Theor Appl Climatol (2009) 98:119–135 DOI 10.1007/s00704-008-0093-6 ORIGINAL PAPER Temporal rainfall variability in the Lake Victoria Basin in East Africa during the twentieth century Michael Kizza & Allan Rodhe & Chong-Yu Xu & Henry K. Ntale & Sven Halldin Received: 17 April 2008 / Accepted: 6 December 2008 / Published online: 20 January 2009 # Springer-Verlag 2009 Abstract Water resources systems are designed and operated on assumption of stationary hydrology. Existence of trends and other changes in the data invalidates this assumption, and detection of the changes in hydrological time series should help us revise the approaches used in assessing, designing and operating our systems. In addition, trend and step change studies help us understand the impact of man’s activities (e.g. urbanisation, deforestation, dam construction, agricultural activities, etc.) on the hydrological cycle. Trends and step changes in the seasonal and annual total rainfall for 20 stations in the Lake Victoria basin were analysed. The seasonal rainfall for any station in a given year was defined in two ways: (1) fixed time period where the rainy seasons were taken as occurring from March–May (long rains) and from October–December (short rains); and (2) variable periods where the rainy seasons were taken as the three consecutive months with maximum total rainfall covering the period of January–June (long rains) and July–December (short rains), to take into account the fact that the onset of rainy seasons within the basin varies from year to year and from one station to the next. For each station, sub datasets were derived covering different periods (all available data at the station, 1941– 1980, 1961–1990, 1971–end of each station’s time series). The trends were analysed using the Mann-Kendall method, while the step changes were analysed using the Worsley Likelihood method. The results show that positive trends predominate, with most stations showing trend being located in the northern part of the basin, though this pattern is not conclusive. In all, 17% of the cases have trends, of which 67% are positive. The 1960s represent a significant upward jump in the basin rainfall. Seasonal rainfall analysis shows that the short rains tend to have more trends than the long rains. The impact of the varying month of onset of the rainy season is that the results from analyzing the fixedperiod and variable-period time series are rarely the same, meaning the two series have different characteristics. It may be argued that the variable-period time series are more reliable as a basis for analysing trends and step changes, since these time series reflect more closely the actual variability in rainy seasons from one year to the next. The fixed-period analysis would, on the other hand, find more practical use in planning. 1 Introduction M. Kizza (*) : H. K. Ntale Faculty of Technology, Makerere University, P.O. Box 7062, Kampala, Uganda e-mail: michael.kizza@hyd.uu.se A. Rodhe : S. Halldin Department of Earth Sciences, Uppsala University, Villavägen 16, 752 36 Uppsala, Sweden C.-Y. Xu Department of Geosciences, University of Oslo, P.O. Box 1047, Blindern, NO-0316 Oslo, Norway Lake Victoria basin in East Africa has an abundance of natural resources and provides services like fishing, transport, agriculture, domestic and industrial water supply, as well as hydropower (Ntiba et al. 2001). The lake basin is one of the most densely populated in Africa with more than 30 million people living around it and drawing their livelihoods directly or indirectly from its resources. The lake is also one of the main sources of the Nile River, which is a key lifeline for Sudan and Egypt who depend almost entirely on the river for water supply (Sutcliffe and 120 Parks 1999). In addition, the lake and its basin have a rich diversity of flora and fauna that are dependent on it for survival. For example the wetlands surrounding the lake serve as breeding grounds for fish and birds. However, 80% of the input into the lake’s water balance is rainfall over its surface, leading some researchers to describe it as ‘atmosphere controlled’ (Flohn and Burkhardt 1985; Yin and Nicholson 2002; Tate et al. 2004). This essentially means that the variability of rainfall over the lake plays a key role in the fluctuation of the lake levels. The levels have exhibited large and rapid changes in response to rainfall anomalies over the last century (Conway 2002; Mistry and Conway 2003). Changes in the lake basin rainfall regime have far-reaching ecological, environmental, hydrological and socio-economic effects. Reduction in lake water levels affects plant and animal habitats, impairs navigation, reduces hydropower generation, fish catch and also reduces access to clean water. Several studies have been aimed at assessing the spatial and temporal variability in the region (Rodhe and Virji 1976; Ogallo 1979; Ogallo 1989; Nyenzi 1990). Rodhe and Virji (1976) examined the existence of trends and periodicities in East African annual rainfall data. Computations by Rodhe and Virji revealed that apart from stations in northern Kenya, most stations did not show any appreciable trends. Spectral analysis also showed a 5– 6 years spectral peak for rainfall in the Lake Victoria region. Ogallo (1979) also noted that most of the annual rainfall series in the region indicated an oscillatory characteristic with no significant trend. Ogallo (1989) used monthly records from over 90 stations in East Africa to study the dominant spatial and temporal modes of seasonal variation using rotated principal component analysis for the period 1922–1983. He demonstrated the dominant effect of nearby large water bodies like Lake Victoria and the Indian Ocean on the seasonal rainfall patterns in the region. In particular, he showed that the Lake Victoria region has a distinctive rainfall regime in East Africa as a whole. Between 2002 and 2006, water levels in Lake Victoria dropped to pre-1961 values despite having remained at higher values for over 40 years (LVBC 2006). This has had profound environmental and socio-economic impact on all activities that depend on the lake resources mentioned above. The reasons for this drastic drop are not yet fully understood but a reduction of the total rainfall input into the lake and its basin has been identified as one of the possibilities. However, for effective planning and management of the regional water resources, there is need to constantly update the knowledge of temporal variability of rainfall in the lake basin. A number of years have passed since the temporal studies were carried out and more data have become available. There is clearly a need M. Kizza et al. to carry out temporal analysis using updated datasets in order to analyse the current trends in precipitation in the region. The objective of this paper was to investigate the temporal distribution of rainfall in the Lake Victoria basin on seasonal to annual time scales. The aim was to use an updated dataset