South African winter rainfall zone shifts: A comparison of seasonality metrics for Cape Town from 1841–1899 and 1933–2020

Mounting evidence across South Africa’s southwestern winter rainfall zone (WRZ) reflects consistent drying since ~ 1980, and projected trends suggest this will continue. However, limited evidence exists for changes in the region’s rainfall seasonality. To improve our understanding of these WRZ drying trends, especially within the context of Cape Town’s 2015–2017 “Day Zero” drought, it is necessary to explore long-term rainfall seasonality trends. Thus, we use the longest WRZ meteorological record from the South African Astronomical Observatory (SAAO) in Cape Town to investigate rainfall seasonality shifts during 1841–2020. Consistent with recorded poleward migrations of the subtropical high-pressure belt and mid-latitude westerlies, known drivers behind the drought and drying trends, calculated trends demonstrate strengthening of WRZ conditions, primarily from a later start-date trend leading to a shorter wet-season. Long-term drying trends are quantified for the wet and dry seasons; however, analysis of trend evolution reveals much variability, reflecting that drying has only persisted since ~ 1892. Comparative analyses of the first and last 59 years of 1841–2020 reveal a rainfall decline of almost 10% across both seasons—highlighting that the extreme “Day Zero” drought was not only driven by wet-season rainfall declines. Results demonstrate that these drying trends were consistently driven by a long-term decline in rain day counts and a more recent decline in average rainfall per rain day. Correspondence between our results and projected rainfall seasonality trends suggests the trends we quantified will likely continue; thus, improvements and continuation of existing water conservation and management strategies are imperative for Cape Town.


Introduction
The southwestern region of South Africa's winter rainfall zone (WRZ; Fig. 1) experienced below-average rainfall during the 2015-2017 April-September winter wet seasons, which resulted in the worst drought and water shortages across the region since 1904 (Botai et al. 2017;Wolski 2018). These deficits, which were most pronounced for the autumn (March-May) and spring transition (September-November) seasons, led to supply dam water levels dropping to ~ 20% capacity during May 2018 (Burls et al. 2019;Pascale et al. 2020). For the ~ 4 million inhabitants of the City of Cape Town, this culminated in an imminent threat that dam levels would fall below ~ 10%, marking the level at which the city's municipal water supply would have been disconnected, before the 2018 winter wet season (Sousa et al. 2018). The fear surrounding this gathered much attention, with media and municipal authorities terming this event "Day Zero", and forced the city to enforce strict water restrictions, from pre-drought levels of ~ 200L per person per day to only 50L per person per day (Muller 2018;Wolski 2018). Fortunately, "Day Zero" was averted as good early winter rains and near-average rainfall for the 2018 winter season led to dam levels rising to ~ 70% by October 2018 (Sousa et al. 2018;Burls et al. 2019). This crisis highlighted that Cape Town is extremely vulnerable to multi-year droughts, which are expected to increase in frequency and magnitude under anthropogenically induced climate change (Otto et al. 2018;Mahlalela et al. 2019;Pascale et al. 2020).
Many studies have investigated climatological factors driving the 2015-2017 drought. Although poor watermanagement practices and infrastructure deficiencies exacerbated this water crisis (Muller 2018), the 2015-2017 rainfall deficit was the main driver (Otto et al. 2018;Wolski et al. 2021). Several studies demonstrate that this rainfall deficit, particularly during the transition seasons, corresponds to trends of annual and seasonal rainfall declines, which are more pronounced for recent decades, from ~ 1980 (du Plessis and Schloms 2017; Sousa et al. 2018;Mahlalela et al. 2019;Jury 2020;Ndebele et al. 2020;Wolski et al. 2021). Odoulami et al. (2021) demonstrate that the 2015-2017 rainfall deficits correspond to a decrease in winter rain-bearing circulation types, including cold fronts, and an increase in dry circulation types, including the South Atlantic Anticyclone (SAA) and associated ridging anticyclones. These may form part of a trend in changing regional synoptic climatology, quantified by Lennard and Hegerl (2015) to have started in ~ 1979. Moisture transport to the southwestern Cape region and the amount of rain falling on days with rain-bearing circulation types was also demonstrated to be reduced during the period 2015-2017 (Odoulami et al. 2021). Burls et al. (2019) found similar results, with a decline, since 1979, in the duration of rainfall events and rainfall amount associated with cold fronts for the April-September winter wet season. It is established that these changes are linked to hemispheric-scale expansion of the Hadley cell, which has in turn driven a poleward expansion of the subtropical high-pressure belt and a poleward displacement of the westerlies moisture corridor (Sousa et al. 2018;Burls et al. 2019;Mahlalela et al. 2019;Odoulami et al. 2021).
