The results presented in this study are preliminary and utilize a technique still in development.

Introduction

Meteorologists convening at climate outlook forums in Africa have developed a strategy for forecasting the precipitation for the upcoming rainfall seasons. The results of these forums are probabilistic forecasts of rainfall being above-normal, normal, or below-normal. Research has been undertaken to develop applications for these probabilistic forecasts in an attempt to help predict the likely outcomes of the coming rainfall season. This research will attempt to fuse information from historical precipitation models, evapotranspiration models, and crop-specific properties to develop forecast estimates of a crop’s water needs being met by precipitation.

For southern Africa the outlook is produced as two separate three-month probabilistic forecasts for the October-November-December (OND) period and the January-February-March (JFM) period. Unfortunately for much of the region the growth cycle of any crop will span portions of these two periods. It follows then that the forecast for an individual crop will be some combination of the forecast for each of the two periods.

The Collaborative Historical African Rainfall Model (CHARM) provides a reasonable historical model for monthly precipitation over the entire continent at a 0.1-degree spatial resolution. The monthly accumulations calculated by CHARM can be temporally decomposed to provide dekadal rainfall estimates which are constrained to match the monthly sums. The extensive historical information provided by CHARM allows for an assessment of 36 years of precipitation data. Incorporation of the probabilistic forecasts with known seasonal information can result in a product which users can explore the relationship between explicit rainfall accumulations and the likelihood of occurrence.

The goal of this research is to use the seasonal precipitation information provided by the CHARM data to weight the tri-monthly forecasts for southern Africa and arrive at a spatially variable crop-cycle probability forecast. These spatially varying forecasts can be blended with distribution parameters developed from the CHARM dataset to arrive at new distribution parameters describing the probability distribution function for rainfall accumulation for the coming rainy season.

This research can be separated into two distinct activities. The first is the use of the CHARM dataset to collect seasonal totals for 35 seasons (1961-62 – 1995-96) which are used to estimate distribution parameters as well as determine the appropriate weights for blending the tri-monthly forecasts. The second activity is the modifying of the forecasts and generating new distribution parameters that reflect the forecast probabilities.

The first activity involves using the CHARM data to determine the start-of-season (SOS) at each location. The SOS is defined as the first dekad that receives 25-mm followed by two dekads with a sum of 20-mm. Scanning for the SOS began with the first dekad of October to try and capture only those regions that would be covered by the SARCOF forecast. With the SOS defined for each of the 35 growing seasons in the CHARM datset, rainfall is then accumulated for the appropriate number of dekads in the growing period. This study used periods of 8, 9, 11, 12, and 15 dekads to coincide with crop information for different strains of maize, millet and sorghum. For each of the 35 seasons, the percentage of growing period accumulation that fell during OND was recorded. The average of this value for the 35 seasons determined the weight of the OND forecast in the seasonal blending. One note about this is that the 8-dekad weight would be the largest of all the growing period weights at a location. Since the SOS is constant, a longer season will naturally have more rainfall in the JFM period because a larger amount of the growing season occurs after the OND period. The weights will reflect this and give the OND forecast less weight in the blending.

Based on the SOS information it is possible to accumulate rainfall for the given interval and arrive at a growth period accumulation. This is defined as the amount of rainfall that occurs during the crop cycle following the SOS. So while the growth period total is not crop specific, in that the total for 90-day maize will be identical to the total for 90-day sorghum, it is specific to the length of the growing period, in that the total for 90-day maize will be different from 120-day maize. With the growth period total collected for each of the 35 seasons in the CHARM climatology, gamma distribution parameters can be estimated at each location to describe the precipitation likelihood.

Integrating the water requirement weights with the forecasts is a clean process. Since the weights at any location are constrained to sum to 1.0, and the periodic forecasts sum to 100%, the result of the weighted forecast is a crop cycle forecast that sums to 100%. The resulting equation (1) represents a fusion of the two forecasts with the resulting crop cycle forecast equaling the weighted sum of the original periodic forecasts.

In this equation CCF is the spatially-dependent Crop Cycle Forecast. The spatially-dependent and crop-dependent weights derived for each period from the water requirement data are represented by the w term. Finally, the climate outlook forum forecast for each period are represented by the F term on the right of the equation.

The CCF can then be used in a Forecast Interpretation Tool (FIT) algorithm that uses Monte Carlo techniques to adjust the crop cycle distribution parameters to match the forecast probabilities. The result of this is a set of updated parameters that can describe the precipitation probability distribution function for the upcoming season, incorporating the forecast probabilities and the length of the growing season of the desired crop. These updated parameters are crop specific, as the periodic weights are specific to a certain growing period type. Implementation of the parameters can provide farmers with an idea of the probability of their crop requirements being met.

