Long-term trends and changes in rainfall magnitude and duration in a semi-arid catchment, South Africa

Changes in rainfall have implications for the hydrological cycle and water resources as rainfall is the main driver of variability in the water balance (Warburton & Schulze 2005). Future warming is likely to be greatest over the interior of semi-arid margins of the Sahara and central southern Africa (Alfaro-Pelico 2010). This will exacerbate the changes in rainfall patterns and trends in southern Africa which is inherently characterised by highly variable climate. This makes it crucial to understand changes in rainfall to provide information that can assist in planning and management of water resources. Long-term changes in rainfall should also be examined in support of efforts to adapt to climate change. Fauchereau et al. (2003) reported cumulative rainfall anomalies over the summer season which did not show a definitive trend, though variations of the interannual variability of rainfall over the 20th century were noted in southern Africa. Daron (2014) reported a decrease in autumn rainfall across most of the southern Africa region except for southwest South Africa and northern central regions with increased rainfall of up to 60 or 70 mm. Increase in summer rainfall in South Africa and Namibia, and decrease at the border between Botswana, Namibia, Zimbabwe and Zambia were also reported in the latter study. Mazvimavi (2010) analysed annual and total rainfall for the beginning and mid to end of the rainy season from 40 stations for the presence of trends using the Mann–Kendall (MK) test and quantile regression analysis in Zimbabwe. The findings did not demonstrate evidence of changes in rainfall. Jury (2013) obtained mixed-up rainfall trends in southern Africa with declining trends in eastern South Africa and Madagascar based on observed, reanalysed, and model-simulated climate data sets. Mosase & Ahiablame (2018) determined upward trends for both annual and seasonal rainfall in most parts of the Limpopo River Basin, except for the winter season which showed a decreasing trend based on the MK test.

Warburton & Schulze (2005) reported that over the latter half of the 20th century, median annual rainfall decreased markedly over the Limpopo and into the border of South Africa with Botswana. Rainfall trends analysed by Kruger & Nxumalo (2017) for the period 1921–2015 indicated an increase in rainfall over the west of South Africa and a decrease in rainfall in some places in the far north-east of South Africa. DEA (2013) obtained mixed signals for spatial patterns of rainfall indices in most seasons, but there were reductions in autumn precipitation and fewer rainy days in summer and winter in most stations in Limpopo and parts of northern Mpumalanga Provinces, South Africa. Banda et al. (2021) identified dominant decreasing trends for annual rainfall and dominant positive trends for summer season rainfall in the Rietspruit sub-basin, South Africa based on MK and Sen's slope estimator. Kabanda & Nenwiini (2013) used MK and Sen's slope estimator to analyse trends and variability of rainfall magnitude in Vhembe District, South Africa using data from 23 stations for the period 1979–2009. Dominant downward and statistically significant trends across the semi-arid zone with stations in the humid zone showing no significant trends were obtained. Odiyo et al. (2015) reported dominant decreasing seasonal and annual rainfall trends based on MK and linear regression in the Luvuvhu River Catchment.

Conventional trends analysis approaches such as linear regression mischaracterise the trends at upper and lower quantiles which poses a risk of underestimating the urgency of climate adaptation or deriving inappropriate intervention strategies (Lausier & Jain 2018). They should therefore be supplemented by other approaches such as quantile regression. Quantile regression enables determination of rainfall changes along the whole range of quantiles (low, median, and high) and hence identification of changes in low-, median-, and high-rainfall events. This study focused on determining trends and changes in annual and seasonal rainfall magnitude and duration based on MK and quantile regression in a semi-arid catchment. Though there have been studies on the determination of rainfall trends in semi-arid areas, they mostly did not apply trend analysis based on quantile regression. Quantile regression enables the determination of rainfall changes along the whole range of quantiles (low, median, and high) and hence the identification of changes in low-, median-, and high-rainfall events. Aguilar et al. (2009) pointed out that a study that examines possible changes over time of rainfall during years with low and high rainfall contributes toward the development of appropriate adaptation measures for such changes. The application of quantile regression in this study enabled the identification of heteroscedastic changes, and thus is more informative as compared to linear regression and MK which only identify monotonic trends. Furthermore, the application of more than one method for trend analysis improves the reasonableness of the results. The findings of this study contribute valuable input for use in developing measures to adapt to changes in rainfall. Analysis of seasonal rainfall trends is also limited in semi-arid areas despite the fact that agriculture in these areas is mostly rainfed and dependent on seasonal rainfall. The results of this study are therefore crucial in understanding the implications of seasonal rainfall changes on rainfed agriculture in semi-arid areas. Information on historical trends in southern Africa has been considered to be scanty and fragmented which was a major concern in the second assessment report of the Intergovernmental Panel on Climate Change (Marumbwa et al. 2019). This study therefore also contributed to closing this gap through an analysis of historic trends for annual and seasonal rainfall magnitude and duration in the semi-arid catchment in southern Africa.