Despite numerous studies exploring rainfall trends and the underlying mechanisms that drove the Cape Town "Day Zero" drought, there is a limited understanding of rainfall seasonality trends, such as wet-season start-and end-date and duration trends, which may have contributed to the occurrence of this drought. To understand this, it is necessary to explore long-term rainfall seasonality changes. Although recent research investigated rainfall seasonality trends for several WRZ locations including Cape Town (Roffe et al. 2021a), the analysis period of 1987-2016 is too short and ends too early to effectively understand how rainfall seasonality trends contributed to driving this drought. Derived from a ratio of monthly rainfall/temperature (Roffe et al. 2021b) and a percentile-based seasonality metric (Roffe et al. 2020) and spanning 1841-2020, we therefore present a long-term rainfall seasonality record for the South African Astronomical Observatory (SAAO) located in Cape Town (Fig. 1), the core region of the drought and WRZ. Using this record, we explore how rainfall seasonality has shifted since 1841 as a means to better understand Cape Town's rainfall seasonality dynamics in relation to the 2015-2017 "Day Zero" water crisis. Given the occurrence of this recent extreme drought and since projections robustly reflect that such droughts will occur more frequently and with higher intensity in the future (Pascale et al. 2020), there is a clear need for such a study exploring long-term rainfall seasonality trends as this can inform management strategies for Fig. 1 Location of the South African Astronomical Observatory (SAAO) meteorological station, at 33°56′4"S, 18°28′38.3"E and 12 m.asl, in relation to the Cape Town City centre within the Western Cape of South Africa, and within the winter rainfall zone (WRZ estimated boundary is from Roffe et al. 2021b). The topography across the region and Cape Town's main water supply dams are also depicted rainfall sensitive activities (e.g. water resource management) across Cape Town.

Study area: Cape Town's rainfall climatology
While most of South Africa is characterized by a summer rainfall regime, Cape Town is situated southwest within the country's spatially heterogeneous WRZ, which extends across the southwestern Cape and north along the west coast to the Namibia border ( Fig. 1; Roffe et al. 2021b). Hence, Cape Town experiences a winter rainfall regime, typically termed a Mediterranean-like climate , with warm, dry summers and cool, wet winters, receiving ~ 80% of annual rainfall during extended winter months, from April-September (Mahlalela et al. 2019). Located windward of the Cape Fold Mountain Belt ( Fig. 1), Cape Town receives an annual rainfall total averaging ~ 600 mm. yr −1 (Ndebele et al. 2020), which varies interannually in relation to the strength and position of the westerlies-the main winter rainfall moisture source (Sousa et al. 2018). During winter, when the westerlies are situated farther north, mid-latitude cyclone cold fronts, tracking eastwards from the southwest Atlantic Ocean, more frequently reach Cape Town and are responsible for ~ 89% of winter rainfall across the southwestern Cape region (Burls et al. 2019). Frontless troughs in the westerly wave, which sometimes break off forming cut-off lows, provide much of the remaining rainfall, primarily during autumn and spring months (Favre et al. 2013;Omar and Abiodun 2020). Long, narrow moisture plumes, termed atmospheric rivers, also produce some of the remaining rainfall . As the SAA and westerlies shift south during summer, fewer temperate weather systems reach Cape Town and the surrounding regions, and warm, dry conditions are promoted by subsidence associated with the SAA (Sousa et al. 2018;Ndarana et al. 2021). SAA ridging anticyclones, occurring year-round but more frequently during warmer months, also typically promote dry summer conditions across the southwestern Cape (Mahlalela et al. 2019;Ndarana et al. 2021). Although the summer period is dry, some rain is delivered to Cape Town and the surrounding regions from convective storms (Lennard and Hegerl 2015), cut-off lows (Favre et al. 2013), and atmospheric rivers (De Kock et al. 2021).
Winter rainfall across Cape Town and the southwestern Cape displays considerable interannual to interdecadal variability, making it susceptible to dry periods (Dieppois et al. 2016;Sousa et al. 2018;Mahlalela et al. 2019). This is driven by several factors, including anomalies in the South Atlantic sea surface temperatures and Southern Ocean sea ice (Reason et al. 2002;Reason and Jagadheesa 2005;Blamey and Reason 2007), and large-scale modes of variability including the Southern Annular Mode (Reason and Rouault 2005;Mahlalela et al. 2019) and El Niño Southern Oscillation (Reason et al. 2000;Philippon et al. 2012). These factors influence the position and strength of the westerlies storm track and associated moisture fluxes to the region (Sousa et al. 2018), which in turn determines rainfall variability primarily through changes in the number and intensity of cold fronts making landfall (Burls et al. 2019).

Data and pre-processing
Records of daily rainfall and temperature, spanning 1841-1899 and 1933-2020, from the SAAO meteorological station were used to explore rainfall seasonality shifts. Data for 1841-1899 were digitized from meteorological registers held at the SAAO (see Picas et al. 2019;Ndebele et al. 2020). From these registers, sub-daily temperature readings, used to calculate mean daily temperatures, span 1834-1899 (Picas et al. 2019); however, daily rainfall records span 1841-1899 (Ndebele et al. 2020); thus, we only use data for the overlapping period 1841-1899. Data for the SAAO for 1933-2020 were obtained from the South African Weather Service (SAWS). Although daily rainfall records span 1900-2020, daily maximum (T max ) and minimum (T min ) temperature records span 1933-2020; thus, our analyses only consider data for 1933-2020. Overall, our analyses consider the periods of 1841-1899 and 1933-2020 as the prior represents the longest period of hard copy digitized archival records containing daily rainfall and temperature variables, while the latter represents the longest period of records containing these same variables stored by SAWS.
