Assessing current and future spatiotemporal precipitation variability and trends over Uganda, East Africa, based on CHIRPS and regional climate model datasets

The lack of reliable rainfall projection records remains a major challenge to Uganda. In the advent of extreme wetness or drought events, reliable rainfall estimates for local planning and adaptation are essential. The present study used two main datasets to conduct a historical analysis from 1981 to 2019, coupled with future projections under representative concentration pathway (RCP 8.5) for the period 2020–2050. Historical analysis revealed bimodal annual rainfall patterns for March–May (MAM) and September–November (SON) gradients representing heavier to lighter rainfall events, respectively, over the study area. Investigation of recent trends in rainfall patterns revealed an upward trend from 2010 onwards in annual and seasonal rainfall. Moreover, results for future projections show wet conditions are projected to occur over the study area between the months of April/May and October. Contrarily, March is likely to experience a reduction in wet conditions. Mann–Kendall test employed to make future projections of rainfall depicted decreasing patterns during MAM season whilst increasing tendencies with strong shift was highlighted for SON season over the study region. Meanwhile, annual projections indicate huge variations with linear trends showing a marginal increase as compared to historical trends. Findings would serve as baseline print to propel further studies that could delve into impact analysis of drought extreme events which pose significant threats to the agricultural sector which is heavily reliant on rainfall.


Introduction
Many African nations rely heavily on rainfall for agricultural activities, hydroelectric production and water supply for their day-to-day activities. A variability in rainfall occurrence (i.e., below normal leading to drought or above normal resulting to flood incidences) has far-reaching consequences on the economic stability of many regions. Thus, an accurate rainfall quantification remains a paramount process for sustainable development. Meanwhile, recent decades have been characterized by the emergence of extreme Communicated by Stephanie Fiedler.
1 3 climatic events over many countries in both hemispheres, mainly due to global warming (Alexander et al. 2006;Seneviratne et al. 2013;Sillmann et al. 2013). Unprecedented changes in global climatic conditions have been induced by anthropogenic activities, resulting from an increase in the concentration of greenhouse gases (GHGs) in the atmosphere (IPCC 2013).
Presently, the outcome of several studies has shown developing countries would bear the unfortunate impacts of extreme weather events such as drought, floods, heatwaves, tropical cyclones, and wildfires as compared to mid-latitude and northern hemispheric nations (Seneviratne et al. 2012;Niang et al. 2014;Reliefweb 2020;Eckstein et al. 2020). Climate hazards would influence agricultural productivity (food security) in the sub-Saharan region (Parry et al. 2005;Schlenker and Lobell 2010). This threat could be further exacerbated by the increasing population, which is estimated to increase at approximately 4.8% per annum (FAO 2013). A recent report by FAO (2017) states about one-third of the human population is at risk of undernourishment in the East African region compared to other areas. This may be partly due to the declining tendency of 'long rains' which occurs between March and May (MAM) (Williams and Funk 2011; Lyon and Dewitt 2012;Liebmann et al. 2014). The situation is worsened by the uncertainties in future projections that continue to exhibit a 'paradox' scenario . For instance, Tierney et al. (2015) show future projection of MAM rainfall is likely to continue exhibiting observed negative trends while other studies Ongoma et al. 2018) reported increasing trends of precipitation towards the end of the twenty-first century. Such situations continue to pose confusion to all relevant stakeholders, thereby inhibiting progress planning and development policy.
Attributions to the uncertainties in the projections studies point to a number of factors such as the systematic and unsystematic biases in model datasets or methods accounting for natural climate variability such as El Nino-Southern Oscillation or warming of tropical Oceans (Giannini et al. 2005;Eyring 2019). Other studies, e.g., Christensen et al. (2008), Teutschbein and Seibert (2010) and Giorgi and Gutowski (2015), show that discrepancies in rainfall projections are sourced from the parameterization schemes in the global climate models (GCMs), such as those used in the fifth Coupling Modelling Inter-comparison Project (CMIP5) (Teutschbein and Seibert 2013). Thus, a number of recent studies Endris et al. 2013) point to the need for consideration of employing dynamically downscaled regional climate models (RCMs), such as those from Coordinated Regional Climate Downscaling Experiment Program (CORDEX) for future climate prognoses. Recent studies have proven the better performance of RCMs in simulating the East African climate with the prospect of improvement in impact analysis (Endris et al. 2013;Osima et al. 2018;Ayugi et al. 2020a).
In Uganda, existing studies show average rainfall in the country has decreased by 12% (Ssentongo et al. 2018;Alex et al. 2019). The occurrence of single large-scale events like droughts, floods, variable onset and offset of rainfall, long dry spells is notable and is consistent with the prediction of IPCC (2014). An average annual rainfall range of 500.0-2500.0 mm is recorded in Uganda, whereas spatial variability in this range is large (Basalirwa 1995;McSweeney et al. 2010), as noted in other parts of East Africa (Hession and Moore 2011). With such changes in rainfall amount and trends, information on spatiotemporal changes would be valuable in developing preparedness measures as well as provision of early warning systems (Omondi et al. 2014). This is of particular importance as crop food production by local communities is directly affected leading to a reduction in the income of approximately 69% of the Ugandan subsistence population (Gollin et al. 2016). Interestingly, despite the notable observations of changes in climate patterns over the study area, the existing ground-based datasets are sparsely distributed, hence, cannot capture local changes in far remote regions with limited or no gauge stations (Kizza et al. 2009;Diem et al. 2014).
