 2001; Pinzón et al., 2002; Pinzón 2002). This is a very exciting resource that will greatly enhance the RVF prediction capacities by improving the baseline accuracy of the monitoring products. The USGS FEWS NET program has two sources of gridded rainfall estimates: the Collaborative Historical African Rainfall Model (CHARM), and satellite-based rainfall estimates provided by our partners at the CPC (Herman et al., 1997;   Xie and Arkin, 1997; Xie and Arkin, 1998). The CHARM time-series runs from 1961 through 1996, while the NDVI extends from 1982 to the present. Average monthly/0.1 o CHARM and maximum monthly NDVI cubes will be put together by Chris Funk, and made available to all collaborators via ftp.

Proposed Methodology:  1. A study of monthly NDVI and CHARM precipitation predictability will be carried by the IRI (Matayo Indeje and Neil Ward) using an SVD/regression technique similar to that used in the NDVI test case (Figure 3). 2. The USGS (Chris Funk in consultation with Molly Brown at NASA GIMMS) will examine the skill of forecasts based solely on the NDVI/precipitation relationships. 3. Matayo and Neil will provide cross-validated time series of monthly maximum NDVI and precipitation to Chris, and he will attempt to improve on either 1 or 2 by combining results.

Proposed Timeline : An experimental forecast in time for the August RVF workshop and GHACOF12 meeting. Operational forecasts in place before short rains (October-November).

Caveats and Comments: Quantification of the NDVI-precipitation relationship (Step 2) will help create realistic expectations for NDVI forecasts. Where this relationship is weak, forecast skill should be low. The CPC is working on a consistent GPI-only dataset for the 1991-2003 period, this should prove useful when available.

Streamflow and Soil Moisture Forecasting

One challenge to stream flow forecasting is the difficulty of satisfying both flood and hydroelectric power users. Flood alerts require daily information, while hydroelectric power applications require longer leads and accumulations. One is based on weather, and the other, climate. Regional climate models have been suggested as a useful way to bridge this gap (Dilley, 2003). Embedded regional models could usefully extend the 3-day prognostics currently available for the SFM. Another useful approach might be forecasting streamflow and soil moisture values directly, through statistical downscaling of IRI precipitation forecasts. A third approach would use current streamflow from the SFM and forecast fields from the IRI in a combined ‘short lag’ forecasting scheme. These three cases correspond to ‘extended monitoring’, ‘long lag forecasting’, and ‘short lag’ forecasting scenarios.

Background: The USGS Streamflow Model (SFM) (Artan, 2001) has been used operationally for several years in the Horn and Southern Africa. The SFM is a distributed parameter soil water balance model within a GIS framework. Jim Verdin, Guleid Artan, Kwabena Asante, Miguel Restrepo and Hussein Gadain have provided model development and training expertise. Kwabena Asante has also developed models and GIS tools to estimate inundated areas, contributing substantially to the Mozambique flood atlas. Jim Verdin is a member of the WMO Commission for Hydrology's Working Group on Hydrologic Forecasting and Prediction. At the IRI Sankar Arumugam and Upmanu Lall (Sankarasubramanian and Lall, 2003) have been producing probabilisitic forecasts of seasonal stream flows based on IRI forecast terciles, and integrating these forecasts with regional policy decisions. In South Africa, Willem Landman, Simon Mason, and William Tennant have been producing categorical streamflow forecasts using downscaled GCM output (Landman et al., 2001). Another salient publication in this area is by Mark Jury (Jury, 2002), which demonstrates that simple multivariate regression models can produce effective predictions of South African streamflow on seasonal timescales. Tamuka Magadzire’s masters thesis at UCSB focused on remote sensing of flood waters in the Limpopo and Zambezi rivers.