Before performing any statistical analyses, data quality was checked and cleaning was performed where necessary. Details of these processes undertaken for data for 1841-1899 are documented elsewhere (see Picas et al. 2019;Ndebele et al. 2020;Picas and Grab 2020). Given the absence of local rainfall and temperature reference datasets, prior to 1900, data gaps for 1878-1880 remain. Besides this, the dataset is complete. Before discussing the quality control and cleaning process for data for 1933-2020, note that we followed the processes undertaken for 1841-1899 as closely as possible to ensure consistency, and raw data for 1933-2020 were near-complete, with ~ 8% missing T max and T min values and ~ 5% missing rainfall values. Initially, dates were checked for temporal consistency and values were rounded to one decimal place for uniformity. Thereafter, raw data were checked to identify errors and outliers. Typographic errors, negative rainfall values, accumulated rainfall values, where T min ≥ T max were among the errors identified (Dyson 2009;Durre et al. 2010;Ndebele et al. 2020;Picas and Grab 2020;van der Walt and Fitchett 2020). Outliers were identified using boxplots (for temperature), histograms (for rainfall), and tolerance tests, involving checking for rainfall and temperature values outside upper/lower limits for the SAAO (Aguilar et al. 2003;Durre et al. 2010). Once validated, through comparison with surrounding stations records, true erroneous/outlier values were deleted and recorded as missing values. These were estimated or replaced on a daily scale using nearby station records within a 15-km radius and with a Pearson correlation coefficient value > 0.70 for existing data . Missing temperature data were replaced with data from the closest station, or, if not possible, estimated using a five-day weighted average (Picas and Grab 2020;van der Walt and Fitchett 2020). Missing daily rainfall readings were estimated by a weighted amount, considering the common period rainfall at the SAAO and the nearest station (Ndebele et al. 2020). Estimation was only undertaken for missing SAAO values corresponding to nonzero rainfall values for the nearest station, while all other missing SAAO values were set to zero (Ndebele et al. 2020). Based on a weighted ratio, accumulated rainfall values, labelled by SAWS or identified when large rainfall values followed immediately after missing values, were disaggregated across rain days identified for the closest station (Ndebele et al. 2020).
It is necessary, following data quality control and cleaning, to determine data homogeneity (Aguilar et al. 2003;Wolski et al. 2021). However, due to a lack of metadata, on changes in data collection methods or station relocation, for instance, and no suitable reference series prior to 1900, homogenization was not possible for 1841-1899 (Picas et al. 2019). Thus, homogeneity was not tested.

Quantifying rainfall seasonality
To quantify rainfall seasonality characteristics for the SAAO, a ratio of monthly rainfall/temperature (Roffe et al. 2021b) and a percentile-based wet-season start-and end-date metric (Roffe et al. 2020) were applied using a January-December climatological year. Although numerous metrics have been applied across South Africa (Roffe et al. 2019), these were applied because they accurately characterize and statistically discriminate South African summer-, winter-, and year-round rainfall regimes (Roffe et al. 2020(Roffe et al. , 2021b. Mean monthly rainfall and temperature, calculated using daily rainfall and temperature records, were used to quantify annual seasonality scores (termed score hereafter) following the ratio method (see Roffe et al. 2021b for method details). This score, representing the unitless m-coefficient of the least squares linear regression equation, models the relationship between rainfall and temperature, where a negative (positive) relationship corresponds to high rainfall during cooler, winter (warmer, summer) months (Roffe et al. 2021b). Resulting scores have positive/negative signs, indicative of seasonality timing, with varying magnitudes quantifying the degree of seasonality, which represents seasonality strength and measures the contrast in length and rainfall amount of wet and dry seasons (Roffe et al. 2021b). Based on this, scores are classified as summer-, winter-, or year-round rainfall regimes, where scores < -0.30 represent WRZ conditions (Roffe et al. 2021b).
Using daily rainfall records, wet-and dry-season characteristics were quantified using a percentile-based wet-season start-and end-date metric (see Roffe et al. 2020, 2021a for method details). Using this metric, summer-, winter-, and year-round rainfall regimes are distinguished primarily through the wet-season timing and duration, where WRZ conditions are classified when the wet season is shorter than nine months (~ 270 days) and includes the winter solstice, but not the summer solstice (Roffe et al. 2020). For the wet season, we quantified annual start and end dates, length, total rainfall, number of rain days and daily rainfall rate (Roffe et al. 2020(Roffe et al. , 2021a. As wet-season start and end dates broadly equate to dry-season end and start dates, respectively, and can be used to infer the dry-season length, only the dry-season total rainfall, number of rain days, and daily rainfall rate were calculated.

Analyses for rainfall seasonality shifts
Rainfall seasonality shifts were analysed through trend and comparative analyses. For all tests, statistical significance was determined using p values, with the alpha level set to 5%.
Trends were calculated for 1841-1899, 1933-2020, and 1841-2020 using the nonparametric Mann-Kendall (MK) trend test with the Sen's slope (ss) estimator, calculating the trend magnitude (Mahlalela et al. 2019;Hekmatzadeh et al. 2020). Considering parametric test assumptions, this test and all the nonparametric tests discussed below were applied as they are robust to outliers and have no data distribution assumptions (Hekmatzadeh et al. 2020).