To improve diagnosis and foster accurate forecast of climatic events, recent studies have gravitated towards RCMs and satellite-derived precipitation estimate (SPE) products as a way of detecting and projecting changes in climate incidences (Tian et al. 2010;Kidd et al. 2012;Nikulin et al. 2012). In addition, applications of SPE products in understanding extreme weather occurrences and analysis such as flooding or drought events have gained tremendous weight across the globe (Toté et al. 2015;Gebrechorkos et al. 2017). The present study sought to conduct an in-depth historical analysis of rainfall patterns over the study area using datasets from Climate Hazard Group Infrared Precipitation with Station (CHIRPS) while future projection was evaluated using regional climate model obtained from multi-model ensemble from the Rossby Centre regional Atmospheric model (RCA4). We characterized recent changes in rainfall patterns and examined future projections using model datasets with fine spatial resolution. Subsequent sections of the present study constitute Sect. 2 which entails description of the study area, datasets and methods. Section 3 presents the main results while Sect. 4 elucidates the summary, conclusion, and recommendations based on the findings.

Study area
Uganda is located in East Africa. The geographical coordinates are within longitude 29° E to 35.2° E and latitude 4.5° N to 1.5° S (Fig. 1). Neighboring countries include Kenya, South Sudan, Democratic Republic of Congo, Tanzania and Rwanda. Complex topography and numerous physical features ranging from high mountainous ranges, large lakes and rivers, rich highlands to plain lands characterize the study domain. For instance, Mount Elgon and Mount Rwenzori with approximately 4321 m and 5109 m in height respectively are situated within the borders of the study area (Bowden and Semazzi 2007). Other geomorphological features within the study area are large water bodies such as Lake Victoria and river Nile which generate meso-scale circulation within the region (Indeje et al. 2001;Ogwang et al. 2014). The rainfall climatology is mostly influenced by the seasonal oscillation of Inter-Tropical Convergence Zone (ITCZ) (Nicholson 2018), monsoon winds and subtropical anticyclones (Basalirwa 1995, Nicholson et al. 2017. Thus, most parts of the study area receive bimodal rainfall patterns with 'long rains' occurring in MAM whilst 'short rains' witnessed from September to November (SON) (Nsubuga et al. 2014;Ojara et al. 2020). This, however, results in a unimodal pattern as the distance from the equator increases steadily. There is also another rainfall band mostly in the northern parts of the country occurring between June and August. This condition is often attributed to the influx of moist westerlies from Congo basin (Basalirwa 1995). On the other hand, the temperature climatology is mostly warm temperate during the year with peaks of highs, experienced during the December-February (DJF) period and lows during the June-August (JJA) period (Omondi et al. 2014). More details regarding the circulation patterns are well described in studies conducted by Nicholson et al. (2018) and Camberlin (2018).

Data
This study utilized monthly precipitation datasets obtained from Climate Hazard Group Infrared Precipitation with Station (CHIRPS.v2) , as well as multi-model ensemble mean (MME) of five selected regional climate models (RCMs). The models were as follows: Model for Interdisciplinary Research on Climate (MIROC5), Commonwealth Scientific and Industrial Research Organization (CSIRO), Institute Pierre Simon Laplace Model CM5A-MR (IPSL-CM5A-MR), Max Planck Institute Earth System Model at base resolution (MPI-ESM-LR) and European community Earth-System (EC-EARTH). The listed RCMs simulations outputs were derived from the dynamical downscaling of CMIP5 GCMs using Rossby Centre regional  ). The RCA4 is a product of major enhancement on RCA3 based on model experimental design. Unden et al. (2002) and (Strandberg et al. 2014) gave detailed account regarding the physics of RCA4 model. The RCA4 simulation outputs are available on CORDEX-Africa domain at a spatial resolution of ~ 50 km × 50 km and temporal coverage ranging from 1951 to 2005 for historical runs and projections from 2006 to 2100.