Extended Monitoring: Through the efforts of the IRI, DMCN and SAWS, regional climate modeling capabilities are now available in the Horn and Southern Africa. IRI, furthermore, has been providing a series of training sessions; building regional capacity. Statistical downscaling approaches can be used (Landman et al., 2001) to correct for bias in modeled precipitation fields, which may then be incorporated in the SFM data stream, extending the forecast projections beyond the three-day limitation of the Air Force model currently being used by FEWS NET. This would provide an enhanced capacity to anticipate and prepare for floods. Guleid Artan, Kwabena Asante, Miguel Restrepo and Hussein Gadain could work with Matayo Indeje, Willem Landman and Simon Mason, creating the links between regional climate forecasts and the SFM for the Horn and Southern Africa. Kwabena Asante has produced high resolution inundated area maps for the Limpopo basin which correspond to various flood categories – these maps, keyed to IRI rainfall forecasts and the FEWS-SFM simulated flood categories – could be very effective tools for hazard preparedness in both the Horn and SADC countries.

Lagged Forecasts : Driving the SFM with the CHARM and RFE time-series provides a 40 year time series of daily soil moisture and streamflow. Because of the uncertainties in the rainfall data and difficulties associated with paramaterizing the SFM in regions with sparse stream data, the SFM streamflows are reported as percentile ranks. Four variables derived from the SFM appear to be likely candidates for forecasting efforts

• Monthly max daily percentile rank streamflow
• Monthly average daily percentile rank streamflow
• Monthly Average soil moisture percent (% of water holding capacity)
• Monthly number of ‘wet soil days’ (# days above XX% soil moisture percent)

This list will need to be refined in advisement with RVF and water resource management ewater resource management experts.   Time series of the above variables at watershed spatial scale could be forecast directly through statistical means (Sankarasubramanian and Lall, 2003) based on IRI forecast fields. Statistical comparisons between monthly/seasonal rainfall totals and monthly/seasonal SFM products should also be carried out. If reasonably strong statistical relationships are found to exist between precipitation and SFM variables, then the estimation of these products as a function of observed and forecast precipitation should be evaluated as an input to the RVF project. Example: Next month’s soil moisture might be strongly linked to this month’s and next month’s precipitation. A good short lag forecast might be obtained by combining this month’s observed precipitation and next month’s forecast precipitation. The accuracy of this approach should be compared to a ‘pure’ forecast strategy, similar to the results obtained for NDVI (Figure 3).

Potential Focus Regions :   Current USGS data resources suggest that the Awash and Limpopo rivers are logical starting points for these activities. A visiting professor from Ethiopia will be at the EDC from the last week of July through the first week of August to work with EDC hydrologists. Hussein Gadain has agreed to callibrate the SFM for the Awash basin and provide SFM data for this region as an input for RVF-related activities. Reports forwarded by Tamuka Magadzire suggest the portions of the Limpopo basin (near Beitbridge) have been experiencing substantial water shortages, so this region might be a natural focuse for activities tied the upcoming SARCOF.

Timeline: The same cross-validated monthly 0.1 o precipitation fields produced by the IRI for the NDVI project can be used to estimate SFM products. Expiremental forecasts based on these results could be presented at the GHACOF12’s flood forecasting workshop.

Macroscopic Crop Forecasting (WRSI)

Crop health, in drought prone areas, is strongly linked to precipitation (which is not so surprising). Cereal yields and production in Ethiopia and Zimbabwe, for example, show strong connections to seasonal precipitation totals (Figure 4). Studies of the relationship between precipitation and WRSI in SADC crop growing regions (Figure 5) reveal a linear relationship between WRSI and precip correlations and total seasonal rainfall for regions with seasonal (September-March) precipitation totals under 800 mm. Beyond 800 mm there is little correlation between end-of-season WRSI and seasonal precipitation. Analyses of this sort will help define areas in which we can expect reasonable levels of skill for a WRSI forecast. For both NDVI and WRSI, it will likely be desirable to screen out areas of persistent high rainfall.