Accounting for the data gap, the nonparametric sequential Mann-Kendall test (SQMKT) was applied only to records spanning 1841-1899 and 1933-2020 to explore the evolution of trends, to determine the start of statistically significant trends if present, and to determine whether change points exist, following George and Athira (2020) and Hekmatzadeh et al. (2020). Using this method, two indicators, termed U(t) and U′(t), are defined based on ranks of the prograde, forward time series and retrograde, backward time series, respectively (Sneyers 1991). Abrupt change points, representing shifts in data distribution properties including the mean, median and variance (Beaulieu et al. 2012), are identified by plotting U(t) and U′(t), where the intersection of the two curves indicates a potential abrupt change point location (Sneyers 1991). Changes in the trend direction are also evident from U(t) line direction changes (Hekmatzadeh et al. 2020). Abrupt change points are considered statistically significant when Z > 1.96 (or Z < -1.96) and the same applies for trends; however, trends are only statistically significant when this occurs for U(t) values (Sneyers 1991;Hekmatzadeh et al. 2020). Where statistically significant change points exist, trends were additionally calculated, using the abovementioned trend test, for periods before and after the change point (Ndebele et al. 2020).
To determine whether any distinct changes occurred for the rainfall seasonality characteristics, a comparative analysis was performed for the first and last 59 years of the study period, spanning 1841-1899 (historical period: HP) and 1962-2020 (recent period: RP). This was undertaken using the nonparametric Wilcoxon paired samples test (WPST), a test determining whether differences in the distribution and median exist between two data samples, boxplots and additional descriptive statistics (e.g. standard deviation).
The very weak insignificant score trend of -0.0020.yr −1 (z = -0.87, p = 0.385) for 1841-1899 indicates a tendency towards a stronger degree of seasonality as scores tend away from zero, towards more unevenly distributed rainfall (Fig. 2a). The SQMKT results reflect strong variability, evident from numerous insignificant change points and the highly variable nature of U(t), thus indicating no clear score trend (Fig. 3a). A reduction (an increase) in the wet-season (dry-season) length similarly reflects a tendency towards a stronger degree of seasonality. This trend of -0.1d.yr −1 (z = -0.34, p = 0.734) is similarly weak and insignificant, and numerous insignificant change points with a highly variable nature of U(t) also indicate strong variability and no clear trend (Figs. 2d,3d). This is the case for the very weak insignificant wet-season start-(0.1d.yr −1 , z = 0.35, p = 0.724) and end-date (0.1d.yr −1 , z = 0.81, p = 0.420) trends, reflecting later start and end dates (Figs. 2b,c,3b,c). However, the wet-season end dates were characterized by a significant change point for 1879 (Fig. 3c). Despite being weaker than the trend prior to 1878 (given 1878-1880 missing data), the trend towards earlier end dates (-0.6d.yr −1 , z = -0.56, p = 0.576) from 1881 may have, together with the overall later start-date trend, contributed to driving the shorter wetseason trend until 1899 as this trend is more pronounced from ~ 1881 (Fig. 3c, d, Table 1).
Consistent with the score trends detected for 1841-1899 and 1933-2020, the trend for 1841-2020 reflects a decreasing trend of -0.0006.yr −1 (z = -1.25, p = 0.212; Fig. 5a), indicating that since 1841 the scores have generally tended towards stronger WRZ conditions. The near-significant decreasing trend (-0.1d.yr −1 ; z = -1.72, p = 0.085) for the wet-season length for 1841-2020 is similarly consistent with the decreasing trends for 1841-1899 and 1933-2020 (Fig. 5d). Together with an increasing dry-season length, this also reflects a trend towards stronger WRZ conditions. Notably, the SQMKT results for the scores and wet-season length appear to broadly track each other in direction throughout 1841-2020 (Fig. 3a, d), thus highlighting good agreement between the two methods (Roffe et al. 2021a). The U(t) line reflects cyclic patterns, with periods of stronger and weaker seasonality (Fig. 3a, d); though further analysis in terms of length and drivers of cycles is beyond the scope of this paper. The near-significant later wet-season startdate trend of 0.1d.yr −1 (z = 1.82, p = 0.068) for 1841-2020 also corresponds to the trends quantified for 1841-1899 and 1933-2020 (Fig. 5b). No trend exists in the wet-season end dates for 1841-2020 (Fig. 5c), which corresponds to large variability in the end dates for 1841-1899 and 1933-2020 (Fig. 3c). This highlights that since 1841, and particularly from 1933, the later start-date trend has primarily driven the shorter wet-season trend. Given this, it is notable that the cyclic patterns reflected by the start-date U(t) line broadly support later (earlier) start dates for periods with shorter (longer) wet seasons, while relatively little correspondence exists between the end dates and wet-season length (Fig. 3b,  c, d).