Both CHIRPS and MME datasets of five better performing RCMs were recently evaluated by inferring their performance over the study domain (Ayugi et al. , 2020a. These models were appraised over the broader Greater Horn of Africa (GHA) against observed datasets using various scalar accuracy measures to assess their capability in reproducing fundamental precipitation characteristics over the study domain. The aforementioned studies used mean seasonal, annual, and inter-annual variations as a way of assessing their skillful simulation of precipitation over the region. Besides, a detailed statistical evaluation was employed to compare the model's performance. They included correlation coefficient (CC), mean bias error (MBE), and root mean square difference (RMSD), amongst the reanalysis and simulated precipitation cycle by the RCA4 models. Finally, the v2 against in-situ datasets based on (a) monthly distribution presented using the ECDF plots for 1981-2017 over Uganda. The second plot, b shows the annual cycle of precipitation products during the study period derived from sixteen available groundbased datasets distributed across the country, as shown in Fig. 1 Table 1   model's skill to simulate observed precipitation was tested using skill score, thereby identifying the five out of ten models evaluated in that study. Consequently, the CHIRPS data cover the period 1981-2019. The CHIRPS datasets were validated against the available ground-based datasets to ascertain their performance on a monthly and annual time scale ( Fig. 2; Table 2). The MME of five RCMs was employed to delineate the future trends and variability of climatic features over the study domain. Historical analysis for validation on the performance over the study domain was performed during 1981-2005, while projections for future climatic trend and variability were assessed under high emission scenario of Representative Concentration Pathways (RCP 8.5) for the 2020-2050 period. A summary of all model datasets used is shown in Table 1, indicating the type, source and resolution. All datasets were re-gridded using the bilinear interpolation technique to 0.5° × 0.5° spatial resolution in the present study. This was aimed to achieve uniform grids for analysis since the gridded datasets were of varying resolutions.

Validation of datasets
First, the present study validated the performance of both CHIRPS datasets and RCMs by calculating CC, RMSD and MBE. These metrics have been employed by so many studies in evaluating model simulation of climate variables (Wilks 2006;Chai and Draxler 2014;Ayugi et al. 2020a). The mathematical formulas of the metrics employed are shown in -the following equations: where M and O are the model simulated and observed values, respectively. I refers to the simulated and observed pairs and N is the total number of such pairs being evaluated.
Taylor diagram and empirical cumulative distribution function (ECDF) were used to show the comparison of the aforementioned datasets over the study area. Taylor diagram is a graphical illustration showing the similarity of two patterns in terms of their CC, centered RMSD, and the amplitude of their variations (represented by the standard deviation) (Taylor 2001). All the three metrics were presented on one plot as illustrated mathematically in Eq. 4. The plot was useful in evaluating multiple aspects of complex models or in gauging the relative skill of many different models (IPCC 2001).
where E is centered RMSDC is correlation coefficient m and o are standard deviation for the model and reference or observed datasets, respectively. A cumulative distribution F(x) can be defined as the proportion of observations lying below a certain value x as employed by Akinsanola (2017). The cumulative distribution of CHIRPS was compared with that of ground stations.

Spatiotemporal analysis
The study computed seasonal mean monthly rainfall by averaging 3 months total during the rainy seasons (i.e., MAM and SON). In addition, mean annual rainfall was derived by computing the average of all the monthly rainfall values over the study area. The evaluation of spatiotemporal patterns of rainfall for historical and future projections was conducted using anomaly test, thus, probability density function (PDF).
Anomaly is calculated from the following wquation:

Trend analysis
Precipitation trends are computed by fitting a linear model, using nonparametric Mann-Kendall test. The magnitude of change is computed using Sen's slope technique. The two main approaches were used to examine past and projected tendencies of precipitation. First, we employed the Theil-Sen slope technique to appraise the long duration tendencies. This method was used to evaluate the magnitude of the slope of the linear trend for a given data (Sen 1968). The method was considered to be effective since it is not influenced by any extreme distribution and does not entail any normal distribution of the residuals. Numerous studies have utilized this approach to examine the linear tendencies of hydroclimatic variables across various domains (Wang et al. 2018;Mumo et al. 2019;Ongoma et al. 2020). Mathematical expression explaining this approach is presented as follows: where i is the number of the time steps, X k and X l are the data points at point k and l respectively. In this case, k must be greater than l . In case of only one datum in each period, then n can be expressed as; n = n(n−1) 2 . When there are multiple data in one or more time periods, then n < n(n−1) 2 , where n represents the time steps. The n values of SSE are arranged from smallest to largest. The Sen's slope estimator is calculated as The trend is indicated by the sign of the Q med while its value portrays the degree of the slope. The confidence interval of the Q med at certain probability is then obtained to check if the slope is statistically significant at zero. The confidence interval of the calculated gradient is shown in Eq. (8) (Hollander et al. 2013): where the variance of S v is illustrated in Eq. 11. z 1−a∕2 is the tabulated value obtained from the t table. According to Gilbert (1987), the lower and upper limits levels of the significant bands,Q min , and Q max , are M 1 th largest and the M 2 + 1 th largest of the N ordered slope estimates. In this case, M 1 = N−C a 2 and M 2 = N+C a 2 . In this study, the slope will be considered statistically significant at ( = 0.01) and if the two limits ( Q min and Q max ) have the same sign.