Background: The 2002/2003 season saw considerable use of the WRSI time-series as a basis for analysis and forecasting in Southern Africa. Historical composites of El Niño years (Funk et al., 2002a) identified strong impacts in a food insecure region (Zimbabwe), confirming many previous studies (e.g. Cane et al., 1994; Makarau and Jury, 1997, …) with a greater degree of spatial resolution. Comparisons of lagged precipitation-rainfall relationships (Husak et al., 2002) showed a strong negative relationship between early rains and late rains during El Niño years, suggesting that an early onset of rains in 2002 might forebode ill. Statistical forecasts based on October reanalysis SST, 200 hPa winds and precipitable water (Funk et al., 2002) anticipated reduced WRSI values in Southern Zimbabwe, Southern Mozambique and Northeastern Republic of South Africa (Figure 6). In mid-season, the RRSU, FEWS NET and DMC-H   pioneered an exciting new application of the updated Climate Outlook Forum tercile probabilities (Figure 7) calculating the probability of each location receiving at least 75% of its crop water requirement for the remainder of a season. This report opened an exciting new avenue of research – pointing out the possible connections between climate forecasts and early warning products. Since the close of the season the IRI and the RRSU have been busy organizing activities for the next season, and a lively discussion regarding tailored agricultural products has taken place (Archer, 2003). In the Horn, a WRSI analysis has revealed a preponderance of low fequency ‘trend’ variabiltity when compared to El Niño impacts (Figure 7). A recent report (Funk et al., 2003a) highlighted the impact of this trend in Ethiopia and made yield projections based on the strong relationship between observed April-May rainfall in the long cycle crop growing region and national Meher yield data. Such projections could benefit from the inclusion of forecast information.

Data:   A 40 year, 0.1 o   time-series of CHARM and RFE-based end-of-season WRSI is available for the the GHA and SADC regions. These time series have been validated through comparison with independent yield data (Verdin and Klaver, 2002) and site visits. These validation efforts suggest a good correspondence with field-based yield estimates (r~0.8).

Hypothesis: End-of-season WRSI can be effectively forecast as a function of both observed and forecast precipitation. This combined forecast/monitoring approach performs better than either approach applied independently. Expressed numerically, our hypothesis is that ‘seasonal WRSI’ = (Seasonal Precipitation/Seasonal Crop Water Requirement) is a useful indicator of yields that may be effectively predicted.

Methodology:   Two critical assumptions in this work are that: i) the interannual variability of a crop’s water requirement is small compared to the interannual variability of precipitation, and ii) the impact of intraseasonal variability does not prevent the seasonal WRSI indicator from being an effective estimator of yields. Tamuka Magadzire, Elijah Mukhala and Greg Husak are currently examining total seasonal precipitation and crop water requirement distributions in Southern Africa. The interannual variability of precipitation and water requirements4 will be evaluated. The similarity of seasonal WRSI and dekadal WRSI patterns needs to be evaluated. The relationship between seasonal WRSI and independent yield data should be ascertained. Assuming these results are favorable, a product merging observed precipitation and IRI/DMCH forecast precipitation could then be deployed.

Timeline :   The RRSU and DMCH have expressed the desire to follow up their previous ground breaking study (Figure 7) with a WRSI interpretation at the next SARCOF. Since the new COF probability map will benefit greatly from downscaling and training activities sponsored by IRI, this is an excellent opportunity to coordinate activities. The USGS has agreed to make the necessary modifications to the FEWS NET AgroClimatology Toolkit, enabling a joint RRSU/DMC presentation based on the final COF map. Following the SARCOF it might be desirable to update the WRSI forecast map on a monthly basis. At a minimum this could combine observed and forecast precipitation, where the forecast is based on the COF map. It might be desirable to have the forecast component updated as well.

Caveats and Comments : As for NDVI, care should be taken when making projections in regions with high precipitation. One frequently consulted resource, the USGS WRSI extended-to-end-of-season maps, could benefit greatly from innovations in this area, since current WRSI conditions are extended through the use of climatological rainfall fields.

Summary: Component Based Estimation Procedures?

One basic science question common to the topics discussed here is whether it is better to directly forecast tailored products or estimate them as a function of precipitation. While the case for the former is compelling (simpler and fewer parameters) and should definitely be considered, there are several appealing aspects about using a precipitation-based approach:

• Precipitation can act as a lingua franca, allowing the early warning and forecast communities to communicate and collaborate.
  
• A precipitation-based approach allows early warning and forecast institutions to remain focused on an enhanced core product (rainfall), confident that improvements will adhere to derived tailored products.
  