Although the trend in wet-season totals for 1841-2020 indicates a near-significant decline of -0.4 mm.yr −1 (z = -1.67, p = 0.096; Fig. 5e), this trend has only persisted since 1881, or more likely since 1892 (Fig. 3e, Table 1). The significant decreasing trend of -0.1d.yr −1 (z = -2.52, p = 0.012) calculated for the wet-season rain day counts for 1841-2020 corresponds to the decreasing trends quantified for 1841-1899 and 1933-2020 (Fig. 5f). No trend was detected for the wet-season daily rainfall rate for 1841-2020 (Fig. 5g), despite significant increasing trends quantified for 1841-1899 and 1933-2020. This is primarily because the series was characterized by relatively high variability, as is evident from the U(t) line and numerous significant and nonsignificant change points (Fig. 3g, Table 1).
While the trend for 1841/42-2019/20 indicates a nearsignificant decline of -0.1 mm.yr −1 (z = -1.91, p = 0.056) for the dry-season totals (Fig. 5h), this trend has only persisted at a significant rate of -0.3 mm.yr −1 (z = -2.03, p = 0.042) since 1955/56 (Fig. 3h, Table 1). No trend was calculated for 1841/42-2019/20 for the dry-season rain day counts and daily rainfall rate (Fig. 5i, j). Despite a significant decreasing trend of -0.1d.yr −1 (z = -2.34, p = 0.019) from 1949/50 for the dry-season rain day counts, the no trend result broadly corresponds to relatively high variability evident from the U(t) line for 1841/42-1898/99 and the significant change point detected (Fig. 3i, Table 1). For the dry-season daily Fig. 2 Time series and corresponding linear trend lines of (a) seasonality scores, (b) wet-season start dates, (c) wet-season end dates, (d) wet-season lengths, (e) wet-season total rainfall, (f) number of wetseason rain days, (g) wet-season daily rainfall rate, (h) dry-season total rainfall, (i) dry-season number of rain days, and (j) dry-season daily rainfall rate for 1841-1899 (or 1841/42-1898/99 for the dryseason variables). Here, ss represents the Sen's slope, z represents the Mann-Kendall statistic, and p represents the Mann-Kendall test p value. Statistically significant p values (p < 0.05) are denoted in bold ◂ rainfall rate, this corresponds to high variability throughout 1841-2020, evident from the U(t) line and numerous significant and non-significant change points (Fig. 3j).
Although we do not further explore cyclic patterns for the various rainfall seasonality variables, it is worth noting similarity in trend evolution. This exists between the wetand dry-season totals (Fig. 3e, h), the wet-and dry-season rain day counts (Fig. 3f, i), and the wet-and dry-season daily rainfall rate (Fig. 3g, j); though, there appears to be stronger, more distinct cyclicity for the wet-and dry-season totals. The structure, and specifically the timing, of cyclicity appears to change from the earlier (1841-1899) to later period (1933-2020), a result which is similarly evident for SAAO annual and seasonal rainfall totals (Ndebele et al. 2020). As such, the similarity in trend structure across the wet and dry seasons highlights that despite differences in their timing, there is much similarity in their rainfall characteristics and trends.

Statistically comparing historical (1841-1899) and recent (1962-2020) period rainfall seasonality characteristics
Significant median and data distribution differences in the various rainfall seasonality variables for 1841-1899 (HP) and 1962-2020 (RP) were calculated, using the WPST, for the wet-season rainfall totals, its rain day counts, and the dry-season rainfall totals (Fig. 6e, f, h). For the remaining variables, the differences are insignificant (Fig. 6a, b, c, d, g, i, j). Despite this, data distribution differences exist between the HP and RP for most variables-evident from visual inspection of the boxplots, focusing mainly on the interquartile range (IQR), representing the behaviour and spread of most data values, and mean and median values (Fig. 6). Notwithstanding the statistical significance level, the differences detected generally agree with trend results for 1841-2020 (Figs. 5, 6, Table 2), reflecting that the mean, median, and data distribution differences broadly represent the results of these trends. The mean (HP = -1.21, RP = -1.23) and median (HP = -1.16, RP = -1.19) score values reflect slightly stronger RP scores (Fig. 6a, Table 2), indicating a shift towards a slightly stronger degree of seasonality for the RP. This corresponds to a shorter (longer) RP wet-season (dry-season) length, supported by the plotted IQR and smaller RP minimum (HP = 139 days, RP = 115 days), mean (HP = 194 days, RP = 183 days), median (HP = 192 days, RP = 184 days), and maximum (HP = 294 days, RP = 240 days) values (Fig. 6d, Table 2). Although notable differences exist in the distribution of the HP and RP wet-season end-date datasets, particularly reflected by larger interannual variability (HP coefficient of variability [CV] = 7.5%, RP CV = 9.9%) and larger IQR (HP = 25 days, RP = 39 days) values for the RP, the mean (HP = 12 October, Julian day [JD] = 285, RP = 10 October, JD = 283) and median (HP = 10 October, JD = 283, RP = 11 October, JD = 284) values are very similar and reflect little to no end-date shift (Fig. 6c, Table 2). This confirms that the wet-season (and dry-season) duration shift is primarily due to later RP wet-season start-dates, evident from the plotted IQR, and later RP mean (HP = 3 April, JD = 93, RP = 11 April, JD = 101) and median (HP = 11 April, JD = 101, RP = 17 April, JD = 107) values (Fig. 6b, Table 2).
