Several sources of operational climate forecasts, disseminated by both local and international institutions, were considered for use in this study. As a number of historical seasons were to be used in developing and testing the yield forecasting methodology, an archive of previously disseminated forecasts was required. The suitability of the forecasts was assessed in terms of several potentially limiting factors, these being, the lead time (seasonal), the number of historical seasons archived and the availability of corresponding observed daily data for use in forecast downscaling and verification.
The most suitable set of seasonal climate forecasts found was that produced by Landman and Klopper (1998) for the 1981/1982 to 1995/1996 seasons. These forecasts of seasonal rainfall were produced in an exercise to validate the statistical rainfall model used by SAWS in its operational forecasts. This model was developed using canonical correlation analysis, a regression-based technique considered to be at the top of the regression modeling hierarchy (Barnett and Preisendorfer 1987). The predictand in the model is December to March (summer) rainfall and the predictors are sea surface temperature anomalies from the global oceans between 45° N and 45° S for each of the four preceding three-month seasons (Landman and Klopper 1998). For large parts of the country, the December to March period forms a major portion of the annual rainfall (Tyson 1986; Schulze 1997). The rainfall model has not changed significantly since the study of Landman and Klopper (1998), apart from ongoing refinements. However, the format in which the forecasts are presented has changed from deterministic (single possible outcome) to probabilistic (three possible outcomes) format. Although the use of probabilistic forecasts is now generally encouraged in applications research (because it conveys associated risk), it was considered appropriate to use the deterministic forecasts published in Landman and Klopper (1998), as a relatively large number of seasons (15) were represented. For only three of these seasons (1993/1994 to 1995/1996), corresponding observed daily rainfall data were not readily available for use in this study. The rainfall forecasts were categorical in that rainfall was forecast to be either below normal, near normal or above normal. The rainfall forecasts were made for six regions of relatively homogenous (summer) rainfall distribution, these regions covering most of the country. The regions were originally defined by Mason (1998) and then updated by Landman and Klopper (1998). The observed categorical rainfall for each region and forecast period was determined by Landman and Klopper (1998).
The rainfall forecasts required both spatial and temporal downscaling in order to develop rainfall inputs to the CERES-Maize model. In the spatial domain, the rainfall forecasts needed to be downscaled from relatively large rainfall regions to QC scale, while in the temporal domain, they required downscaling from categorical rainfall for four month periods to daily rainfall values. Different methods of downscaling the rainfall forecasts were considered. The analogue season downscaling technique was selected as it is a relatively simple and robust technique in which the authors had previous experience in applying (Hallowes et al. 1999; Lumsden et al. 1999). The data/information required to apply this technique were available, whereas this was not the case for some other methods (e.g. applying a stochastic rainfall generator). The procedure adopted for downscaling rainfall forecasts (for the 1981/1982 to 1992/1993 seasons) was as follows: The categorical rainfall forecasts for a particular rainfall region were assumed to apply to each of the catchments falling within that region, i.e. an above normal forecast for the rainfall region implied an above normal forecast for each of the catchments in that region. The seasonal (December to March) forecasts for a catchment were then downscaled to daily values of rainfall by selecting all historical seasons in that catchment's rainfall record that represented the forecast concerned, i.e. if the forecast for a season was for above-normal rainfall, then all historical seasons experiencing above-normal rainfall were selected to represent the rainfall record for that season. For each catchment, a single rainfall station having a rainfall record representative of the catchment, had previously been selected. Above normal, near normal and below normal classes of rainfall corresponded to the upper, middle and lower terciles, respectively, of the long-term probability distribution of seasonal (December to March) rainfall. For each catchment, thirty seasons (1950/1951 to 1979/1980) of observed rainfall were extracted from the QC climate database to serve as analogues to represent the above normal, near normal and below normal rainfall terciles. Data were extracted for this period as it allowed for an equal number of seasons to represent each tercile, i.e. ten seasons per tercile. Only seasons prior to the first season forecasted (1981/1982) were considered for use as analogue seasons. When preparing CERES-Maize climate input files to represent a seasonal forecast, ten individual climate files were created, each of these corresponding to a different analogue season. This implied that there were multiple yield outcomes for a season, which could then be considered a forecast yield distribution.
Was this article helpful?