This research will investigate two outcomes of the CCF for two different growing period lengths. The first line of investigation will analyze the shift of the median of the distribution. The median is an approximate idea of the seasonal expectations as it is the level that is expected to be exceeded half the time. The second line of investigation is to show the change in probability of receiving more than 400mm in the growth period. This 400mm level is selected as an approximate crop need for maize, although the actual need is dependent on many other variables. By investigating this probability it is possible to see how the forecast will affect the probability of crop requirements being met, and thus, an effective crop yield. This analysis will be performed for the 90-day and 120-day growing periods.

An initial look at the properties of the region for the different growing seasons based on the CHARM climatology will serve as a strong introduction to the characteristics of this area. Figure 1 shows the median rainfall for the two crop cycle lengths. These median values are based on the climatologic conditions and will be altered by the forecast. The impact of the forecast on these values will depend on the distribution at each location and how the forecast alters the distribution parameters.

Figure 1. Median rainfall for nine-dekad (left) and 12-dekad (right) growing periods for southern Africa. (note: the western side of the continent has been shaved because there was no crop information for those locations)

In looking at the maps there is a noticeable swath of higher rain extending from the southeast to the northwest of this region. The topographic regions are well defined, as is the increase in precipitation over the additional 30 days of the longer growing period. It should be noted that the patches of high rainfall in the southwest are artifacts of the distribution estimation parameters as the probability of receiving no rainfall is not calculated included as a part of this study. For much of this study area the probability of a completely dry event is zero, or nearly zero, for a nine-dekad growing period or longer.

An alternative method to evaluating the ability of a location to sustain agriculture is to measure the likelihood it has of exceeding the demands of the crop. For this research a crop need of 400mm is assumed, because it is an approximate average for the nine-dekad growing cycle. The probability of exceeding this value for a 90-day and 120-day growing period is mapped in Figure 2.

Figure 2. The probability of precipitation exceeding 400mm for a nine-dekad (left) and 12-dekad (right) crop growing period for southern Africa.

The maps in Figure 2 clearly show the rain swath through the tropics and extending into the eastern part of South Africa. Areas in black will rarely or never achieve 400mm in a season according to the CHARM. In the cyan/green locations it would be expected that have the seasons will achieve 400mm, while the red areas will achieve 400mm in nearly every season. From the map it is clear that many areas will consistently receive this base amount of rainfall, although it should be noted that in these regions the crop demand is typically greater due to more evapotranspiration. The location which receive marginal rainfall are those which have poor crop yields leaving the people of that area in need of aid. How the forecast affects these probabilities will, in some ways, indicate the security of these locations for the coming season. If a location dramatically improves its chances of receiving 400mm it is more likely that there will be good yields there, while a location that sees a marked decrease in this probability, combined with a marginal climatic chance, will be a candidate for aid.

The maps and implications of the forecast for 2004-05 will be presented in the next section. Discussion of the similarities and differences of the output distributions from climatology will also be defined.

The results of this section show that the forecasts describe probabilities that differ very little from climatology. In this section, the results of the difference in median and probability of being over 400mm will be discussed for the 90-day and 120-day growing periods. The discussion will focus on the areas most severely impacted by the forecasts as well as the similarity these forecasts have to climatology in a practical sense.

Maps presented in this section represent adjustments to climatologic conditions presented in the previous section. These adjustments are a function of the forecast weights – which are themselves a function of SOS and distribution of rainfall within a specified period – the probabilistic forecasts and the seasonal precipitation defined by the distribution parameters. Two locations with the same probabilistic forecasts for OND and JFM may show different impacts resulting from the forecast weights or the rainfall distribution that are unique to each location. The presented maps show the spatial effects of the forecast and highlight those areas most seriously impacted by the SARCOF forecasts.

The change in the median (Figure 3) value was generally small throughout southern Africa with the largest anomalies being an increase of about 20mm in areas of northern Zimbabwe, Zambia and Western Tanzania. Notable decreases of nearly 20mm occurred in western Tanzania, Malawi, southern Zimbabwe and throughout South Africa.

Figure 3. The change in median rainfall, in millimeters, as a result of the SARCOF forecasts based on a nine-dekad (left) and 12-dekad (right) growing period.

The difference between the shorter and longer growth cycles manifests itself in different ways for different locations. This manifestation is a combination of the distribution parameters for a location as well as the forecast weights, and the forecast itself.

In northern DRC there is a larger change in the median for a shorter growing period, due to the fact that the weight on the OND forecast is greater for the 90-day growing period, and the OND forecast calls for increased chances of below-normal rains, while the JFM forecast calls for increased chances of above-normal rains, balancing out the OND forecast for a longer growing period.