Changes also occurred in the magnitude and intensity of dry-season rainfall, and the rain day counts, with the HP dryseason wetter than the RP dry-season (Fig. 6h, i, j, Table 2). The plotted IQR and mean (HP = 113.2 mm, RP = 104.2 mm) and median (HP = 110.6 mm, RP = 103.0 mm) values reflect lower RP dry-season totals (Fig. 6h, Table 2). The RP dryseason rain day counts are marginally lower than those for the HP, reflected by the plotted IQR (Fig. 6i) and mean (HP = 29 days, RP = 27 days) and median (HP = 30 days, RP = 27 days) values ( Table 2). The HP dry-season daily rainfall rate was only slightly higher than that for the RP, as demonstrated by the plotted IQR and mean (HP = 4.2 mm. d −1 , RP = 4.1 mm.d −1 ) and median (HP = 3.8 mm.d −1 , RP = 3.7 mm.d −1 ) values (Fig. 6j, Table 2). Fig. 3 Sequential Mann-Kendall test of (a) seasonality scores, (b) wet-season start dates, (c) wet-season end dates, (d) wet-season lengths, (e) wet-season total rainfall, (f) number of wet-season rain days, (g) wet-season daily rainfall rate, (h) dry-season total rainfall, (i) dry-season number of rain days, and (j) dry-season daily rainfall rate for 1841-2020 (or 1841/42-2019/20 for the dry-season variables). Solid black lines represent the U(t) values, dashed black lines represent the U'(t) results, and horizontal dashed grey lines depict the z value confidence intervals at ± 1.96 ◂

Discussion
Using the SAAO meteorological record, we present the longest station-based rainfall seasonality record for Cape Town, within South Africa's WRZ. This long-term record is particularly valuable for trend analysis as trend evolution can be explored meaningfully and in detail as has been done here using the SQMKT (Ndebele et al. 2020), and the influence of interdecadal variability on trend magnitude and direction is limited compared to that for shorter records (e.g. Roffe et al. 2021a). While this is the case, it is important to acknowledge that the quasi-periodicities evident from the SQMKT plots can make detection and interpretation of long-term trends difficult, thus necessitating consideration of trend evolution for long-term records ). Although we do not further explore these periodicities, as this is beyond our paper scope, these periodicities, especially for the wet-and dry-season totals, would likely be similar to those detected by Dieppois et al. (2016), Ndebele et al. (2020) and Wolski et al. (2021). Nevertheless, our results reflect that rainfall seasonality, expressed here by variables measuring the degree of seasonality and the duration, intensity, magnitude, and timing of wet-and dry-season rainfall, has, for 1841-2020, undergone a significant change for some of the variables considered. Note that the rainfall seasonality trends quantified here are not representative of the full WRZ given that rainfall seasonality characteristics demonstrate spatial heterogeneity (Roffe et al. 2020(Roffe et al. , 2021b. Rather these trends tentatively extend across the far southwestern Cape region, containing Cape Town's supply dams (Fig. 1). This is argued based on results from Wolski et al. (2021), as trends for stations across the southwestern WRZ region they defined, based on interannual rainfall variability patterns, are largely consistent, particularly in direction, throughout the various temporal periods they considered. As rainfall seasonality characteristics represent surface responses of weather systems (Lennard and Hegerl 2015), the changes quantified here are likely largely driven by weather system changes which are primarily driven by hemispheric-scale atmospheric circulation changes (Sousa et al. 2018;Burls et al. 2019). As our aim is only to explore how Cape Town's rainfall seasonality characteristics have changed during 1841-2020, we only tentatively discuss trend drivers.
Results from the ratio and percentile-based seasonality metric demonstrate strong agreement, reflecting a trend towards stronger WRZ conditions. This trend and agreement, evident from the score and wet-season (and dryseason) length trend directions and the SQMKT results thereof, exists throughout 1841-2020, though trends are stronger for 1933-2020. The changes in seasonality timing are thus a degree of seasonality increase, with scores tending away from zero, and a wet-season (dry-season) shortening (lengthening), a result consistent with observed and projected trends for the southwestern Cape (Li et al. 2016;Pascale et al. 2016;du Plessis and Schloms 2017;Dunning et al. 2018;Jury 2020;Ndebele et al. 2020;Wolski et al. 2021). Although rainfall declines detected across the southwestern Cape for spring (September-November) or months therein are argued to reflect earlier end-dates (du Plessis and Schloms 2017; Jury 2020; Ndebele et al. 2020), our results do not necessarily reflect end-date changes during 1841-2020. Instead, the end-dates varied substantially, and this variability has seemingly increased for recent decades. The trend in wet-season duration is thus primarily driven by later wet-season (dry-season) start dates (end dates), which is near-significant for 1841-2020 and significant for 1933-2020. This is supported by our comparative analysis and by later wet-season start-date and autumn (March-May) drying trends quantified for recent decades, from ~ 1980, for Cape Town and the southwestern Cape, respectively (Mahlalela et al. 2019;Roffe et al. 2021a;Wolski et al. 2021). Considering drivers behind the wet-season (and dry-season) timing changes, it is likely that the later start-date trend is linked to increases in the frequency of post-frontal SAA ridging high-pressure systems recorded for 1979-2017, which is likely driven by Hadley cell expansion and descending branch intensification, especially in Autumn (Grise et al. 2018;Burls et al. 2019;Mahlobo et al. 2019). This is supported by Mahlalela et al. (2019) who link dry early winter (April-May) periods, which coincide with the timing of most SAAO start dates, to an increased frequency of SAA ridging high-pressure systems, across the southwestern Cape. Further, Mahlalela et al. (2019) note that the driest winters during 1979-2017 occurred after 2000, which corresponds to a significant later start-date trend, from ~ 2007 according to our SQMKT results.