with records covering the period 1903– 2006 to test the presence of significant trends in the rainfall data. The approach was to test for trends in the seasonal and annual rainfall data for selected stations in the lake basin. The study attempted to address issues of dependence of trend test results on the period of study by dividing the primary rainfall series for each station into different subseries, carrying out trend tests on them and then making comparisons between the results. In this study, we also carried out temporal analysis of total rainfall for the two periods that correspond to the rainy seasons in East Africa; the ‘long rains’ and the ‘short rains’. The seasons were defined in two ways. One was the fixed-time period where the long rain season occurs in March to May and the short rain season was assumed to occur from October to December. The second was the maximum 3-month total rainfall with the first covering the period January to June (for the long rains) and the second covering the period July to December (for the short rains). 2 Study region and data 2.1 Study area Lake Victoria is the largest lake in Africa and the second largest lake in the world. The lake is located between latitudes 0o20’N–3oS and longitudes 31o40’E–34o53’E (Fig. 1). The lake basin area is 194,000 km2 and the lake surface area is about 68,800 km2 or 35% of the basin. The lake surface is shared between Kenya (6%), Uganda (43%) and Tanzania (51%) while its basin includes parts of Burundi and Rwanda. It is located in a continental sag between the two arms of the Great Rift Valley system, with high mountains ranges on the east and west (Kilimanjaro, Kenya and Rwenzori). The altitude of the lake surface is about 1,135 m amsl while the basin is made of a series of stepped plateaus with an average elevation of 2,700 m but rising to 4,000 m or more in the highland areas. The general climate of the lake basin ranges from a modified equatorial type with substantial rainfall occurring throughout the year, especially over the lake and its vicinity to a semiarid type characterised by intermittent droughts over some areas located even within short distances from the lake shore. Climate variability at different time scales in the lake basin is influenced by both large-scale and mesoscale circulations. Temporal rainfall variability in the Lake Victoria Basin 121 Fig. 1 Lake Victoria basin and its location in Africa (inset) and the rainfall stations used in the current study 2.2 General circulation and rainfall variability The diurnal, seasonal and inter-annual variability of Lake Victoria (and East Africa generally) climate results from a complex interaction between the inter-tropical convergence zone (ITCZ), El Nino/Southern Oscillation (ENSO), Quasibiennial Oscillation (QBO), large-scale monsoonal winds, meso-scale circulations and extra-tropical weather systems (Ogallo 1988; Mutai et al. 1998; Nicholson and Yin 2002). The wind and pressure patterns that govern the region’s climate include three principal air streams and three convergence zones namely; the Congo airstream with a westerly and southwesterly flow, the southeast monsoon and the northeast monsoon (Trewartha 1981; Nicholson 1996). The monsoons are thermally stable, and associated with subsiding air and are, therefore, relatively dry which partly accounts for the relatively arid conditions in much of the area. The Congo air mass is humid, thermally unstable and, therefore associated with rainfall. The Congo air mass significantly boosts convection and overall rainfall amounts received, especially over the western and northwestern parts of the Lake (Nicholson 1996). The three airstreams are separated by two convergence zones; the ITCZ which separates the monsoons and the Congo air boundary which separates the Indian Ocean easterlies and Atlantic Ocean westerlies (Trewartha 1981). A third convergence zone aloft separates the dry, stable northerly flow from Sahara and the moister southerly flow. The seasonal climate patterns follow the seasonal N–S movement of the ITCZ which lags the seasonal migration of the sun and results in a bimodal rainfall distribution; the March–May rainfall period (long rains) and the October– December rainfall period (short rains). The northeast (NE) and southeast (SE) monsoon winds also modify the seasonal climate of East Africa (Mukabana and Piekle 1996). The NE monsoon air stream occurs during the Southern Hemisphere summer and, after traversing over Egypt and Sudan, is warm and dry. On the other hand, the SE monsoon air stream occurs when the sun is north of the equator. It is cool and moist after picking up maritime moisture from the Indian Ocean and is responsible for large-scale precipitation over the lake basin. The QBO is a quasi-periodic oscillation of the equatorial zonal wind between easterlies and westerlies in the tropical stratosphere with a mean period of 28–29 months (Indeje et al. 2000). Inter-annual variability corresponds to the ENSO variability. El Niño years are usually associated with above normal rainfall amounts in the short rainfall season in most of the region (Indeje et al. 2000). However, arguments remain with regard to the relative importance of Indian Ocean versus Pacific Ocean forcing of East African rainfall (Mistry and Conway 2003; Latif et al. 1999). Meso-scale circulations due to orography, lake surface temperature and other factors have also been shown to influence rainfall variability in the Lake Victoria basin (Mukabana and Piekle 1996; Nicholson and Yin 2002; Anyah et al. 2006) 122 M. Kizza et al. 2.3 Rainfall data availability Rainfall in the lake basin has been recorded since the start of the twentieth century using manual rain gauges and, more recently, some automatic recording gauges. The rainfall data for the current study were collected from various sources including the hydro-meteorological database of the World Meteorological Organisation (WMO 1982), meteorological departments in Kenya, Tanzania and Uganda as well as from our correspondence with various researchers in the region. The data format was either as the raw daily values or aggregated monthly values. An assessment of the number of stations in the basin with time (Fig. 2) shows that from just a few stations at the turn of the twentieth century, the number grew to over 400 at the peak in the 1970s. Most of the stations are concentrated in Kenya, Tanzania and Uganda. The southwest of the basin has very few stations, which are mainly located in Rwanda and Burundi where political problems have resulted in few current data being available. Similarly, the records from the northwest of the basin, which is mainly part of Uganda, were interrupted for long periods in the late 1970s and 1980s. There has been a general decline in the rain gauge network coverage since the 1970s. The drop in network coverage is an familiar pattern especially in developing countries where insufficient funding, inadequate institutional frameworks, a lack of appreciation of the worth of longterm data and sometimes political turmoil over the recent years have resulted in a marked decline of national hydrometeorological gauging network coverage (Sene and Farquharson 1998; Sawunyama and Hughes 2008). 