Second, the study further employed Mann-Kendall (MK); (Mann 1945;Kendall 1975) and Sequential Mann-Kendall (SQMK;Sneyer 1990) to determine the significance of the trend and possible abrupt changes in the time series. The MK test is a rank-based non-parametric method that checks the existence of a trend in a time series against the null hypothesis of no trend. Several existing pieces of literature have applied the MK test (Ongoma and Chen 2017;Ayugi et al. 2018Ayugi et al. , 2020b. Standardized MK trend statistics is calculated using the mathematical expression shown in the following equation: where x i and x j are sequential data for the i th and j th terms, n is the sample size and (7) when n is an even number .
A hypothesis is set as follows: H 0 null hypothesis signifies no trend. Alternative hypothesis, H 1 indicates the presence of trend, either increasing or decreasing monotonic trend.
The variance was calculated using the following equation: The probability associated with S and the sample size n is calculated to assess the significance of the trend. The scores of Z values also show the significance of the trend where the negative and positive scores of Z values denote downward and upward trends, respectively. A two-tailed test, at a given α level of significance, H 1 is accepted if where Z 1− ∕2 is calculated from the standard normal distribution tables. The probability associated with MK and sample size n is computed to statistically quantify the significance of the trend. The normalized test statistic, Z, is calculated using the following equation: The trend is considered to be decreasing if Z is negative. For sequential Mann-Kendall (SQMK) test, forward sequential statistic: u (t) and backward sequential statistic u′ (t) by Sneyers (1990) from the progressive analysis of MK test was used to investigate the change in trend of rainfall with time. In the computation, the test compared the relative magnitudes of data instead of the data values directly. In this case, u (t) is the standardized variable that has a unit standard deviation and a zero mean. The progressive MK values u (t) and u′ (t) were calculated using the MK test for each data set, from the start to the end of the study period. In the plot of sequential MK, the confidence limits of the standard normal Z values are at α = 5%. The upper and lower confidence limits, therefore, correspond to + 1.96 and − 1.96, respectively. A significant trend is noted if the progressive MK values cross either confidence limit lines at the 5% significance level. (12)

Validation of datasets
distribution and annual cycle during the study period. These were derived from sixteen available ground-based datasets distributed across the country as shown in Fig. 1. A summary of statistical metrics detailing the models' performance against observed datasets is shown in Table 2. Our results demonstrate that CHIRPS.v2 can reliably reproduce rainfall climatology over the study area. There is a strong agreement between CHIRPS datasets and ground-based on monthly distribution, with ECDF showing homogeneous patterns as observed along with the frequency distribution. Moreover, the CHIRPS products captured the annual cycle's bimodal patterns and seasonal peaks, as observed using in-situ datasets ( Fig. 2b and Table 1). However, aslight overestimation was observed in April when the study region experienced the highest rainfall amount. Interestingly, the month of October and November depicted an underestimation of ground-based data by CHIRPS.v2. Despite the reliability of CHIRPS.v2 in reproducing rainfall information, contrary performance was noted at Kitgum station. The CHIRPS.v2 showed a weak correlation (CC = 0.21) and high amplitude with RMSD = 99.99 mm/ month and Bias = 32.99 (Table 2). This calls for further investigation of the observed outlier performance on this particular station. Overall, the CHIRPS datasets can be employed as an alternative to in-situ datasets in a region characterized by scarcity of ground-based datasets for a timely exploration of ever-increasing climate extremes.
Furthermore, this study sought to validate the performance of the better performing RCA4 model as recommended in a recent study (Ayugi et al. 2020a) against the "observed" CHIRPS.v2 datasets over Uganda during 1981Uganda during -2005. Models were evaluated using robust statistical metrics such as CC, SD, RMSD, and MBE. Figure 3 shows the results of the performance of RCA4 models against the CHIRPS.v2 datasets. It demonstrated previous assessment's substantial similarity to few existing ground-based datasets over the study domain. The validation of the MME displays relatively better performance over the study region. For instance, the correlation coefficient shows 0.9 with CHIRPS. v2, while RMSD depicts a low amplitude of < 50%. Moreover, a relatively low standard deviation (about 0.65 mm/ month) is simulated against the observed value of 1.0 mm/ month. Therefore, the analysis elucidates the multi-model ensemble application for rainfall projections and impact analysis over the study area.
The results for CHIRPS and RCMs validation over the study region agree with many evaluative studies. For instance, the reliability of CHIRPS in reproducing regional rainfall information as compared to other satellite datasets across many regions in the sub-Saharan equatorial region as affirmed in various studies (e.g., Diem et al. 2014Diem et al. 2019bGebrechorkos et al. 2017;Kimani et al. 2017;Nicholson et al. 2019;Ayugi et al. 2019). It should be noted that the study area experiences heavy rains for MAM, and short rains are received for SON (Kizza et al. 2009;Ogwang et al. 2014;Gamoyo et al. 2015). The bimodal pattern is mostly influenced by the oscillation of Intertropical Convergence Zone (ITCZ) (Nicholson 2018;Yang et al. 2015;Nicholson et al. 2018).