• A precipitation-based approach allows various precipitation modeling techniques (statistical, dynamical, etc.) to be used interchangeably5. This is particularly important given that both satellite rainfall estimates and precipitation forecasts are undergoing constant improvements. Future changes in these fields are one of the few certainties.
  

The use of sea surface temperatures in the climate community seems a very successful analog. Climate researchers routinely integrate observed and forecast SST values in making predictions, and the ‘best’ way at deriving these fields is a topic of healthy debate. While it is not at all certain that precipitation can fulfill the same role in tailored products, an evaluation of this hypothesis seems warranted.

1We define ‘long’ lag times as those in which the observed land surface properties (e.g. soil moisture, precipitation, NDVI) contain no useful information regarding potential hazards. At ‘short’ lags the land surface properties can play an important role.

2FEWS NET Report Flood Risk Alert for Southern Somalia: April 30 th 2002
FEWS NET Somalia Weather Advisory for Somalia: Heavy rains may cause flooding in Juba and Shabelle Valley – October 2002

3Nicholson found a log-linear relationship between precipitation and NDVI in the GHA, with the rainfall/NDVI relationship breaking down in regions of high rainfall.

4With precipitation and water requirements calculated over the length of each growing period.

5Though it is likely that the functions mapping precipitation accumulations to tailored products will need re-calibration.

Figure 1. Conceptual diagram of an integrated monitoring/forecast product for an idealized season (this could also be a flood event). At long lags future estimates are based solely on climate forecasts. After the season (or event) is over, observed values provide all the information needed. In between (from shortly before to midway through the event), merged EW/forecast short lag products are likely to be the most effective.

Figure 2. Schematic diagram showing short lag forecast process. Antecedent conditions (current soil moisture, stream flow, crop health and vegetation status) will impact future observations of these values, and provide one source of forecast skill. Slowly evolving boundary conditions, interpreted by ensemble GCM fields and MOS, represent another source of information about the future. Short lag forecasts, at the intersection of these sources of information, should benefit from both.

STATISTICAL NDVI PREDICTION SKILL OVER KENYA
(Orography shown in contour)

Statistical Prediction of NDVI over Eastern Kenya from Global model Rainfall Forecast (correlation coefficient =0.84)

Figure 3. Forecast skill map and NDVI time series for Eastern Kenya October-November NDVI. Produced by Matayo Indeje (IRI) and presented at the Port Said RVF workshop. The methods used here are based on Singular Value Decomposition and multiple linear regression fit, with application of cross-validation technique. In 1994, the October and November NDVI data was missing.

  Figure 4. Comparisons between Ethiopian national Meher yields (1995-2002) and Zimbabwe relative production (% average) and seasonal precipitation. Ethiopia values were based on Central Statistics Authority (CSA) data and combined CHARM (1995) and satellite RFE (1996+) precipitation time series. The 2001 yield value (a dramatic outlier) was excluded due to large (unrealistic?) variations in area planted. Zimbabwe production estimates (1991-1996) were provided by the Regional Remote Sensing Unit and are compared with CHARM precipitation.

Figure 5. Correlations between seasonal precipitation and end-of-season WRSI for SADC crop growing regions. Based on the 1961-1996 CHARM time-series.

Figure 6. Comparison of maize Water Requirement Satisfaction Index forecast (made in mid-November, 2002, in collaboration with the RRSU) with March 2003 end-of-season values for Southern Africa. Values are expressed as a percent of the historic median for each location. The technique (matched filter regression) and cross-validated skills (between 0.6-0.8 for individual regions) were similar to those used in the IRI NDVI forecast (Figure 3).

Figure 7. Chance of receiving 75% of the estimated crop water requirement in February- March-April, conditioned on the updated climate outlook issued by the DMCH. Water requirements were computed using onset of rains and crop phenology. The probability of receiving these amounts were then estimated   using the Forecast Interpretation Tool.   From Interpretation of February-April 2003 forecast update, implications on maize agriculture in the SADC region. Produced jointly by the RRSU, DMCH and FEWS NET.