The study area is the Luvuvhu River Catchment which covers a total area of 5,941 km². It is located in Limpopo Province of South Africa between the longitudes 29 °49′46.16″E and 31 °23′32.02″E and latitudes 22 °17′33.57″S and 23 °17′57.31″S (Figure 1). The catchment is within the Limpopo River Basin which is shared by South Africa, Botswana, Zimbabwe, and Mozambique. Rainfall in the Luvuvhu River Catchment is highly variable and characterised by extreme events, which affect the timing and magnitude of floods and droughts, and runoff processes. The mean annual rainfall varies from 450 mm in the low-lying plains (northern and eastern parts) (Masupha et al. 2016) to more than 1,800 mm in the mountainous areas in the south-western parts of the catchment (Figure 2). The mean annual temperature ranges from <18 °C in the mountainous regions to >24 °C in the north-eastern plains of the catchment (Moeletsi et al. 2018).

Historical rainfall data for hydrological years 1965–2015 (51 years) for nine stations (Figure 1) were obtained from the South African Weather Services (SAWS) and Lynch (2001) database. The length of time series for a proper description of climate conditions as recommended by the World Meteorological Organisation is a minimum of 30 years (Arguez & Vose 2011). In trend analysis studies, at least 50 years of rainfall data have been used in the analysis of long-term trends (for example MacKellar et al. 2014; Mulugeta et al. 2019; Song et al. 2019). The spatial distribution map is based on the mean annual rainfall for the stations in Figure 2. However, the stations used in trend analysis in the study were selected based on the availability of long-term rainfall data and with minimal or no gaps. Stations with ≤1% of missing rainfall data were considered in the study following van Wilgen et al. (2016). Most of these stations are in the upstream area of the Luvuvhu River Catchment. The data were analysed based on the hydrological years. A hydrological year in the study area starts from October of the current year and proceeds to September of the following year at annual scale. At seasonal scale, the season starts from October of the current year and proceeds to March of the following year. For example, the hydrological year 1965/1966 covers the period from October 1965 to September 1966 for the annual time scale while the seasonal time scale for the year 1965/1966 covers October 1965 to March 1966.

Trends were determined for computed total annual and seasonal rainfall magnitude and duration. Total annual rainfall, seasonal rainfall, and number of seasonal rainy days (rainfall duration) for each hydrological year were computed from daily rainfall for each station. The same approach used by Moeletsi et al. (2010) was used to determine seasonal rainfall duration. The procedure includes defining the onset and cessation of the rainy season. Onset was defined as the first day of the season in which 25 mm or more rainfall was accumulated during the previous 10 days, thereafter accumulating more than 20 mm rainfall in the subsequent 20 days (Moeletsi et al. 2010). This is used to determine the beginning of the rainy season in most semi-arid areas of South Africa. Cessation was defined as the end of the rainy season and was obtained by getting the last day in which cumulated rainfall of 25 mm over 10 days occurred. The length of the rainy season (duration of the rainy season) was determined as the difference between cessation and onset. Since rainfall duration was defined based on data from the rainy season (October–March), it was therefore not possible to do trend analysis for rainfall duration at annual scale. Thus, the study only considered seasonal rainfall duration in the trend analysis.