Simultaneously, other studies have also substantiated that RCMs can generate high-resolution projections of climate events in many parts of the world as compared to Global Climate Models (GCMs) (Kidd et al. 2012;Nikulin et al. 2012). Recent evaluative studies on the performance of RCMs over the Greater Horn of Africa (GHA) identified the robust performance of RCMs derived from the Rossby Centre regional Atmospheric Model (RCA4) (Endris et al. 2013;Kisembe et al. 2019;Ayugi et al. 2020a). The validation of model performance over the study area reaffirms the trustworthiness and reliability of the RCMs in projecting the possible future changes amid climate change and global warming. Thus, this study employs the MME of the five models in the projection of precipitation change using the representative concentration pathways, proposed by Riahi et al. (2007).

Spatiotemporal variability of annual and seasonal rainfall
This analysis aimed to show the spatial and temporal patterns of rainfall change over Uganda during the recent period 1981-2019. The results are depicted in Figs. 4, 5,  respectively. Figure 4 shows an overview of the annual precipitation cycle and the rate of change per year over Uganda based on CHIRPS.v2 datasets for 1981-2019. It is apparent from the figure that the study area experiences bimodal rainfall patterns with a strong gradient of rainfall experienced during MAM period over the years along with the second rainfall band occurring during SON. The mean rainfall for MAM (SON) is 139.20 mm/month (125.71 mm/month).
The results further illustrate that the study area received less rainfall of < 48 mm/month during the DJF season, while a substantial amount of rainfall (107 mm/month) is noted during the JJA period. Remarkably, the yearly variation shows that the year 2012 recorded the highest amount of rainfall, whereas the least amount was noted in the year 2009. The observed anomalies depict a period of dryness (wetness) over the study area. To understand the recent trends experienced over the given period, we employed Theil Sen Slope Estimator to compute the trends. Findings for the rate of change are presented in Fig. 4b. As shown in the figure above, the results display a significant positive increase in October at 0.62 mm/year, while January and July depicted a negative insignificant decreasing pattern of − 0.178 mm/ year (− 0.311) mm/year, respectively. For seasonal tendencies, the MAM season shows a decreasing pattern compared to SON, which is considered a 'short rainy' season over the study region. Further analysis of the spatial distribution of seasonal rainfall (mm) and their respective probability density function (PDF) distribution over Uganda based on CHIRPS.v2 datasets for the period 1981-2019 is shown in Fig. 5. Figure 5a shows spatial patterns of MAM rainfall, while SON rains are presented in Fig. 5b. MAM receives more rainfall distribution compared to SON. Moreover, southern and eastern parts of the region experience a higher magnitude of rainfall, while the northeast depicts less precipitation amount of ≤ 50 mm/ month. The rainfall amounts over the northeastern zone are more pronounced during SON, with larger parts of the study area receiving fewer rainfall amounts than MAM season. Despite the higher rainfall amount experienced during MAM season, analysis of PDF that examines the relative likelihood in the distribution of variable from the mean shows that SON is likely to shift from negative to positive patterns. The results presented provide information on how the region's rainfall has varied both in time and space.
Numerous existing studies have attempted to establish an explanation of monthly variations, the trends, and mechanisms regulating the observed patterns. For instance, the seasonal variability of rainfall is regulated mainly from the complex interaction of weather systems such as; the bi-annual oscillation of ITCZ from North to South, the tropospheric systems of Quasi-biennial Oscillation (QBO), large-scale monsoon winds, and subtropical anticyclones (Mutai et al. 2000;Nsubuga et al. 2011). The ITCZ has a larger influence than the listed underlying mechanisms of seasonal rainfall patterns (Nicholson 2018). Meanwhile, the increasing amount of rainfall during the period of JJA, which was considered as the local dry season, is attributed to moist westerlies originating from the Congo basin resulting in enhanced rain during this season in the north and southwest when other parts of the country are cold and dry (Mchugh 2004;Kizza et al. 2009). On the other hand, Lake Victoria's lake/land breeze plays a significant role in driving the rainfall being experienced during the dry season of DJF (Nsubuga and Rautenbach 2017). The present study's findings correspond to past studies that noted the observed climatology of enhanced rainfall during JJA season (Nsubuga et al. 2014;Yang et al. 2015).