Figure 8. Schema of component based approach to tailored products. If (big if) tailored products can be effectively estimated as a function of forecast and observed precipitation, then the forecast and early warning communities can benefit from a component based approach to development. The early warning and forecast communities can focus on improving their core products (such as rainfall estimates), confident that any improvements in these areas will be passed on to the derived tailored products. Different approaches to rainfall estimation may be swapped in and out. An analogy might be the treatment of sea surface temperature fields and derived products from numerical weather prediction models.

Archer, E., 2003: Concept paper for Tailored Agricultural Products. F. Chris, Ed.

Artan, G., J. Verdin, and K. Asante, 2001: A wide-area flood risk monitoring model. 5th International Workshop on the application of remote sensing in hydrology, Montpelier, France.

Barnston, A., W. Thiao, and V. Kumar, 1996: Long-lead Forecasts of Seasonal Precipitation in Africa using CCA. Weather and Forecasting, 11, 506-520.

Cane, M. A., E. Gidon, and R. W. Buckland, 1994: Forecasting Zimbabwean Maize Yields Using Eastern Equatorial Pacific Sea Surface Temperature. NATURE, 370, 204-205.

Dilley, M., 2003: Rift Valley Fever (RVF) Risk Model:   Design Process. C. Funk, Ed.

Dvorsky, J. R., J. Verdin, and G. Artan, 1999: Flood risk monitoring in the Greater Horn of Africa using Meteosat-based daily rainfall estimates and a distributed parameter soil water balance model within a GIS framework. American Geophysical Union Hydrology Days, Fort Collins, Colorado, Hydrology Days Publications, p. 111-124.

Funk, C. and J. Verdin, 2003: Comparing Satellite Rainfall Estimates and Reanalysis Precipitation Fields with Station Data for Western Kenya. International workshop on crop monitoring and early warning for food security, Nairobi, JRC - FAO.

Funk, C., T. Magadazire, G. Senay, J. Rowland, J. Verdin, and J. Michaelsen, 2002a: Historical El Niño/crop production relationships in Southern Africa. FEWS NET Special Report.

Funk, C., T. Magadazire, G. Husak, J. Verdin, J. Michaelsen, and J. Rowland, 2002b: Forecasts of 2002/2003 Southern Africa Maize Growing Conditions Based on October 2002 Sea Surface Temperature and Climate Fields. FEWS Special Report.

Funk, C., A. Alemu, P. Steffen, G. Senay, J. Rowland, and J. Verdin, 2003a: Estimating Meher Crop Production Using Rainfall in the 'Long Cycle' Region of Ethiopia

Issued.

Funk, C., J. Michaelsen, J. Verdin, G. Artan, G. Husak, G. Senay, H. Gadain, and T. Magadazire, 2003b: The Collaborative Historical African Rainfall Model: Description and Evaluation. International Journal of Climatolgy, 23, 47-66.

Herman, A., V. B. Kumar, P. A. Arkin, and J. V. Kosuky, 1997: Objectively Determined 10-Day African Rainfall Estimates Created for Famine Early Warning Systems. International Journal of Climatolgy, 18, 147-164.

Husak, G., F. C., J. Rowland, J. Michaelsen, and J. Verdin, 2002: Observations between October and Late-season Rainfall for Southern Africa, 2 pp.

Indeje, M., 2003: Statistical NDVI Prediction Over Kenya. Port Said RVF Conference, Port Said, Egypt.

Indeje, M., F. H. M. Semazzi, L. Xie, and L. J. Ogallo, 2001: Mechanistic Model Simulations of the East African Climate using NCAR Regional Climate Model: Influence of large-scale Orography on the Turkana Low-level Jet. Journal of Climate, 14, 2710-2724.

Jury, M., R., 2002: Economic Impacts of Climate Variability in South Africa and Development of Resource Prediction Models. Journal of Applied Meteorology, 41, 46-55.

Jury, M. R., H. M. Mulenga, and S. J. Mason, 1999: Exploratory long-range models to estimate summer climate variability over southern Africa. Journal of Climate, 12, 1892-9.

Landman, W. A. and S. J. Mason, 1999: Operational long-lead prediction of South African rainfall using canonical correlation analysis. International Journal of Climatology, 19, 1073-90.