Our results consistently demonstrate that the wet season has become drier. Wet-season totals have declined since ~ 1881, or ~ 1892 at least, and our comparative analysis results demonstrate that this reduction is one of the most distinct rainfall seasonality changes to have occurred during 1841-2020. The winter wet-season drying trend, being the strongest direct driver of the 2015-2017 "Day Zero" drought, is among the most robustly projected rainfall Fig. 4 Time series and corresponding linear trend lines of (a) seasonality scores, (b) wet-season start dates, (c) wet-season end dates, (d) wet-season lengths, (e) wet-season total rainfall, (f) number of wetseason rain days, (g) wet-season daily rainfall rate, (h) dry-season total rainfall, (i) dry-season number of rain days, and (j) dry-season daily rainfall rate for 1933-2020 (or 1933/34-2019/20 for the dryseason variables). Here, ss represents the Sen's slope, z represents the Mann-Kendall statistic, and p represents the Mann-Kendall test p value. Statistically significant p values (p < 0.05) are denoted in bold ◂ Fig. 5 Time series and corresponding tentative linear trend lines of (a) seasonality scores, (b) wet-season start dates, (c) wet-season end dates, (d) wet-season lengths, (e) wet-season total rainfall, (f) number of wet-season rain days, (g) wet-season daily rainfall rate, (h) dry-season total rainfall, (i) dry-season number of rain days, and (j) dry-season daily rainfall rate for 1841-2020 (or 1841/42-2019/20 for the dry-season variables; dashed blue line). Trendlines for 1841-1899 (or 1841/42-1898/99; solid black line) and 1933-2020 (or 1933/34-2019/20; solid grey line) are also plotted. Here, ss represents the Sen's slope, z represents the Mann-Kendall statistic, and p represents the Mann-Kendall test p value. Statistically significant p values (p < 0.05) are denoted in bold changes in response to anthropogenically induced climate change (Dunning et al. 2018;Mahlalela et al. 2019;Pascale et al. 2020;Lim Kam Sian et al. 2021), and it is significant for recent decades for Cape Town (Roffe et al. 2021a). Numerous studies have attributed this drying trend to an observed poleward expansion of the subtropical high-pressure belt and migration of the westerlies (Sousa et al. 2018;Burls et al. 2019;Mahlalela et al. 2019), and it is very likely that these trends will persist until the end of the century, at least (Engelbrecht et al. 2009;Chavaillaz et al. 2013;Fahad et al. 2020). Results from the literature suggest these trends have manifested as a decline in the frequency of winter rainbearing circulation features (cold fronts in particular) with an increased frequency of dry circulation features such as SAA ridging high-pressure systems (Lennard and Hegerl 2015;Sousa et al. 2018;Burls et al. 2019;Mahlalela et al. 2019;Odoulami et al. 2021). Burls et al. (2019) additionally demonstrate that, due to increased post-frontal SAA ridging, the duration and intensity of cold front rainfall events have declined since ~ 1979. This is linked to a long-term decline, since ~ 1900, in rain day counts and a more recent decline, since ~ 2010, in rainfall intensity for the April-September winter wet-season (Burls et al. 2019). This is supported by our results of a decline in wet-season rain day counts throughout 1841-2020, but more prominently since 1933, and of a more recent daily rainfall rate decline since ~ 2009.