The direction of the trend was determined based on the slope coefficient (SC) (β(θ)₁) following Mazvimavi (2010). A negative (positive) SC indicated that the θth quantile of rainfall magnitude or rainfall duration was decreasing (increasing). The trend was significant if the computed p-value was less than α of 0.05.

The DMA results for only four (Entabeni Bos, Matiwa, Levubu, and Hanglip) of the nine stations are presented in Figure 3 for simplicity. The DMA plots are approximately straight lines and therefore indicate that the rainfall data are consistent, homogenous, and have no errors. The coefficient of determination (R²) values for the stations are above 0.99 indicating the strong linear relationship between each station and other surrounding stations. R² value near 1 indicates a strong relationship between the two variables (Ratner 2009). A study by Nkuna & Odiyo (2011) determined R² values above 0.996 for DMA results for rainfall stations in the Luvuvhu River Catchment. This is comparable with the values obtained in this study. These results, therefore, indicate that the rainfall data are homogenous, consistent, and of good quality, and can further be used for hydrological analysis.

The results of MK analysis for annual and seasonal rainfall magnitude are given in Tables 1 and 2. Klein Australie had statistically increasing trends for annual and seasonal rainfall while Entabeni Bos had statistically decreasing trends for annual rainfall magnitude. For Klein Australie, annual and seasonal rainfall increased by 6.78 and 6.51 mm, respectively, based on Sen's slope. The annual rainfall magnitude for Entabeni Bos decreased by −12.31 mm. Hanglip and Vondo Bos had non-statistically significant increasing trends of annual and seasonal rainfall magnitude. The rest of the stations had decreasing trends for annual and seasonal rainfall magnitude which were not statistically significant. Entabeni Bos, Levubu, Matiwa, and Palmaryville had statistically significant decreasing trends for the duration of seasonal rainfall (Table 3). These stations had a magnitude of change ranging from −0.80 to −0.28.

Table 4 gives the quantile regression results for annual rainfall magnitude for all nine stations. Entabeni Bos and Nooitgedacht had non-statistically significant decreasing trends in all quantiles except quantiles 0.1 and 0.2, respectively, which had non-statistically significant increasing trends. Klein Australie, Vondo Bos, and Hanglip had non-statistically significant increasing trends for low rainfall events at quantiles of 0.1, 0.2, and median annual rainfall (0.5 quantile). Matiwa, Klein Australie, and Vondo Bos had non-statistically significant increasing trends for 0.9 quantile (high-rainfall events). Seasonal rainfall magnitude at Levubu, Entabeni Bos, Matiwa, Nooitgedacht, Palmaryville, and Shefeera stations had decreasing trends that were not statistically significant in most quantiles (Table 5). Entabeni Bos and Matiwa had non-statistically increasing trends for quantiles 0.1, and 0.1 and 0.2, respectively. Palmaryville and Nooitgedacht showed non-statistically significant increasing trends on quantile 0.7. Klein Australie, Vondo Bos, and Hanglip showed non-statistically increasing trends in most quantiles for seasonal rainfall magnitude, with Vondo Bos and Hanglip showing non-statistically decreasing trends for quantiles 0.7 and 0.9, respectively.

The trends for seasonal rainfall duration for Entabeni Bos at quantiles 0.2, 0.5, and 0.7, Levubu and Palmaryville at quantile 0.2 were statistically significant (Table 6). Nooitgedacht, Matiwa, Hanglip, and Shefeera had non-statistically significant increasing trends in duration of seasonal rainfall at quantiles 0.1, 0.9, 0.7, and 0.1 and 0.2 quantiles, respectively. Vondo Bos had non-significant statistically increasing trends at low (0.1 and 0.2), median (0.5), and high (0.9) seasonal rainfall events.