On the linear trend analysis, other existing studies equally noted a declining trend in rainfall during the local wet season of MAM (Funk et al. 2005;Lyon and Dewitt 2012;Liebmann et al. 2014). The season mentioned above is considered as the main growing season for agricultural activities that support 70% of the local economy (McSweeney et al. 2010;Ojara et al. 2020). This decrease would influence food security and livelihoods of the people. In the meantime, the spatial patterns of rainfall that depicted a higher rainfall amount during MAM season over southern and eastern parts could be attributed to the presence of large water bodies, i.e., Lake Victoria and complex geomorphology situated over the regions that receive the highest amount of rainfall (Basalirwa 1995;Indeje et al. 2001;Ogwang et al. 2014). The presence of high elevation along the eastern region produces leeward rain shadows and block the passage of rainbearing disturbances in other areas (Ogwang et al. 2014). On the other hand, the northeast locale is characterized by arid and semi-arid lands (ASALs) with scarce vegetation cover, low precipitation events, high evapotranspiration, strong radiation, and high wind patterns throughout the year. Such factors result to dry anomaly, which significantly impact community livelihoods and other ecosystem processes. Overall, the MAM season is mainly regulated by a combination of mesospheric features and atmospheric circulation, thereby contributing to higher rainfall amount observed as compared to the SON season (Yin and Nicholson 2002). The PDF analysis depicted changes in the shift for seasonal rainfall distribution has been attributed to the alteration of Walker circulation anomalies due to intense warming of Sea surface temperature (SST) along the western Indian Ocean (Lyon and Dewitt 2012). The implication of the changes in the seasonal rains is likely to impact farmers' uncertainties regarding crop growing seasons, which predominantly have been observed for MAM season (Matthew et al. 2015;Adhikari et al. 2015).

Interannual variability of seasonal and annual rainfall
Analysis of rainfall anomalies for seasons over the study region during the 1981-2019 period based on CHIRPS. v2 datasets is presented in Fig. 6. Examining the interannual variability of rainfall aids in the understanding of the main factors controlling the interannual variations. Typical wet and dry years were identified following Makkonen's (2006) recommendations on exceeding the standard deviation of +  (White and Tourre 2003). In this study, we analyzed decadal to show spatial variations of rainfall and to identify the years associated with climatic phenomenon regulating the interannual and interdecadal variability. The spatial anomalies were calculated based on CHIRPS datasets over the last four decades, thus, 1981-1990; 1991-2000; 2001-2010; and 2011-2019. The summary of the rate of change per decade is highlighted in Table 3. From the analysis based on rainfall decadal anomalies, it is apparent that varying tendencies were noted from 1 decade to another across various timescales. For instance, the study region witnessed a pronounced decreasing trend during MAM seasons through the last decade (2011-2019) though SON showed a decreasing trend during the first two decades (1981-1990 and 1991-2000) coupled with a recovery afterward. Moreso, MAM depicted distinct wetting trends covering most parts of the study area during the first three decades (i.e., 1981-1990, 1991-2000, and 2001-2010) although dry trends were observed in 2011-2019 (Fig. 7). The SON rains (Fig. 8) exhibited a reduction trend in the first 2 decades (1981-1990 and 1991-2000), followed by an increase  (2001-2010 and 2011-2019), probably reflecting an increased November rainfall over broader study locale (Spinage 2012). Overall, the observed patterns show a reversal in rainfall patterns over the study region with enhanced rainfall occurrence during the shortrain season whereas reduction was witnessed during the long rain season.
The result from the inter-annual variation analysis shows high spatiotemporal variability, which coincides with the anomalies in remote forcing events. For example, the primary controlling mechanism influencing the observed interannual variations in high/low rainfall events have been linked to the dipole reversal of atmospheric circulation and Indian Ocean sea surface temperatures (Saji et al. 1999). The rains are moderated by weather phenomena such as El Niño Southern Oscillation (ENSO) (Indeje et al. 2000;Ntale and Gan 2003) and the Indian Ocean Dipole (IOD) (Behera et al. 2006). The El Niño Southern Oscillation (ENSO) phenomena are strongly associated with the inter-annual variability of rainfall in this region (Indeje et al. 2000). This is reflected in the present study as years with El Niño events such as 2010 and 2012 were recorded as wet years whereas La Nina events coincided with dry years such as 1984, 1992, and 2008. The inter-annual variability of SON rains is mostly attributed to ENSO and IOD influence.
The decadal anomalies revealed an increase in anomalous events that have been mostly associated with positive Indian Ocean Dipole instead of the teleconnection patterns across the Pacific Ocean (Behera et al. 2006;Muhati et al. 2007). The increasing positive IOD phase results from the heightened warming of the Western Indian Ocean, which alters the changes in Walker Circulation (Behera et al. 2006;Ogwang et al. 2015). Several studies have reported that SON rainfall is mostly influenced by the changes in ENSO and IOD (Saji and Yamagata 2003;Manatsa et al. 2012). On the other hand, MAM rains occurred as a result of three main drivers, the Indian Ocean SST, the seasonal amplitude of the Madden Julian Oscillation (MJO), and the phase of the MacLeod 2019). The observed decline in MAM rainfall in the region has been prominently analyzed and discussed (Funk et al. 2008;Rowell et al. 2015). Williams and Funk (2011) attributed the continuous drying over the EA region to an anthropogenic-forced relatively enhanced warming of Indian Ocean SSTs, which extends the warm pool and Walker circulation westward, leading to a subsidence anomaly causing the drying effects. The observed trends have already affected many communities that rely on rainfall for their socio-economic activities (Funk et al. 2008).
