Rain gauge dataset from 8 319 rainfall stations was used to interpolate a rainfall surface using Empirical Bayesian Kriging (EBK) Regression Prediction geostatistical analysis tool available in ArcGIS Pro 2.7. The rain gauge dataset was received courtesy of the University of KwaZulu-Natal (Lynch 2004) and the dataset was last updated in 2002. Some rainfall stations had data going back to 1871 till the year 2001. Only those rainfall stations that have been active for long enough to be representative of the long-term mean were included (Lynch & Dent 1990). Mean annual rainfall values were extracted for each station and used as the dependent variable in the EBK Regression Prediction tool. This tool was selected as it is a recent and robust geostatistical method which considers local variation in explanatory variables and is in essence a hybrid between regression and kriging approaches to interpolation. To apply EBK Regression Prediction to the interpolation of Mean Annual Precipitation, six explanatory variables were used to predict precipitation at local scales. The explanatory variables that were selected included standard geographic and topographic variables known to influence rainfall. These datasets were either derived from the Shuttle Radar Topography Mission (SRTM) 90 m Digital Elevation Model (DEM) or created with the same resolution of 90 m pixels. Continentality refers to the climatic phenomenon whereby locations further inland are isolated from the climatic influence of the oceans. With regard to rainfall, the general pattern is that interior areas are further from the moist oceanic source of atmospheric water and are therefore drier. Continentality was included as a simple distance from the coastline, measured in km (a). Latitude in this instance simply represents the distance south from the equator (b). Most generally, rainfall decreases with increasing latitude, and particularly in tropical regions, latitude can influence the duration and intensity of rainfall. Altitude is the height of a point relative to mean sea level (c). The general pattern is that rainfall increases with altitude, as atmospheric water precipitates due to the effects of altitude, increasing altitude decreases air temperature and air density. The downscaling of the South African rainfall surface was a requirement for the fine-scale delineation of the national Strategic Water Source Areas (SWSAs). Because of the shared water catchments that feed water supply systems in South Africa, this surface was generated across South Africa, Lesotho and eSwatini. To incorporate the local effects of topography on rainfall, the Topographic Position Index (TPI) was also included as an explanatory variable. The TPI compares the elevation of each cell in a DEM to the mean elevation of a defined neighbourhood surrounding that cell. By doing so it calculates the degree to which a location is higher or lower than its surrounds and provides a way to highlight mountain ridges and valley bottoms. Positive TPI values represent locations that are higher than the average of their surroundings (mountains), negative values represent locations that are lower than their surroundings (valleys), and values near zero are flat areas. The inclusion of TPI is based on the concept of orographic precipitation where moist air precipitates as it is cooled while rising over a mountain range. Topography theoretically has a strong effect on orographic rainfall, whereby moist air rising over the windward side of a mountain will precipitate rainfall and the drier air on the leeward side creates a rain shadow. As TPI is strongly scale-dependent, it was calculated from the 90 m SRTM DEM using both a 50 km (d) and 250 km search radius (e). By default, the EBK Regression Prediction tool will create random overlapping polygons that incorporate on average around 500 sampling points. Each of these subset polygons will have its own semivariogram to model precipitation. However, the default is a random dataset, and it is recommended to rather select more representative subset polygons that are based on hydrologically or ecologically meaningful units. We used various model performance metrics to determine that the South African Weather Service Rainfall Districts produced the best result. Model performance was assessed using two primary model fit statistics, the Root Mean Square Error (RMSE) and the Continuous Rank Probability Score (CRPS). Both of these statistics are crossvalidation procedures, through which a data point is removed, then the model run with the remaining data points, and then the predicted value for the excluded data point is compared to its actual value. This is done for all the measurements. In both of these statistics, a lower value indicates better model fit. The CRPS is the most important model fit statistic for this analysis, and it was ultimately used to select the best performing model. The RMSE is useful as it is measured in the same units as the response variable, in this case mm precipitation per annum. The EBK Regression Prediction in ArcGIS also provides a Standard Error surface, which shows where the predictions are less certain. The final model with the selected parameters provided an improvement in model fit in comparison to other interpolation options tested. For comparison, using Simple Kriging resulted in an RMSE of 78 mm. Applying Empirical Bayesian Kriging improved this to 74 mm and the final EBK Regression Prediction had an error of only 68 mm. The EBK Regression Prediction process also provides a Standard Error surface that shows the uncertainty related to the predicted values (side graphic). The downscaled rainfall surface was refined as part of the work of the Biodiversity and Land Use (BLU) project which was implemented by SANBI in partnership with a range of public and private organisations and supported by the United Nations Development Programme (UNDP

Strategic Water Source Areas (SWSAs) refer to the 10% of South Africa’s land area that provides a disproportionate 50% of the country’s water runoff. Understanding where these SWSAs are is crucial to planning and management of water resources, including the ecosystems that support water quality and quantity. These areas extend into Lesotho and eSwatini. National SWSAs for surface water have been delineated in various forms over the past 15 years, with increasing precision in each iteration. In 2018, 22 SWSAs were identified based on a generalised 1.7 x 1.7 km resolution Mean Annual Runoff dataset, providing a widely accepted product that gained strong traction with government and non-government audiences, proving effective for building awareness and integrating SWSAs in a range of national policies and frameworks. However, the coarse resolution does not align well with the scales used for implementation at catchment and local levels. So, using best available information and the latest geostatistical approaches, South Africa’s SWSAs for surface water have now been delineated at a finer resolution of 90 x 90 m. The work, which concluded in 2021 resulted in two products (both explained in greater detail in Lotter & Le Maitre (2021)): 1. A downscaled mean annual precipitation surface: through consideration of several explanatory variables in the modelling process (such as elevation, latitude, distance from coast, topographical positional index, and Mean Annual Runoff), the “new” Mean Annual Precipitation surface layer was “interpolated” using a dataset of over 8000 rainfall stations and a Shuttle Radar Topography Mission (SRTM) 90 m Digital Elevation Model (DEM) (available as a stand-alone product). 2. 2021 SWSA layer: This precipitation surface layer was used to delineate fine-scale SWSA boundaries, which were compared with the older 2018 SWSAs for surface water to maintain the 22 SWSAs with similar extent and location to those identified in the 2018 SWSAs. Delineating SWSAs at a finer scale was done across South Africa, Lesotho and eSwatini because of the shared water catchments that feed into water supply systems in South Africa. Both above mentioned layers therefore extend across the three countries.