The land-use changes within a 2 and 5 km radius of Sheefera and Palmaryville stations are presented in Figures 4 and 5, respectively, as examples to give an indication of the distribution of the land-use change. The land-use in the northern part of the 5-km boundary for Shefeera was not included in the study since this portion is beyond the Luvuvhu River Catchment boundary (Figure 5). This portion is within the Nzhelele River Catchment which is on the leeward side of the Soutpansberg Mountain. The Soutpansberg Mountain forms a rainfall divide between the Soutpansberg Mountain and Nzhelele River Catchment resulting in contrasting rainfall behaviour. Land-use change portions for Vondo Bos, Matiwa, and Entabeni Bos (5 km only) which were also outside the Luvuvhu River Catchment were therefore also excluded. The areas within 2 km radius of Klein Australie, Vondo Bos, Entabeni Bos, Matiwa, Nooitgedacht, and Sheefera had more than 50% of their area with no change in land use (Table 7). The dominant land use in these areas is planted forest, with 92.02% of Sheefera covered by planted forest. Other minor land uses (<10%) were not included in Tables 7 and 8, except for cases where the dominant land use was <10%. More than 50% of the land use within a 2 km radius of Hanglip, Matiwa, and Palmaryville had changed between 1990 and 2018 through the changes varied from one area to another. The land use dominantly changed from thicket/dense bush to natural wooded land, indigenous forest to thicket/dense bush, and natural wooded land to built-up residential for Hanglip, Matiwa, and Palmaryville, respectively. Palmaryville is the only station located in an area dominated by built-up residential areas and this increased by 9.98%.

Areas within the 5 km radius of the rainfall stations were dominated by no land-use change with percentage area of more than 50% for all stations, except for Hanglip (Table 8). Planted forest was the dominant land use within 5 km of Klein Australie, Vondo Bos, Entabeni Bos, Matiwa, Nooitgedacht, and Sheefera stations. It was only the area within a 5 km radius of Hanglip which had more than 50% land-use change, dominated by a change from indigenous forest to thicket/dense bush.

There was a negative correlation between annual and seasonal rainfall magnitude, rainfall duration trends with altitude. The R values ranged from −0.58 to −0.03, −0.72 to −0.003, and −024 to −0.06 (Table 9), with most of them indicating a weak negative correlation (R<−0.5). Quantile 0.9 for annual rainfall magnitude and quantile 0.1 for seasonal rainfall magnitude had a weak positive correlation (R<0.5) with R values of 0.28 and 0.31, respectively. The R values between annual and seasonal rainfall magnitude, and rainfall duration trends with percentage land-use changes mostly indicated a weak negative correlation (R<−0.5) (Table 10). Quantiles 0.7 and 0.9, quantile 0.7 for seasonal rainfall for rainfall duration as well all MK values had a weak positive correlation (R<0.5) between rainfall trends and land-use change. Rainfall trends for quantile 0.9 and land-use change had a good positive correlation (R>0.5).

MK and quantile regression had shown dominant non-statistically significant decreasing trends for long-term annual and seasonal rainfall magnitude in five stations (Levubu, Matiwa, Nooitgedacht, Shefeera, and Palmaryville) in the study area. The findings of this study are similar to those of earlier studies. Kruger & Nxumalo (2017) reported a decrease in rainfall in the northern and north-eastern parts of South Africa, where the study area is also located. Kruger (2006) reported decreases in annual precipitation in northern Limpopo Province of South Africa. Non-statistically significant trends are dominant in southern Africa due to highly variable year-to-year rainfall as indicated in studies such as DEA (2013), Chikoore (2016), and Odiyo et al. (2020). However, rainfall variability is associated with more widespread and intense droughts in southern Africa (Fauchereau et al. 2003; Tfwala et al. 2018). This increases the vulnerability of the southern Africa region to climate change impacts. This implies that although the trends were not statistically significant measures to adapt to changing climatic conditions, they are still essential to reduce climate change impacts. This is therefore crucial when making decisions related to climate change in semi-arid areas of southern Africa.