2.4 Dataset and data properties For this study, monthly rainfall records have been compiled for 20 stations. The main factor in selecting stations for inclusion in the temporal analysis was the length of records, 450 Number of stations with data 400 350 300 250 200 150 100 50 0 1900 1910 1920 1930 1940 1950 1960 Years 1970 1980 1990 2000 Fig. 2 Variation of number of rainfall stations in the Lake Victoria basin with time which was set to 50 years or more whenever possible. This number of stations was considered representative of the large basin area because of the strongly coherent patterns of variability throughout the region (Yin and Nicholson 1998). Almost all stations had some periods of missing data ranging from a few days to several years, whose gaps were filled using linear regression with nearby stations that have highly correlated rainfall records. However, many of the stations still had other constraints that made them unsuitable for the analysis. The first constraint was the length of the number of missing records, which was set to 5 months or less in order to minimise the uncertainty related to estimating the missing values. In cases where the station has several nearby stations with available records for use in estimation of missing values at a given station, this constraint was relaxed on the assumption that using many stations reduces the uncertainty in the estimated value. Other constraints included availability of recent records for assessment of the rainfall trends in recent years (since 2000) and how they fit into the overall pattern as well as ensuring a sufficient spread of stations around the basin. The locations of the stations that were used in the current study are shown in Fig. 1, and Fig. 3 is a chronogram detailing the data availability for each station with time. The summary of the key statistics of the dataset used in the study is shown in Table 1. It is seen that the mean annual rainfall varies between 2,037 mm for Bukoba and 847 mm for Musoma. In general, stations on the north to north eastern part of the basin receive more rainfall than those in the southern part. For the yearly rainfall, the standard deviation varies between 339 mm and 168 mm (for Bungoma and Mbarara respectively) while the coefficient of variation (CV) varies within a range of 0.24 and 0.13 for Ngudu and Bukoba respectively. The average CV is 0.19, indicating that the rainfall varies considerably from one year to the next. On average the dataset contains 65 years of records. Jinja has the longest records with 96 years while Rulenge has the shortest with 28 years. The maximum annual rainfall amount of 2,736 mm was observed in Bukoba while the minimum of 400 mm was observed in Ngudu. The maximum annual rainfall for five stations (Jinja, Bukoba, Biharamulo, Mwanza and Ngudu) occurred in 1961 while that of Buvuma occurred in 1963. The minimum annual rainfall for five stations has occurred in years after 2000. The range in the annual varies between 1,561 and 872 mm. Figure 4 shows a boxplot showing the variation of the annual data at each of the stations. A visual inspection of the annual time series from 12 selected stations from the study (Fig. 5) reinforces the commonly held view that the amount of rainfall received in the 1960s was above average. Other periods with above average rainfall conditions include the late 1970s to early 1980s as well as the late 1990s. Temporal rainfall variability in the Lake Victoria Basin Station Fig. 3 Chronograms showing years with complete monthly data, i.e. for which 12 monthly values are available 123 Entebbe (20) Kamenyamigo (19) Mbarara (18) Kabale (17) Bukoba (16) Rulenge (15) Biharamulo (14) Kahunda (13) Mwanza (12) Ngudu (11) Mugumu (10) Musoma (9) Shirati (8) Sotik (7) Kericho (6) Eldoret (5) Kitale (4) Bungoma (3) Buvuma (2) Jinja (1) 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 Year 2.5 Temporal distribution of the rainfall records A key feature of the dataset is that the number of years with data available for analysis varies with time period. Jinja has the time series with earliest available rainfall records starting in 1903 while the Kahunda records are only available since 1971. Half of the stations used have records from before 1930; 13 stations have more than 50 years of record and only one station has less than 40 years of record. A total of 17 of the 20 stations have data for the period after 2000 making it possible to assess the recent rainfall trends in the basin. Kericho (ending in 1986), Kitale (ending in 1988) and Biharamulo (ending in 1995) are the only stations with records ending before the year 2000. 3 Methodology Change in a data series can occur in various ways, e.g. steadily (a trend), abruptly (a step change) or in a more complex form. The change may affect the mean, median, variance or any other aspect of the time series. There are Table 1 Properties of the annual rainfall data Station name (number in Fig. 1) WMO number Annual mean (mm) CV Skewness Maximum (mm) Minimum (mm) Start year No. of years Jinja(1) Buvuma(2) Bungoma (3) Kitale (4) Eldoret(5) Kericho(6) Sotik(7) Shirati(8) Musoma(9) Mugumu(10) Ngudu(11) Mwanza(12) Kahunda(13 Biharamulo(14 Rulenge(15 Bukoba(16) Kabale(17) Mbarara(18 Kamenyamigo(19 Entebbe(20 8933043 8933005 8934134 8934008 8935133 9035003 9035013 9133002 9133000 9134033 9233005 9232009 9232027 9231000 923001 9131002 9129000 9030003 9031026 8932066 1,170 1,584 1,515 1,305 1,073 1,826 1,365 901 847 1,100 868 1,083 1,175 986 994 2,037 1,012 923 998 1,617 0.17 0.16 0.22 0.17 0.21 0.16 0.16 0.23 0.22 0.24 0.24 0.20 0.20 0.17 0.15 0.13 0.14 0.19 0.23 0.19 0.22 −0.29 −0.35 −0.18 0.60 0.44 0.15 0.56 0.13 −0.71 0.43 0.04 0.40 0.61 0.74 0.83 0.06 0.40 0.00 0.90 1,731 2,131 2,130 1,740 1,716 2,485 2,021 1,424 1,390 1,545 1,444 1,543 1,777 1,599 1,421 2,736 1,282 1,520 1,450 2,679 726 913 725 861 590 1251 751 536 421 324 400 671 707 624 769 1523 727 529 556 1117 1903 1930 1963 1929 1957 1927 1926 1944 1922 1966 1930 1950 1971 1922 1971 1922 1944 1903 1952 1944 96 68 41 59 47 59 78 57 84 33 70 56 29 69 28 82 57 89 41 58 124 M. Kizza et al. 2500 2000 Rainfall (mm) Fig. 4 Box-plot for the annual data for the study rainfall stations. Each box plot shows the median, lower and upper quartiles in the main box indicating the main variation in the data. The whiskers show the full range of the data while the ‘crosses’ represent data that might be considered as outliers 1500 1000 many approaches that can be used to detect trends and other forms of non-stationarity in time series data. The methods are broadly classified as parametric and non-parametric procedures. Non-parametric approaches find wide use in hydrological studies because there is no requirement of making assumptions on the