Evidently, the dry-season has also become drier. This trend is significant from 1933, or more specifically from ~ 1955/56, and our comparative analysis results robustly reflect reduced RP dry-season totals. Our results also reflect a significant decline (since ~ 1949/50 at least), in the dry-season rain day counts and during recent decades, from ~ 1996/97 or ~ 2001/02 at least, the daily rainfall rate has also declined. Thus, together with the wet-season rainfall decline, the dry-season rainfall decline has contributed to an increasing duration, frequency, and intensity of drought periods for Cape Town and the southwestern Cape region recorded since ~ 1985 (Botai et al. 2017;De Kock et al. 2021). Notably, dry-season rainfall has received limited research attention, despite its importance in mitigating drought impacts through a contribution to increasing dam levels, especially from large dry-season rainfall events (De Kock et al. 2021;Wolski et al. 2021). Despite this, much evidence exists for long-term drying during the dry-season, or months therein (Ndebele et al. 2020;De Kock et al. 2021;Wolski et al. 2021). Although limited evidence exists for changes in the dry-season rain day counts and daily rainfall rate, for Cape Town and the southwestern Cape, Mackellar et al. (2014) demonstrate declines in summer (December-February) rain day counts for 1960-2010. Considering drivers behind the dry-season rainfall changes, it is likely that these are consistent with the hemispheric-scale drivers of wet-season rainfall changes. Not only is this suggested by De Kock et al. (2021) and Wolski et al. (2021), but similar patterns of cyclicity across our SQMKT wet-and dry-season results reflect similar trend drivers. Thus, the dry-season trends are similarly linked to a stronger and farther south and expanded SAA, induced by Hadley cell expansion and descending branch intensification (Sousa et al. 2018;Mahlobo et al. 2019;Jury 2020;De Kock et al. 2021;Wolski et al. 2021). Declines in the rainfall contributions from atmospheric rivers and cut-off lows are detected for dryseason months from 1979/80 and are similarly consistent  Odoulami et al. 2021). Again, it should be highlighted that these hemispheric dynamics, identified as drivers behind dry-season rainfall changes, are expected to change in a similar manner in the future (Engelbrecht et al. 2009;Chavaillaz et al. 2013;Fahad et al. 2020).
Strong consistency between the trend, SQMKT and comparative analysis results provides corroboration for the nature of seasonality trends quantified here, while strong agreement with seasonality projections, especially for recent decades, suggests these trends may continue (Pascale et al. 2016;Dunning et al. 2018). The consistency and robustness of observed and projected trajectories of the Hadley cell and mid-latitude westerlies (Yin 2005;Lu et al. 2007;Chavaillaz et al. 2013;Nguyen et al. 2015;Sousa et al. 2018;Mahlobo Fig. 6 Boxplots comparing the statistical distribution of historical (1841-1899) and recent   , our results of drier conditions during the wet and dry seasons, which are supported by copious evidence from the literature, demonstrate that Cape Town will likely experience increasing water stress. As such, water resource management and planning need to focus on reducing the risks and vulnerability associated with changing rainfall regimes. Water conservation and management strategies, such as recycling, water supply diversification (e.g. desalination and groundwater extraction), ecosystem-based approaches (e.g. removal of invasive vegetation, ensuring more water reaches dams), and awareness initiatives, implemented not only in response to the 2015-2017 "Day Zero" water crisis, but in general, must continue and be improved upon (e.g. Booysen et al. 2019).
As for any study, there are limitations, and here, these arise from the use of long-term data. For the SAAO record, limitations hinder establishment of dataset homogeneity and include no existing reference series for 1841-1899, little metadata found for instrumental type, changes in instrumentation and location, and recording methods and times (Lakhraj-Govender et al. 2017;Picas et al. 2019). The rainfall record has, for the most part, been based on manually derived daily rain gauge readings from the SAAO gardens (Ndebele et al. 2020); however, during 2009 an automatic station was installed (Glass 2018). For the temperature record, several changes in recording times and methods occurred throughout 1841-2020 (Lakhraj-Govender et al. 2017;Picas and Grab 2020). Recent records, from 2009, are captured hourly by automatic stations, whereas earlier records were manually captured at varying times (Lakhraj-Govender et al. 2017;Picas and Grab 2020). With no station relocation records, it can be assumed that the SAAO monitoring station has remained in the same location (Lakhraj-Govender et al. 2017;Ndebele et al. 2020;Picas and Grab 2020). Thus, the main limitation, besides that of the data gap for 1900-1933, underlying our results is that our records were not homogenized to account for changes in instrumental type and recording times. This is a limitation to note for our comparison of the historic 1841-1899 records to those obtained from SAWS; however, the daily rainfall and temperature records for the SAAO have, for the most part, been collected in a relatively consistent manner, adding confidence in our results. Further, records for 1841-1899 were corrected for detectible errors and are comparable to recent records, which increases robustness of this record (Ndebele et al. 2020;Picas and Grab 2020), and results from Lakhraj-Govender et al. (2017) reflect little change in trend magnitude and no change in trend direction for the SAAO 1933-2013 temperature series following homogenization adjustments. Hence, this improves confidence in our results.

Conclusion
Based on the longest record of rainfall seasonality characteristics for Cape Town within South Africa's WRZ, we demonstrate value in using long-term records to explore rainfall regime trends in detail. This is particularly evident from our SQMKT application which demonstrates a complex temporal pattern for Cape Town's rainfall seasonality characteristics for 1841-2020 and reveals much value in exploring trend evolution, beyond simply considering overall trend magnitudes and directions for arbitrarily defined periods. Together with evidence from projected rainfall seasonality trends for Cape Town, which correspond to our findings, we provide an example of the likely rainfall seasonality patterns to occur across Cape Town in future decades. Our findings particularly highlight trends towards shorter, later starting and drier wet seasons together with longer and drier dry seasons and have important implications for rainfall sensitive activities. Thus, our findings can contribute to water resource management strategies for Cape Town and the broader southwestern Cape region, a region already vulnerable to rainfall regime changes. If scientific evidence is considered properly for such strategies, then the impacts of future extreme multiyear droughts, which are projected with much robustness (Otto et al. 2018;Pascale et al. 2020), may be mitigated considerably.