The local impact of the interannual variability observed had immense impact on the livelihoods of people. For example, the periods of negative anomalies have resulted in droughts that have been identified to align with ENSOrelated anomalies. The impact of drought has remarkably altered water resources, ecosystem balance, and impacted agricultural activities. Previous work by Apuuli et al. (2000) noted how people were displaced during drought years with

Historical monotonic trends analysis
The present study utilized the Mann-Kendall trend test to detect the possible significant and abrupt changes in rainfall patterns over the study area for the given period . The results for MAM, annual, and SON time series show Z score of 0.315, 0.629, and 0.677 (Table 4), which is below the threshold value of 1.96, signifying slight positive tendencies. However, the variance (S) shows negative values for all the analysis indicating a reduction in rainfall for the given period. Figure 8 demonstrates results for sequential Mann-Kendall statistic values of progressive u (t) (solid red line) and retrogressive u′ (t) (black dotted line), derived from CHIRPS.v2 precipitation datasets for (a) MAM, (b) SON, and (c) annual mean over Uganda during 1981-2019 period. Generally, the SQMK indicates an upward trend in annual and seasonal rainfall over the study area which is indicated using progressive statistic.  Essentially, the analysis demonstrates a positive insignificant trend even though the amplitude varies from one season to another. For instance, the MAM season (Fig. 9a) showed a decreasing pattern at the start of the study period until 1984 when it experienced a noteworthy decrease with a reversed upward trend, experienced thereafter which was eventually sustained until the end of the study period. During this period, change occurred in 2011 as evidenced by the intersection of u (t) and u′ (t).
On the other hand, the SON season (Fig. 9b) experienced several changes, with three major intersections encountered across the study duration. The annual rainfall record presented similar patterns as the MAM season with three changes occurring in 2010, 2013, and 2018 where the forward and retrograde line crossed each other (Fig. 9c). The results are in congruence with previous findings (Funk et al. 2008;Williams and Funk 2011;Lyon and Dewitt 2012) that reported an abrupt shift in rainfall tendency from wet years to an almost continuous period well-below average rainfall over the study domain. This trend has been consistent in the three decades, following the early years of 1980s, as well as series of dry years in 2010.
Fluctuations in water resources are evident based on the decline in water levels of Lake Victoria mainly as a result of the impact of rainfall variations (Kull 2006). The decreasing trend in MAM total rainfall impacted negatively through a reduction in the number of wet days, thereby affecting the cropping cycle and maturity of staple foods. In some regions, farmers had noticed seasonal rains delayed. Again, they noticed rains received as well as rainfall duration during such periods were short, intensive and erratic. Such scenarios had immense impact on the overall output of food supply and water table. This in effect, calls for the implementation of rapid growing crops and systems that required less water as an adaptive mechanism to cope with changes associated with rainfall. Farmers and all relevant stakeholders need to shift the planting season from MAM to SON which currently experiences much rainfall occurrence as compared to MAM.

Future projections
This study used the MME ensemble to depict the future projections under RCP8.5 in a bid to demonstrate the nearterm changes in rainfall over the study area. Figure 10 shows the monthly changes of rainfall in the future for 2020-2050 period under RCP8.5, relative to the baseline data for 1981-2019 over Uganda. Results show that wet conditions are projected to occur over the study area between April-May and SON season. Contrarily, the month of March is likely to witness a reduction in wet conditions. Projected patterns show further reduction as compared to the baseline period in rainfall during the months of June-August (JJA) and December-February (DJF). Further analysis of projected linear trends of seasonal and annual rainfall over Uganda for the given period (2020-2050), relative to 1981-2019 is shown in Figs. 11, 12. The distribution and analysis were to explain projected linear patterns, relative to the baseline period, which highlights the expected changes and their respective magnitude over the study area. Linear trends (Fig. 11) and PDF (Fig. 12) for seasonal precipitation illustrate decreasing patterns during the long rainy season (Figs. 11a and 12a) whilst increasing tendencies with substantial shift is depicted during the SON season (Figs. 11b and 12b) over the study region. The resultant implication is an upsurge in rainfall amount received during SON over the study domain. Meanwhile, annual projections (Figs. 11c and 12c) indicated huge variations, with linear trends showing a marginal increase as compared to historical tendencies.
It should be noted that projections were derived from the multi-model ensemble of five better performing RCMs over the study region (Endris et al. 2013;Ayugi et al. 2020a). This was to enhance the confidence in projections due to reduced inter-model uncertainty and minimum biases. The results of the annual cycle (Fig. 10) agree with existing studies on projected annual rainfall across the East African (EA) region. For instance, Ongoma et al. (2018) projected an increase in rainfall during the month of April and September while May to August was likely to witness a reduction in wet conditions. The observed and projected dry conditions during JJA could be attributed to changes in above-normal sea-level pressure over Bombay and Indian drought in July through to September, thereby contributing to dry conditions over Uganda, and other parts of EA region (Camberlin 1997;Patricola and Cook 2011). Nevertheless, the projected wet conditions during the April/May period could be ascribed to a strong Somali jet that is southerly over the Horn of Africa and turns westerly into the Arabian Sea, thereby transporting moisture over EA (Hastenrath et al. 2011). The analysis for linear trends as presented in Figs. 11, 12 which depicts negative trends for annual and MAM season could be attributed to reduced rainfall trends during the long rainy season which may account for the higher rainfall amount recorded as compared to the SON season. The projected decrease agrees with recent studies conducted by Rowell et al. (2015) and Tierney et al. (2015) that equally asserted that the projected long rains would continue to exhibit decreasing patterns until 2060 before it depicts recovery trends towards the end of the century. Apparently, other studies demonstrate correspondingly opposite trends over the EA region with increasing patterns being reported contrary to observed reduced trends recorded Ongoma et al. 2018). This discrepancy in projected rainfall trends is termed as the East Africa "climate paradox". This further calls for more evaluative studies to ascertain clarity in projected patterns .