distribution form. An important issue to deal with, when we attempt to test for existence of trends in a series is the inherent variability of hydrometeorological data (WMO 2000; Burn and Hag Elnur 2002). If the data series is sufficiently long so that natural cycles cancel out each other, then variability is not an important issue. However, the length of the records is usually not sufficiently long to support this assumption. We therefore have to develop a rigorous procedure for detection of trends. A systematic procedure was adopted that involves three related stages. – – The first stage was to select the stations to be studied (see section 2.3). The second step was to test for presence of trends in the rainfall data. Two trend methods (Mann-Kendall and linear regression tests) and one step-jump (Worsely likelihood ratio test) method were used. On application of the procedure, it was discovered that, for the current dataset, the Mann-Kendall and linear regression tests have similar power and give very similar results. Therefore, for the test for trend, only the results from the Mann-Kendal test are presented in this paper. The null hypothesis for the Mann-Kendall test is that there is no trend in the data while that for Worsely likelihood ir M ati us om M a ug um u N gu du M w an Ka za h Bi und ha a ra m ul o R ul en ge Bu ko ba Ka ba l Ka Mb e m ara en r ya a m ig En o te bb e So tik Sh Ji n Bu ja vu Bu ma ng om a Ki ta le El do r Ke et ric ho 500 – ratio test is that there is no change in mean of the data series for two different periods. The two methods are described in detail in sections 3.1 and 3.2. The third step was to determine the significance of the detected trends. This was achieved by carrying out resampling analysis using bootstrapping which helps to avoid the need for strict adherence of the data to test assumptions. The bootstrap resampling technique is described in section 3.3. 3.1 Mann-Kendall test This is a rank based method which is non-parametric and is based on an alternative measure of correlation called Kendall’s τ. Mann (1945) originally used this test and Kendall (1975) subsequently derived the test statistic distribution. It is robust to the effect of extremes (for example highly skewed data) and to deviation from a linear relationship. It has been used by other researchers in similar applications (Hamed and Rao 1998; Burn and Hag Elnur 2002; Helsel and Hirsch 2002; Xu et al. 2003). Helsel and Hirsch (2002) give a procedure for carrying out the Mann-Kendall test which involves computation of the standardised test statistic S given by S¼ n1 X n X sgn Xj Xi ð1Þ i¼1 j¼iþ1 where Xi and Xj are sequential data values, n is the dataset record length and sgn(Xj–Xi) is +1, 0 and −1 for Xj–Xi Temporal rainfall variability in the Lake Victoria Basin 125 Fig. 5 Annual rainfall series of selected study stations greater than, equal to or less than 0 respectively. The significance level, which indicates the strength of the trend, was determined by resampling analysis while the Kendall’s correlation coefficient (Hirsch et al. 1982), a measure of the strength of the correlation, was calculated as purpose of the test is to determine the mean of a time series after m observations (Worsley 1979). i ¼ 1; 2; . . . ; m ð3Þ t ¼ 2S=ðnðn 1ÞÞ E ðxi Þ ¼ m þ Δ i ¼ m þ 1; m þ 2; . . . ; n ð4Þ ð2Þ A positive value of τ indicates increasing trend and vice versa. 3.2 Worsely likelihood ratio test This method tests whether the means in two parts of a record are different. It also estimates the most likely time of change (in case the null hypothesis is rejected). The test assumes that the data are normally distributed and the E ðxi Þ ¼ m where μ is the mean prior to the change and Δ is the change in mean. The cumulative deviations from the mean Sk are calculated as: S0 ¼ 0 Sk ¼ k X i¼1 ðxi mÞ k ¼ 1; 2; . . . ; n ð5Þ 126 M. Kizza et al. The values of Sk are then weighted according to their position in the time series. h i k ðn k Þ0:5 Sk ð6Þ Zk ¼ s where s is the sample standard deviation (assumed to be equal for the two groups). The test statistic W is: n2 1=2 V ¼ maxZk ð7Þ W ¼ V 1v 2 The critical values for different significance levels for the test have been derived by Worsley (1979). A negative value of W indicates that the latter part of the record has a higher mean than the earlier part and vice versa. 3.3 Estimation of the significance levels of the test statistics In order minimize the effect of the test assumptions (like form and constancy of the distribution, independence) on the results, a bootstrap sampling strategy was adopted to compute the significance levels for the two test methods. In this case, the original data series is sampled with replacement to give a new series that has the same number of values as the original series but may contain more than one of some values in the original series but none of the other values (Davidson and Hinkley 1997; WMO 2000). The rationale behind this approach is that if there is no trend (using the Mann-Kendall test) or step jump (using the Worsley Likelihood ratio test) under the null hypothesis of no trend in the data, shuffling the data should not change the gradient very much. The data are shuffled many times and after each shuffle, the test statistic of the generated series is recalculated. The test statistic of the original series is then compared with that of the generated data to determine the significance level. Assuming that the test statistic of each of the generated series is estimated as Tk (which can be ordered as T1 ≤ T2 ≤…≤ TS), and assuming that the original test statistic is T0 and Tk ≤ T0 ≤ Tk+1, then the probability of the test statistic being less or equal to T0 under the null hypothesis is approximated as p¼ k N ð8Þ where N is the number of times a series is resampled. If we assume that large values of T indicate departure from the null hypothesis, the significance level is estimated from 100 2 minðp; 1 pÞ% ð9Þ A critical issue to address when using resampling methods is the number of samples that should be generated, which depends on the level of significance required and on the degree of change seen in the data. Usually, a more accurate estimate of the significance is achieved with more samples. On the other hand, when using permutation testing, all permutations (n! where n is the series length) could be generated. These are typically too many. However, 100–2,000 samples are usually recommended as sufficient and 1,000 samples were used for the current study. 