Mosase & Ahiablame (2018) reported changes in rainfall magnitude ranging from 0.001 to 0.46 mm and −0.2 to −0.0003 mm for increasing and decreasing rainfall trends, respectively, for stations within the Limpopo River Basin. These ranges differ from the values obtained in this study due to the fact that Mosase & Ahiablame (2018) covered a river basin while this study was on a catchment scale where the orographic effect of the Soutpansberg Mountains influences rainfall behaviour. Kabanda & Nenwiini (2013) obtained Sen's slope values of −12.5 mm for Entabeni Bos based on annual rainfall for 40 years of period (1970–2009). This is comparable to the value of −12.31 mm for annual rainfall magnitude for Entabeni Bos obtained in this study.

The study identified statistically significant decreasing trends for seasonal rainfall duration for four stations based on MK (Table 3 and Figure 5), whereas quantile regression (Table 6) displayed statistically significant trends for some quantiles for three stations (Entabeni Bos at 0.2, 0.5, and 0.7, and Levubu and Palmaryville at 0.2). DEA (2013) showed that there has been a marginal reduction in rainfall for the summer months in the northern part of South Africa, where the study area is located. Mackellar et al. (2014) pointed out a significant decrease in a number of rainy days over the central and north-eastern parts of South Africa. Decreasing trends for seasonal rainfall duration imply that rainfall progressively decreased during the crop growing season, thereby increasing the vulnerability of crop production and crop failure as noted by Matata et al. (2019).

There were deviations in the trends from quantile regression for low and high quantiles for some stations when compared to those obtained from MK. In cases where decreasing trends dominated, there were deviations in annual rainfall magnitude (Entabeni Bos, Nooitgedacht, and Matiwa at 0.1, 0.2, and 0.9, respectively), seasonal rainfall magnitude (Entabeni Bos and Shefeera at 0.1, Matiwa at 0.1 and 0.2, Palmaryville and Nooitgedacht at 0.7 quantiles, respectively), and seasonal rainfall duration (Nooitgedacht, Matiwa, Hanglip and Shefeera at 0.1, 0.9, 0.7, and 0.1 and 0.2 quantiles, respectively). For stations with dominant increasing trends, there were also deviations in annual rainfall magnitude (quantile 0.7 for Vondo and quantiles 0.7 and 0.9 for Hanglip), seasonal rainfall magnitude (Vondo Bos and Hanglip at quantiles 0.7 and 0.9, respectively), and rainfall duration (Vondo Bos at quantile 0.7). This indicates that low- and high-rainfall events had different behaviours as captured by quantile regression. These findings are similar to those of Lausier & Jain (2018) who indicated that rainfall trends at low and high quantiles deviated from those of the linear regression method. The low- and high-rainfall events signify drought and flood periods, respectively, meaning that if these trends are not accurately identified, improper strategies for mitigating or preventing their impacts may be developed leading to inaccurate decision-making and planning. Accurate determination of rainfall trends is also essential in understanding and supporting efforts to adapt to climate change. This is of crucial importance, particularly in the Luvuvhu River Catchment where there is evidence of changes in rainfall that are linked to climate change and variability.

The correlation analysis indicated variations (weak negative and positive correlations) in the relationship of long-term rainfall trends with both altitude and land-use change. The correlations were also not statistically significant. This indicates that there was no clear relationship between long-term rainfall trends and these variables. Several studies have indicated variations in the relationship between rainfall and altitude. Qing et al. (2011) noted that the effect of altitude on rainfall is not always positive and reported both negative and positive correlations of rainfall with altitude in different regions of China. Hu et al. (2021) found negative and positive correlations between extreme precipitation trends and altitude in the Tibetan Plateau and its surrounding area. These were linked to the complex precipitation mechanism due to prevailing atmospheric circulations, large-scale weather systems, and complex topography which affect the local patterns of precipitation. The negative correlation of summer precipitation with altitude was associated with increased global warming by Lu et al. (2008) in China. Dahri et al. (2016) noted that inadequate observations affected the determination of the relationship between altitude and precipitation due to spatial variability of precipitation in the Indus basin. Qing et al. (2011) indicated that additional topographic variables such as slope, orientation, and exposure, need to be included in predicting the relationship between precipitation trends and rainfall. This indicates that there are several factors that influence rainfall trends. Nenwiini (2017) reported that rainfall in the study area is influenced by the moisture-laden air from the Indian Ocean driven by the prevailing south easterly winds and extreme topographic diversity with altitudinal changes over short distances. The rainfall trends in this study are likely to be influenced by atmospheric circulations, and weather systems in addition to altitude, resulting in their variations. Evidence of global warming in the Luvuvhu River Catchment has been reported by Odiyo et al. (2020) and it was indicated that it is impacting the rainfall behaviour. Kabanda & Nenwiini (2013) have shown that due to diverse rainfall zones in this study area, rainfall is highly variable within local settings. This also influences rainfall trends.