Landman, W. A. and G. L., 2001: Statistical recalibration of GCM forecasts over southern Africa using model output statistics. Journal of Climate, 15, 2038-2055.

Landman, W. A., S. J. Mason, and W. J. Tennant, 2001: Statistical downscaling of GCM simulations to streamflow. Journal of Hydrology, 252, 221-236.

Makarau, A. and M. R. Jury, 1997: Predictability of Zimbabwe summer rainfall. International Journal of Climatology, 17, 1421-32.

Mason, S. J., 1998: Seasonal forecasting of South African rainfall using a non-linear discriminant analysis model. International Journal of Climatology, 18, 147-64.

Mutai, C. C. and M. N. Ward, 2000: East African Rainfall and the Tropical Circulation/Convection at Interannual and Intraseasonal Timescales. Journal of Climate, 13, 3915-3939.

Mutai, C. C., M. N. Ward, and A. W. Colman, 1998a: Towards the prediction of the East Africa short rains based on sea-surface temperature-atmosphere coupling. International Journal of Climatology, 18, 975-97.

Mutai, C. C., M. N. Ward, and C. A.W., 1998b: Prediction of East Africa Short Rains Rooted in Evidence for Widespread SST-forced Variability During October-December. International Journal of Climatolgy, 18, 975-997.

Nicholson, S. E. and J. Kim, 1997: The Relationship of the El Nino-Southern Oscillation to Africa Rainfall. International Journal of Climatolgy, 17, 345-375.

Nicholson, S. E., M. L. Davenport, and A. R. Malo, 1990: A comparison of the vegetation response to rainfall in the Sahel and East Africa, using Normalized Difference Vegetation Index from NOAA AVHRR. Climatic Change, 17, 209-241.

Pinzón, J. E.,   M. Brown, and C. J. Tucker (2002). "Nonparametric solar zenith angle correction of NDVI time series from NOAA-AVHRR radiometers using empirical mode decomposition." IEEE Transactions on Geoscience and Remote Sensing, submitted.

Pinzón, J. E. (2002). "Using HHT to successfully uncouple seasonal and interannual components in remotely sensed data." SCI 2002 Proceedings, July 14 - 18, Orlando, FL.

Pinzón, J. E., J. F. Pierce, and C. J. Tucker (2001).   "Analysis of remote sensing data using Hilbert-Huang transform." SCI 2001 Proceedings, July 22 - 25, Orlando, FL.

Sankarasubramanian, A. and U. Lall, 2003: Flood Quantiles and Changing Climate: Seasonal forecasts and causal relations. Water Resources Research, 39.

Senay, G. B. and J. Verdin, 2001: Using a GIS-Based Water Balance Model to Assess Regional Crop Performance. Fifth International Workshop on Application of Remote   Sensing in Hydrology, Montpellier, France, UNESCO IAHS.

Tennant, W., 1999: Numerical forecasting of monthly climate in southern Africa. International Journal of Climatology, 19, 1319-36.

Verdin, J. and R. Klaver, 2002a: Grid cell based crop water accounting for the Famine Early Warning System. Hydrological Processes, 16.

Verdin, J., C. Funk, J. Klaver, and D. Roberts, 1999: Exploring the correlation between Southern African NDVI and ENSO sea surface temperatures: results for the 1998 growing season. International Journal of Remote Sensing, 20, 2117-2124.

Ward, N., 1998: Diagnosis and short-lead time prediction of summer rainfall in tropical North Africa at interannual and multidecadal timescales. Journal of Climate, 11, 3167-91.

Xie, p. and P. A. Arkin, 1992: A 17 year Monthly Analysis based on gauge observations, satellite estimates, and numerical model outputs. Bulletin of the American Meteorological Society, 78, 2539-58.

Xie, P. and P. A. Arkin, 1997: Global Precipitation: A 17-Year Monthly Analysis Based on Gauge Observations, Satellite Estimates and Numerical Model Outputs. Bulletin of the American Meteorological Society, 78, 2539-2558.

Xie, p. and P. A. Arkin, 1998: Global monthly precipitation estimates from satellite-observed outgoing longwave radiation. Journal of Climate, 11, 137-64.