3.4 Analysis framework Several datasets were derived from the primary monthly dataset for purposes of analysis of different aspects of the rainfall time series. The annual rainfall total was used for testing whether there have been trends in the overall totals. Analysis of the seasonal rainfall trends was divided into the short rains and long rains. The long rains and short rains have been variously quoted by researchers as occurring from March to May and October to December respectively. The March-May (referred to as MarMay) and October–December (referred to as OctDec) rainfall totals were used as the second pair of variables in the current study. On the other hand, it is also known that the onset of the rainy season varies from year to year and the actual rainy season may fall outside the above months in some years. An additional pair of variables where the maximum 3-month total rainfall for each 6-month period (January–June for the long rains and July–December for the short rains) was also calculated. The January-June 3 month rainfall variable is hereinafter referred to as JanJun3 while the JulyDecember rainfall variable is referred to as JulDec3. Comparisons between the ‘fixed period’ seasonal rainfall trends and the ‘variable period’ trends give a more clear understanding of the inter-annual temporal variability as well as giving some insight into the changes of the seasons within the basin. The assumption that the two 6month periods represent a discontinuous break in the two seasonal rainfall peaks is supported by plots of the longterm median monthly rainfall shown in Fig. 6. Each of the five datasets (annual, MarMay, OctDec, JanJun3, JulDec3) was further subdivided into different sub-periods in order to test for trend in the different periods. These are 1. Whole period of records available to test for general trends in the data 2. The period 1941–1980 to test for the impact of the heavy rains in the 1960s 3. The period 1961–1990 that is a WMO recommended baseline period for climate studies 4. The years 1971 to the last year of record for a given station. This was aimed at testing for the trends in the recent years in relation to more recent data excluding the 1961 rainfall event Temporal rainfall variability in the Lake Victoria Basin 127 Fig. 6 Long term median monthly rainfall for selected stations in the study area 3.5 Serial correlation 4 Results The existence of serial correlation in the data complicates the identification of trends. For example, a positive serial correlation can increase the expected number of false positive outcomes for the Mann-Kendall test (Burn and Hag Elnur 2002). Serial correlation coefficients for lag 1 and lag 2 years in the annual rainfall series for each station used in this study were computed and tested for their significance at the 5% level. The assumption was that after a lag of 3 years, any correlation in the data is not due to serial correlation especially if lag 1 and 2 correlations are not significant. The results revealed that only one station showed a significant serial correlation with a lag of 1 year. Therefore, no further action was taken for the whole series as independence held for the majority of the stations. The results below are presented for annual and seasonal (long and short rainy seasons) analyses. First we carry out an assessment of the mean rainfall variation for all stations in the basin to identify periods of significant departure from the long-term mean (trend or step) including El Niño years. We then present results of analyses at the individual stations for the different cases that were introduced in section 3.4. 4.1 Annual and seasonal rainfall variation The pattern of the MarMay rains is much closer to that of the total annual rains than the pattern of the OctDec rains (Fig. 7). The mean rainfall for the annual total, MarMay total and OctDec total are 1,202, 465 and 314 mm respectively. On average, the MarMay and OctDec rainfall 128 M. Kizza et al. totals account for 65% of the mean rainfall in the basin. If the 3-month maximum rainfall (JanJun3 and JulDec3) is used instead, the average contribution of the rainy season to the total annual rainfall increases to 71%. The MarMay rains contribute 39% of the total annual rainfall while the OctDec rains contribute 26% though there are considerable fluctuations (23%–50% for MarMay rains and 15%–46% for OctDec rains). Years of anomalously high rainfall can be identified in all plots in Fig. 7. For total annual rainfall, the years include 1937, 1941, 1947, 1951, 1961, 1963, 1977, 1989, 1997 and 2001. For MarMay rainfall, the years that have high rainfall include 1931, 1942, 1951, 1963, 1970, 1981 and 2002, while for the OctDec rainfall, the years include 1941, 1951, 1961, 1963, 1972, 1982, 1989 and 1997. The principal driving mechanism of these extreme rainfall events has been established as a dipole reversal in atmospheric circulation and Indian Ocean sea surface temperatures (Conway 2002). Hydrometeorological anomalies in the region (especially the 1961 and 1997 events) have received considerable research attention in trying to understand their dynamics, spatial and temporal nature as well as their hydrological impacts (Kite 1981; Flohn 1987; Latif et al. 1999; Webster et al. 1999; Conway 2002). The 1961 rainfall event resulted in 2.5 m increase in the water level of Lake Victoria which caused widespread flooding. On the other hand, the 1997 rains caused rise of only 1.7 m in the Lake water level but with similar flooding effects. 4.2 Annual time series results 4.2.1 Case I (all available data at each station) Annual data for 6 stations (Jinja, Eldoret, Sotik, Musoma, Ngudu and Entebbe) show a positive trend (Fig. 8, A-1). Of the stations with positive trend, 5 are located in the north to north eastern part of the basin and only one (Ngudu) is located in the south. Only one station (Bungoma) has a negative trend. A similar pattern is followed by the step change results with all stations that have trend in the annual data also having step changes (Fig. 8, B-1). The years when the step changes occurred are: 1993 (Jinja), 1999 (Bungoma), 2000 (Eldoret), 1962 (Sotik), 1949 (Musoma), 1959 (Ngudu), 1987 (Entebbe). Total Annual Rainfall Rainfall (mm) 1600 1200 800 800 Rainfall (mm) March-May Total Rainfall 600 400 200 800 October-December Total Rainfall 600 400 200 0 20 10 1930 1940 1950 1960 1970 1980 1990 2000 No. of stations Rainfall (mm) Fig. 7 Total rainfall (annual, March–May, and October–December) (continuous line) for the study stations with the average (dash-dash line) and the 5 year moving average (dash-dot line) superimposed. The lower panel shows the number of stations used to compute the mean A closer examination of the plots in Fig. 7 shows that a large portion of the variability in the annual rainfall is contributed by the OctDec rainfall. Spectral analysis using the Fast Fourier Transform shows that the total annual series has peaks at 2.4, 3.5, 5.2 and 6.5 years. The MarMay series has peaks at 4.0, 5.2, and 6.5 years while the OctDec series has peaks at 2.4, 3.0, 5.2, 6.5 years. In the annual rainfall series, the dominant time scale of variability is 5.2 years which corresponds with the dominant time scale for the ENSO phenomena (Nicholson 1996). The 2.4 year peak can be associated with the quasibiennial oscillation (Rodhe and Virji 1976). Temporal rainfall variability in the Lake Victoria Basin 129 Fig. 8 Trend results (marked A) and step change results (marked B) for the annual time series for each of the four cases studied (all years, 1941–1980, 1961– 1990, 1971–end) 4.2.2 Case II (1941–1980) There is no evidence of trends in the data apart from Ngudu which has a positive trend (Fig. 8, A-2). This is also true for the step change results with only Ngudu showing a positive jump (Fig. 8, B-2) 4.2.3 Case III (1961–1990) The series for Bukoba and Kamenyamigo show significant negative trend (Fig. 8, A-3) with no trends detected at all the other stations. For step changes, four stations (Jinja, Musoma, Ngudu and Kamenyamigo) show significant negative step jumps occurring in 1961 and 1962 (Fig. 8, B-3). 