Conversely, the substantial increment in precipitation during SON rains could be attributed to changes in teleconnection patterns such as ENSO and IOD (Nicholson and Kim 1997;Indeje et al. 2000;Endris et al. 2019). These studies illustrated a higher warming rates over the western Indian Ocean than the eastern Indian Ocean. This may lead to intensified positive IOD occurrences, leading to stronger spatial coherence in precipitation patterns. The decisive shift, indicating an increase in rainfall event, despite projected future warming, agrees with a study conducted by Kent et al. (2015) that revealed a lack of correlation between uncertainty in global mean temperature and projected endof-twenty-first-century change in precipitation. Moreover, the study further noted that uncertainty in regional precipitation over the study region is predominantly related to spatial shifts in convection and convergence, associated with processes such as SST patterns and land-sea thermal contrast change. The conclusion in various studies that attempts to elucidate shifts in rainfall projections highlights the complexity of regional rainfall fluctuations, resulting in uncertainty about the magnitude of impacts, related to extremes weather events. Meanwhile, the projected reductions in rainfall amounts for MAM seasons and annual precipitations could be associated with the weakening of Walker Circulation over Indian and Pacific Ocean basins (Tierney et al. 2015).
In summary, the projected increase in April/May rainfall may be a relief to farmers who entirely depend on rainfall for their agricultural activities. The increased wetness would enhance farm outputs, thereby improving the overall productivity at the national level. However, the projected reduction in JJA rainfall, attributed to Congo westerlies would adversely impact communities, mostly located in southwestern Uganda (Diem et al. 2019a). For instance, the crop regions along the western belt would be affected by the changes in climatic conditions. Further, rapid population growth and an expansion of farming and pastoralism under drier climate regimes would dramatically increase the number of vulnerable communities for the next 30 years. The changes projected could also heighten community conflicts in the use of limited natural resources. Continuous monitoring of rainfall and its related impacts remains a crucial task with short term policies being formulated to reflect the present changes from time to time.

Conclusion and recommendation
Rainfall remains one of the most important climatic variables supporting the economies and livelihoods of many African countries. Slight variation or changes in its trends translate to massive impact on agriculture, energy, transportation and other climate-sensitive sectors of the economy. This present study, thus, sought to examine recent trends and possible future projection of rainfall over Uganda. The study utilized rainfall datasets obtained from Climate Hazard Group Infrared Precipitation with Station (CHIRPS) and multi-model ensemble (MME) of five regional climate (RCMs) datasets. The listed RCMs simulation outputs were derived from the dynamical downscaling of CMIP5 GCMs using Rossby Climate Modelling Atmospheric Centre (RCA4), originally developed by the Swedish Meteorological and Hydrological Institute (SMHI) under the CORDEX initiative. The CHIRPS dataset was employed for historical analysis while MME was employed for future projections under RCP8.5 scenario.
The results for multi-year monthly climatology show occurrence of bimodal rainfall patterns with strong gradient of rainfall experienced for the March-May season (MAM). The second rainfall band, on the other hand, occurred in September-November (SON). Results further revealed the study area received less rainfall of < 48 mm/month during the December-February season (DJF) while substantial amount of rainfall (107 mm/month) was recorded for the June-August season (JJA). An analysis of recent trends of rainfall revealed an upward trend in annual and seasonal rainfall over the study area after 2010. Furthermore, results for future projections show wet conditions are projected to occur over study area between April-May and October. On the contrary, the month of March is likely to witness a reduction in wet conditions based on study findings. Linear trends for seasonal rainfall show decreasing patterns for MAM season whilst increasing tendencies with strong shift are depicted during SON over the study region. Meanwhile, annual projections indicated huge variations with linear trends, illustrating a marginal increase as compared to historical tendencies.
The results of this study would contribute to the everpresent debate on projected expected changes in the region's rainfall variation which continues to draw much attention due to the many conflicting projection patterns. The results can be utilized by relevant stakeholders, including policymakers and farmers who largely depend on rainfed climatic patterns in the wake of global warming. Decision and policymakers ought to design appropriate policies that reflect all possible outcomes as various studies are inconclusive on the exact direction of future climate in the wake of global warming and climate change.