Similarly, for the large blue swath comprising southern Botswana, northern South Africa, southern Zimbabwe and a strip of southern Mozambique, the effects of a dry forecast for the OND and JFM periods sharply contrasts with the surrounding areas calling for a dry OND followed by a wet JFM. The SOS determines how heavily to weight the different forecasts and these weights, as well as the distribution parameters, are evident in the resulting change in dryness.

Tanzania, is an example of mixed forecasts. This country is sharply split between wet and dry forecasts for the OND and JFM periods. Areas with the largest increase (western parts) or decrease (eastern parts) show locations which had a unanimous forecast in both the OND and JFM periods. Other parts of the country with less extreme changes in the median reflect a location which had a mixture of wet forecast for one period and dry forecast for the other, resulting in a median that is closer to climatology.

Interpretation of these changes in median value should keep in mind the climatological median for each location (Figure 1). It is apparent that the locations with the largest change in the median value are also the locations with large medians to begin with. What this means is that these changes in the median, as a percentage of the median, may be rather consistent across space, or at least show smaller variance than the change in median itself.

Analysis of the change in the probability of receiving 400mm provides some insight into how the forecast will affect a locations ability to support a staple maize crop. This is being expressed as an absolute change in percentage (Figure 4), so the interpretation of this result should be done while considering the climatologic information (Figure 2).

Figure 4. Change in the probability of achieving 400mm during a 9-dekad (left) and 12-dekad (right) growing period. Values are expressed as change in probability multiplied by 100.

The maps in Figure 4 show that the forecasts have very little impact on the probability of a location achieving 400mm for much of southern Africa. An overview of the two maps shows that DRC shows some impacts for the 90-day growing cycle which are not captured in the 120-day map, and that otherwise the spatial patterns of the two products are quite similar. The similarities include a band of increased likelihood of achieving 400mm running east-west through the heart of Angola and Zambia, as well as regions of decreased likelihood in eastern Tanzania, and a larger patch containing portions of South Africa, Zimbabwe and Mozambique.

Looking at the band of increased chance in Angola and Zambia, it is noted that this region shows only a modest chance of achieving 400mm based on climatology. This is good news for these regions as an increase of 10% reflects a strong improvement in their chances of supporting staple maize over a 90- or 120- day growing period.

Alternately, the regions of decreased probability are also those locations that have only a modest chance of achieving 400mm based on climatology, so this reduction represents a significant impact on the chances of supporting staple maize. This may be especially important in the region of South Africa known as the “maize triangle”, which is a producer of much of the regions food and is showing a reduction in this likelihood.

Those areas most impacted in the analysis of achieving 400mm are those regions which show a median value close to 400mm. Regions that typically receive much larger amounts will have little impact in their chance of achieving 400mm, as the base likelihood may be 99%, so a wet forecast may only increase the chance by 0.5% or less. A dry area will also not be affected by the forecast as it may have a 1% chance of achieving 400mm, and so even a wet forecast may only increase that likelihood to 2%, a small increase. Areas that expect to receive close to 400mm should be highlighted by this analysis, and that is borne out by the maps.

Summary and Conclusions

This study seeks to blend climatic information with SARCOF forecasts to analyze the impact of the forecast on southern Africa. The first step is to use the amount of rainfall in the OND period to establish weights for the OND and JFM SARCOF forecasts and arrive at a blend of these two for different growing periods is the first step. This blended forecast is then used to alter the climatic distribution parameters using a Monte Carlo methodology known as the FIT. The output parameters can be used to analyze the effects of the forecast by linking probability and accumulation and comparing climatic conditions with forecast conditions.

Results of this study show pockets of southern Africa which have differences in the median amount, which can be used as an expectation, as a result of the forecast of up to 20mm. Areas of Tanzania, northern Zimbabwe and Zambia all show dramatic increases in the expected amount, while areas of DRC, South Africa, Mozambique and Tanzania show large reductions in the expected rainfall for the growing period.

An alternative to the median approach investigated the probability of a location receiving 400mm of rainfall, an amount used to represent the crop requirements of a staple maize crop grown in the region. This analysis revealed that only a slim band of Angola and Zambia showed an increase in the likelihood of receiving 400mm, while areas of Tanzania, southern Mozambique, southern Zimbabwe and South Africa showed decreases. Interestingly, there was very little change in the probability of receiving 400mm for much of the region.

This study shows that the SARCOF forecasts do little to alter the expectations from climatology. Anomalies of 20mm in the median over a 90-day period for locations that typically receive upwards of twenty times that amount reflect relatively little difference from the climatic conditions. The probability of achieving 400mm exposes areas that are adversely affected by the forecasts and face problems of producing typical crop yields for the coming season. In general the conditions for southern Africa look to extend from typical to slightly worse than typical.