4.2.4 Case IV (1971–end of each station’s series) The rainfall series for three stations (Entebbe, Jinja, and Eldoret) show positive trends while those for Biharamulo and Bukoba show a negative trend (Fig. 8, A-4). The stations showing positive trend are all located in the northern part of the basin while the stations showing negative trend are located to the south. 130 Annual rainfall for two stations (Entebbe and Eldoret) show positive step changes (Fig. 8, B-4) while annual data for three stations (Jinja, Bungoma and Sotik) show negative step changes. 4.3 Long rainfall season results 4.3.1 Case I (all available data at each station) The MarMay rains have two stations (Kericho and Kahunda) with a positive trend and two stations (Buvuma and Eldoret) with a negative trend (Fig. 9, A-1). The JanJun3 rainfall total for three stations (Eldoret, Sotik and Fig. 9 Trend results (marked A) and step change results (marked B) for the long rainfall season time series for each of the four cases studied (all years, 1941– 1980, 1961–1990, 1971–end) M. Kizza et al. Musoma) have a positive trend while two stations (Buvuma and Bungoma) have a negative trend. For stations with a negative trend, only Buvuma shows similar trends in the two time series. The sign of the trend for Eldoret is reversed from one time series to the next (positive for the MarMay rains and negative for JanJun3 rainfall). For the MarMay period, five stations in the northeast of the basin show step jumps (Fig. 9, B-1). Three of the jumps are positive (Bungoma, Kitale, Kericho), while two are negative (Buvuma, Eldoret). The step jump results show that for the JanJun3 period, one station (Biharamulo) has a positive jump while two stations (Bukoba and Bungoma) have negative jumps. Temporal rainfall variability in the Lake Victoria Basin 4.3.2 Case II (1941–1980) Only one station (Kamenyamigo) has a positive trend in the MarMay rains and one station (Kahunda) has a trend (positive) in the JanJun3 rainfall total series (Fig. 9, A-2). Only one station (Kitale) shows step jumps for both the MarMay series and the JanJun3 series (Fig. 9, B-2). However, the signs are reversed with the MarMay series having a positive jump while the JanJun3 rainfall total series has a negative jump. 131 station (Bungoma) shows a negative trend for the OctDec rains (Fig. 10, A-2). These stations are uniformly spread within the basin with no clear spatial pattern. There is no evidence of trend in the JulDec3 time series. None of the OctDec time series shows any significant step changes despite seven of them having significant trends (Fig. 10, B2). However, for the JulDec3 rainfall, two stations show negative jumps (Kericho, Sotik) while Entebbe shows a positive jump. 4.4.3 Case III (1961–1990) 4.3.3 Case III (1961–1990) For the MarMay, Kamenyamigo has a positive trend and Mwanza has a negative trend (Fig. 9, A-3). One station (Kahunda) has a positive trend in the JanJun3 rainfall while Bukoba has a negative trend in the same series. No step jumps are detected in the MarMay series (Fig. 9, B-3). In the JanJun3 rainfall, Bukoba has a positive jump, while Buvuma and Ngudu have positive step jumps. The OctDec rainfall totals for Kamenyaymigo and Kabale show negative trends (Fig. 10, A-3) while there are no trends detected in the JulDec3 rainfall series. The OctDec rainfall series for 10 stations (Jinja, Kitale, Kericho, Shirati, Musoma, Ngudu, Mwanza, Biharamulo, Kamenyamigo and Entebbe) show negative jumps (Fig. 10, B-3). The JulDec3 rainfall for Entebbe, Jinja, Sotik and Mwanza show negative jumps. 4.3.4 Case IV (1971–end of each station’s series) 4.4.4 Case IV (1971-end of each station’s series) The MarMay rains for Entebbe, Bungoma, Kitale, Kericho, Mugumu and Kahunda show positive trend while those for Eldoret and Mwanza show negative trend. The JanJun3 rainfall for Jinja, Sotik, Musoma and Ngudu show positive trend while those for Bungoma, Bukoba and Kabale have negative trend (Fig. 9, A-4). Two stations (Kitale and Kericho) have positive step jumps in their MarMay rains while the series for Kericho has a negative jump. On the other hand, step change results for the JanJun3 series show that the jumps for Bungoma and Nugudu are negative and positive respectively (Fig. 9, B-4). The OctDec rains for Jinja and Eldoret both show a positive trend (Fig. 10, A-4). The JulDec3 rainfall series for Sotik, Musoma and Kabale have negative trends. The OctDec rainfall series for Eldoret has a positive trend (Fig. 10, A-4). For the JulDec3 rainfall, two stations (Sotik and Shirati) have negative jumps while one station (Bungoma) has a positive jump. 4.4 Short rainfall season results 4.4.1 Case I (all available data at each station) The OctDec rains have eight stations with positive trend and one with a negative trend (Fig. 10, A-1). On the other hand the two stations from the JulDec3 show positive trend and one station shows a negative trend.For the OctDec data, five stations show positive step jumps and one station shows a negative jump (Fig. 10, B-1). All the six stations with step jumps also show trend apart from Eldoret. The JulDec3 rainfall shows two stations with positive jumps (Sotik, Ngudu) and one with a negative jump (Bungoma). 4.4.2 Case II (1941–1980) Six stations show evidence of positive trends (Buvuma, Sotik, Ngudu, Biharamulo, Mbarara, Entebbe) and one 5 Discussion The analysis for this study was based on data obtained from different sources. In some cases, data from different sources for a given station were combined to form a single time series. Information on the type of instrument or any instrument changes or changes in the station settings could not be obtained and therefore we could not relate the results to properties but we feel this is not necessary for the validity of the current analysis. Data quality was checked using visual inspection of rainfall plots with time to identify clearly erroneous values and double mass plots to check for non-homogeneity. Twenty rainfall stations were included in the analysis to test for the presence of significant trends and step changes in the Lake Victoria basin. For each of the stations, five time series were derived: annual rainfall totals, March–May (MarMay) rainfall totals, October–December (OctDec) rainfall totals, long rains 3-month maximum rainfall (JanJun3) and the short