The correlation of rainfall trends with percentage land-use change also varied. Since there was dominantly no change in land-use for a percentage area of more than 50% within a 5-km radius for most of the stations, it is likely that land-use change did not significantly influence changes in rainfall in these areas, as indicated by the non-statistically significant correlations. Liang et al. (2019) noted that detecting the impact of land-cover change on local rainfall is difficult due to multiple environmental factors (including advective atmospheric transport, local evapotranspiration, and large-scale climate drivers) that cannot be strictly controlled. Jia et al. (2019) also noted that land and climate interact in complex ways through changes in forcing and multiple biophysical and biogeochemical feedbacks across different spatial and temporal scales. These may also be the cause of the variations in the correlation between rainfall trends and percentage land-use change in the study area. The findings of this study imply that land-use change and topography are therefore just some of the factors influencing rainfall behaviour and in some cases, they may not be adequate to explain the variations of long-term rainfall trends if analysed alone. Detailed studies to understand the interaction of various environmental factors and their influence on rainfall trends are therefore required, though these may be limited by data availability which is typical in semi-arid areas.

The decreasing rainfall trends identified in this study will impact streamflow and water availability for both agriculture and domestic uses. Reduced rainfall in upstream river basin systems affects the availability of water for downstream systems and uses for various economic activities (Nhemachena et al. 2020). Most of the rainfall stations in this study area are in the upstream area of the Luvuvhu River Catchment implying that downstream water uses will be affected due to decreased rainfall. This will be worsened by the semi-arid nature and highly variable rainfall distribution within the Luvuvhu River Catchment. Commercial and subsistence (small scale) forms of farming which rely on water resources within the Luvuvhu River Catchment are therefore vulnerable. The vulnerability of subsistence farming will threaten rural livelihoods and food security. Odiyo et al. (2020) reported that some of the regional water supply schemes in the Luvuvhu River Catchment are already in deficit. The communities that are highly dependent on water supply from these regional schemes will therefore suffer from acute water shortages.

Long-term seasonal and annual changes in rainfall magnitude and duration were determined based on MK and quantile regression methods. Dominant decreasing trends in long-term annual and seasonal rainfall magnitude and seasonal rainfall duration were identified in the study area. This indicated the reduction of annual and seasonal rainfall magnitude and rainfall duration in most of the stations. Decreasing rainfall trends will reduce water availability for agricultural water use and domestic water supply. The results of trends from quantile regression at low (0.1 and 0.2) and high (0.7 and 0.9) quantiles deviated from those of MK for most of the stations. Decision-making should therefore consider that though decreasing trends are dominant, there will be deviations for low and high rainfall events which need to be accounted for. The study indicated varying (weak negative and positive) correlations of long-term rainfall trends with both altitude and land-use change. This implied that land-use change and topography may not be adequate to explain the variations of long-term rainfall trends if analysed alone since there is an interaction of other environmental factors that influence rainfall behaviour. Further studies are therefore required to understand the interaction of these factors and their influence on rainfall trends in semi-arid areas.

South African Weather Services is acknowledged for providing rainfall data used in this study.

The authors declare no conflict of interest.

Part of the results of the study were generated in a Masters study funded by the National Research Foundation (NRF), grant number AEMD160625174151.

Data cannot be made publicly available; readers should contact the corresponding author for details.