rains 3-month maximum rainfall (JulDec3). Analysis was carried out for four time periods, i. 132 M. Kizza et al. Fig. 10 Trend results (marked A) and step change results (marked B) for the short rainfall season time series for each of the four cases studied (all years, 1941–1980, 1961–1990, 1971– end) e. the ‘All years’ case, the 1940–1980 case, the 1961–1990 case and the 1971–end case. Therefore, a total of 400 cases were analysed of which 65 (17%) had significant trends. Of the stations showing significant trend, 43 cases (67%) are positive trends and 22 (33%) are negative (Table 2), which suggests that the positive trends predominate in the basin over the twentieth century. For stations with significant trend based on more than 60 years of recording the trend represents an increase of 2–4 mm per year. This translates to a rainfall increase of about 24% over the twentieth century. Other studies have also found positive trends in the Lake Victoria basin. Rodhe and Virji (1976) did not find evidence of long-term trend in six gauges around the basin which was probably due to the fact that the data used were different from ours. However, Hulme et al. (2001) computed a positive trend giving an increase of between 10–20% or more in the annual rainfall for Lake Victoria basin over the period 1901–95. It is clear that annual rainfall variability in the basin is strongly influenced by variations in the ‘short rains’ which generally occur from October to December. Most of the stations whose annual rainfall data show trends also had significant trends in their October to December rains. For example, for the all-years-case, four of the stations whose annual rainfall shows a positive trend (Fig. 8, A-1) also have a positive trend in their OctDec rainfall (Fig. 10, A-1) Temporal rainfall variability in the Lake Victoria Basin 133 Table 2 Number of stations with significant trend for all the periods considered Time series Positive Negative Total Annual MarMay OctDec JanJun3 JulDec3 Total 10 8 16 7 2 43 5 4 4 5 4 22 15 12 20 12 6 65 suggesting a positive correlation between the two. A scrutiny of Fig. 7 will show that the above-average mean annual rainfall in the region since 1960 can be accounted for by higher than average mean rainfall for the short rainy period, with little or no trend observed in the rainfall for the long rainy period. The total upward trend in the short rains (OctDec) is about 30%. Several studies have shown that the short rains in the region are strongly influenced by the ENSO phenomenon (Mutai et al. 1998; Indeje et al. 2000). Towards the end of the twentieth century, El Niño events tended to be more frequent (WMO 2003), thereby explaining the upward trend in OctDec rainfall. However, other factors like Indian Ocean sea surface temperatures have also been proposed (Mutai et al. 1998; Camberlain et al. 2001; Mistry and Conway 2003). There is a strong similarity between stations showing significant trends and those showing significant step changes in the annual and MarMay time series (Table 3). This suggests that long-term changes in precipitation in the study area are due to the presence of periods with increased precipitation and are not purely monotonic in nature. The OctDec step change results are strongly influenced by results for the 1961–1990 time period, which show a significant negative step change for most of the stations in the basin which can be attributed to the anomalously heavy rains in the early 1961 followed by relatively high rainfall in 1962–1964 (Fig. 10, B-3). The extreme rains in 1961 and 1997 were studied by Conway (2002) who showed that the two events were associated with a dipole-like reversal of Indian Ocean sea surface temperatures. In addition, 1997 was a strong El Niño year. The 1961 and 1997 events were Table 3 Number of stations with significant step jumps for all the periods considered Time series Positive Negative Total Annual MarMay OctDec JanJun3 JulDec3 Total 9 6 6 3 4 28 8 3 11 6 9 37 17 9 17 9 13 65 similar in spatial and temporal characteristics and occurred mainly in the short rains period (October–December). The two events had far reaching hydrological impacts in the regions (including record river flows and flooding) with large socio-economic consequences (Conway et al. 2005). Other years with extreme rainfall include 1937, 1941, 1947, 1951, 1961, 1963, 1977, 1989, 1997 and 2001. The presence of trends in the data can also be classified by location of stations. Using this approach, 28 of the 43 cases with positive trends are located in the northern part of the basin, while 15 are located in the southern part. On the other hand, there are only weaker patterns in the distribution of the negative trends with 13 in the northern part and 9 in the southern part of the basin. The positive step jumps are similarly distributed with 19 positive jumps in the northern part and 9 positive jumps in the south. There are also 26 negative jumps in the north and 11 negative jumps in the south. Therefore, the trends and step jumps are more likely to occur in the northern part of the basin than in the south. 6 Conclusions We have shown that the Lake Victoria basin experienced a predominantly positive trend over the twentieth century. The results are supported by other studies within the basin and also within the East African region generally. This means that assessments of future climate scenarios for the basin should allow for wetter conditions. The magnitudes and sign of the trends depend on the data period used in the analysis and vary by station location with most of the stations with positive trends being located in the northern to north eastern part of the basin. However, the trends only represent long-term conditions and short-term variability may sometimes be more critical in assessing adaptation mechanisms within the basin. The influence of short rains on annual rainfall variability is discernible. Most of the stations whose annual rainfall data had trend also had significant trends in their October to December rains. Step change results show a more balanced picture between positive and negative changes within the basin. The step-change test results show a clear similarity to the trend test results, suggesting that the temporal rainfall patterns are not entirely monotonic but step wise with periods of dry years separated by wet years. The step change results are dominated by two periods with anomalously high rainfall in 1961 and 1997. The trend test results from analysing seasonal time series are quite different when we consider fixed time periods (March to May for the long rains and October to December for the short rains) from when we consider variable time periods representing the three consecutive months with 134 maximum rainfall totals in 6-month periods per year (JanJun3 and JulDec3 series). This reflects the variability of the rainfall seasons and may also reflect shifts in the onset of the rainy season within the basin. Further studies could shed more light on the pattern of such shifts. The current study tested the existence of trends in only the rainfall data. Additional work is needed to address the issue of existence of trends in other hydrologic variables like discharge and evapotranspiration in order to get a clearer picture. Additional work is also needed to check whether the observed trends are linked to climate change or reflect natural variability. Acknowledgements This work was performed within the doctoral study programme of the first author at the Department of Earth Sciences, Uppsala University, Sweden and Faculty of Technology, Makerere University, Uganda. The study was funded by the Swedish International Development Cooperation Agency (Sida) through the Department for Research Cooperation (SAREC, Reference number 75007304). The authors are pleased to acknowledge this financial support. Appreciation is also extended to the Departments of Meteorology in Kenya, Tanzania and Uganda for granting accessibility to the rainfall data used in the study. References Anyah RO, Semazzi FHM, Xie L (2006) Simulated physical mechanisms associated with climate variability over Lake Victoria in East Africa. Mon Wea Rev 134:3588–3609 Burn DH, Hag Elnur MA (2002) Detection of hydrologic trends and variability. J Hydrol 255:107–122 Camberlain P, Janicot S, Poccard I (2001) Seasonality and atmospheric dynamics of the teleconnection between African rainfall and tropical sea-surface temperature: Atlantic vs. ENSO. Int J Clim 21:973–1005 Conway D (2002) Extreme rainfall events and lake level changes in East Africa: recent events and historical precedents. In: Odada EO,Olago DO(eds) The East African Lakes: limnology, paleoclimatology and biodiversity. Kluwer, Dordrecht, The Netherlands, pp 63–92 Conway D, Allison E, Felstead R et al (2005) Rainfall variability in East Africa: implications for natural resources management and livelihoods. Philos Trans R Soc 363:49–54 Davidson AC, Hinkley DV (1997) Bootstrap methods and their application. Cambridge University Press, Cambridge, UK, 582 pp Flohn H (1987) East African rains of 1961/62 and the abrupt change of the White Nile discharge. Paleoecol Africa 18:3–18 Flohn H, Burkhardt T (1985) Nile runoff at Aswan and Lake Victoria: a case of a discontinuous climate time series. Z Gletscherk Glazialgeol 21:125–130 Hamed KH, Rao AR (1998) A modified Mann-Kendall trend test for autocorrelated data. J Hydrol 204:182–196 Helsel DR, Hirsch RM (2002) Hydrological analysis and interpretation: statistical methods in water resources. US Geological Survey, Reston, VA Hirsch RM, Slack JR, Smith RA (1982) Techniques of trend analysis for monthly water quality data. Water Resour Res 18:107–121 Hulme M, Doherty R, Ngara T et al (2001) African climate change: 1900–2100. Clim Res 17:145–168 M. Kizza et al. Indeje M, Semazzi FHM, Ogallo LJ (2000) ENSO signals in East African rainfall seasons. Int J Clim 20:19–46 Kendall MG (1975) Rank correlation methods. Griffith, London Kite GW (1981) recent changes in level of Lake Victoria. Hydrol Sci Bull 26:233–243 Latif M, Dommenget D, Dima M, Grotzner A (1999) The role of Indian Ocean sea surface temperature in forcing East African rainfall anomalies during December–January 1997/98. J Clim 12:3497–3504 LVBC (2006) Special report on the declining of water levels of Lake Victoria, Lake Victoria Basin Commission, Kisumu, Kenya, 15 pp Mann HB (1945) Nonparametric tests against trend. Econometrica 13:245–256 Mistry VV, Conway D (2003) Remote forcing of East African rainfall and relationships with fluctuations in levels of Lake Victoria. Int J Clim 23:67–89 Mukabana JR, Piekle RA (1996) Investigating the influence of synoptic-scale monsoonal winds and mesoscale circulations on diurnal weather patterns over Kenya using a mesoscale numerical model. Mon Wea Rev 124:224–244 Mutai CC, Ward MN, Colman AW (1998) Towards the prediction of the East African short rains based on sea-surface temperatureatmosphere coupling. Int J Clim 18:975–997 Nicholson SE (1996) A review of climate dynamics and climate variability in Eastern Africa. In: Johnson TC, Odada EO (eds) The limnology, climatology and palaeoclimatology of East African Lakes. Gordon and Breach, Melbourne, Australia, pp 25–56 Nicholson SE, Yin X (2002) Mesoscale patterns of rainfall, cloudiness and evaporation over the Great Lakes of East Africa. In: Johnson TC, Odada EO (eds) The limnology, climatology and palaeoclimatology of East African Lakes. Kluwer, Dordrecht, The Netherlands, pp 93–119 Ntiba MJ, Kudoja WM, Mukasa CT (2001) Management issues in the Lake Victoria watershed. Lakes Reserv Res Manage 6:211–216 Nyenzi BS (1990) An analysis of the interannual variability of rainfall over East Africa. Proceedings of the Second Technical Conference on Weather Forecasting in the Eastern and Southern Africa, Nairobi, Kenya, September 1990, pp 36–41 Ogallo L (1979) Rainfall variability in Africa. Mon Wea Rev 107:1133–1139 Ogallo L (1988) Relationships between seasonal rainfall in East Africa and the Southern Oscillation. J Clim 8:31–43 Ogallo L (1989) The spatial and temporal patterns of the East African seasonal rainfall derived from principal component analysis. Int J Clim 9:145–167 Rodhe H, Virji H (1976) Trends and periodicities in East African rainfall data. Mon Wea Rev 104:307–315 Sawunyama T, Hughes DA (2008) Application of satellite-derived rainfall estimates to extend water resource simulation modelling in South Africa. Water SA 34(1):1–9 Sene KJ, Farquharson FAK (1998) Sampling errors for water resources design: the need for improved hydrometry in developing countries. Wat Res Manage 12:121–138 Sutcliffe JV, Parks YP (1999) The hydrology of the Nile. IAHS Special Publication no. 5, IAHS, Wallingford, UK, 179 pp Tate E, Sutcliffe J, Conway D et al (2004) Water balance of Lake Victoria: update to 2000 and climate change modelling to 2100. Hydrol Sci J 49:563–574 Trewartha GT, (1981) The earth’s problem climates. Univ. Of Wisconsin Press, Madison, WI, 371 pp Webster PJ, Moore AM, Loschnigg JP, Lebden RR (1999) Coupled ocean–atmosphere dynamics in the Indian Ocean during 1997– 98. Nature 401:356–360 Temporal rainfall variability in the Lake Victoria Basin WMO (1982) Hydrometeorological survey of the catchments of Lake Victoria, Kyoga and Albert: project findings and recommendations. WMO/UNDP RAF/73/001, WHO, Geneva WMO (2000) Detecting trends and other changes in hydrological data. WMO/TD No. 1013, WHO, Geneva, 168 pp WMO (2003) Climate in the 21st century. Cambridge University Press. Cambridge, UK, 240 pp 135 Worsley KJ (1979) On the likelihood ratio test for a shift in location of normal populations. J Am Stat Assoc 74:365–367 Xu ZX, Takeuchi K, Ishidaira H (2003) Monotonic trend and step changes in Japanese precipitation. J Hydrol 279:144–150 Yin X, Nicholson SE (1998) The water balance of Lake Victoria. Hydrol Sci J 43:789–811 Yin X, Nicholson SE (2002) Interpreting annual rainfall from the levels of Lake Victoria. J Hydrometeorol 3:406–416