Projections of three important seasonal rainfall parameters—total precipitation (), wet-day mean () and wet-day frequency ()—considered to be relevant to crop agriculture were performed. Links between large-scale climate variables and local precipitation in Tanzania were investigated during the March–May (MAM), October–December (OND), and December–April (DJFMA) rainfall seasons. Variables found to have strong links were used to downscale the local precipitation for three future periods; near term, midcentury, and end of century. Downscaling models for , and were calibrated using observed large-scale seasonal rainfall and projected downscaled parameters were obtained based on rainfall simulations from ensembles of GCMs. The models’ skill scores were found to be sensitive to the domain size and number of leading EOFs used. The common EOF method employed in the downscaling modulated the skills depending on the GCMs used. The spread in the rainfall projections was found to be larger in OND and moderate in MAM and DJFMA. The multimodel mean projections in response to two RCPs (RCP4.5 and RCP8.5) suggest a shift toward wetter (drier) conditions () for OND (DJFMA) for all three periods. There is no uniform projection for MAM; some stations are projected to feature wetter and some drier conditions. In the midcentury and end of century, there is an increase of precipitation to about 40% for some areas getting OND rainfall and a decrease of precipitation up to about 10% for some areas getting MAM or DJFMA rainfall. Generally, the magnitude of change strongly differs across the areas.
Agriculture accounts for about 43% of the annual gross domestic products (GDPs) of East African countries (Omamo et al. 2006). Yet the farming in these countries is dominated by small holders who rely on rainfall (United Republic of Tanzania 2014; Waithaka et al. 2013). Rain-fed agriculture has remained the dominant source of staple food production and the livelihood foundation of the majority of the rural poor in sub-Saharan Africa (Cooper et al. 2008). As a consequence the sector has remained susceptible to climate change and variability. Changes in climate between and within the seasons (Rowhani et al. 2011) and changes in the mean climate and the magnitudes and frequency of the extreme events (Moriondo et al. 2011) have a significant impact on crop yields and livelihood as well as food security (United Republic of Tanzania 2014). It is therefore important to have an improved understanding of the influence of climate on agricultural production that is needed to cope with expected changes in temperature and precipitation, and hence to minimize an increasing number of undernourished people in food-insecure regions (Rowhani et al. 2011).
Although some general conclusions about climate change and their impacts have been drawn especially at large scales, the potential damage of climate change at a particular region or farm needs to be assessed under site-specific conditions (Li et al. 2011). Moreover, there is an increasing number of decision makers from different organizations seeking climate predictions of the regional and local changes that will impact people. Since global climate models (GCMs) are at coarse resolution such that they are unable to resolve features at regional or local scales, it is therefore necessary to downscale the GCMs in order to produce scenarios of climate change at the scale required by impact assessment studies.
The downscaling of GCMs in Africa has received relatively less attention as compared to other part of the world (Hulme et al. 2005; Shongwe et al. 2011). Moreover, existing studies on regional climate modeling have mostly focused on southern Africa (e.g., Hewitson and Crane 2006; Landman and Tennant 2000) and West Africa (e.g., Paeth et al. 2005; Penlap et al. 2004), with very few in East Africa (e.g., Moore et al. 2012; Taye et al. 2011). With the exception of the statistical downscaling works done by Tumbo et al. (2010) over Same district in Tanzania and Taye et al. (2011) over Lake Victoria basin, previous downscaling work already done in East Africa/Tanzania is based on dynamical downscaling (e.g., Moore et al. 2012; Omondi et al. 2014), with analyses given on a regional or country level, which is too course to draw conclusions on impacts at local levels. While the work of Tumbo et al. (2010) was based on only one location in Tanzania, in this study we also propose the use empirical–statistical downscaling (ESD) covering different locations in the country representing homogeneous climatic zones (see, e.g., Ng’ongolo and Smyshlyaev 2010), also using several GCMs.
The current study therefore attempts to address the shortcomings of using coarse spatial resolution models for assessment of climate change for impact studies (e.g., agriculture). In particular, we aim to use ESD to explore changes in climate at local level. We rely on the hypothesis that the local rainfall is connected with larger-scale conditions and that we can use GCMs to describe the statistics of year to year variations in seasonal precipitation in Tanzania. The objectives of this work are 1) to assess and establish links between large-scale circulation patterns and local precipitation in Tanzania, 2) to develop ESD models based on the identified links, 3) to assess the skill of the ESD models, 4) to create climate change scenarios using ESD models and GCMs projections, and 5) to identify causes of the spread and differences in the multimodel downscaled results.
Regression-based ESD utilizes systematic dependency transfer functions between large-scale circulation (predictors) variables, , and regional or local climate variables, (predictand), to derive regional climate information (Leung et al. 2003). This function is typically established through training and validating the models using point observations and gridded reanalysis data (Fowler et al. 2007). The derived statistical relationships are then applied to the large-scale climate variables from GCM projections to estimate corresponding future local and regional characteristics (climate change). Practical considerations for developing an ESD scheme involve the selection of appropriate predictor variables, predictor spatial representation, domain size and location, types of transfer functions, and definition of temporal scales such as season (Leung et al. 2003).
The utility of ESD depends upon four assumptions: 1) high-quality large-scale and local data are available for an adequately long period to establish robust relationships in the current climate, 2) the link between predictands and predictors is strong in order to explain satisfactorily the local climate variability, 3) the predictor variable is well simulated by the GCM, and 4) relationships derived from recent climate remain relevant in a future climate (stationarity) (Benestad 2001a; Busuioc et al. 2001; Wilby and Wigley 2000).
The use of ESD has the advantage that 1) it is computationally very inexpensive, 2) it can provide information at smaller scales than the GCM-scale output, 3) it can be used to derive variables that are not available from regional climate models (RCMs), 4) it is able to directly incorporate observations, 5) it provides diagnostics and insights (e.g., a picture of how large and small scales are related), and 6) it can be used to assess the models. The main drawbacks of this method are that 1) ESD has underlying assumptions that can only be partly tested (Willems et al. 2012), 2) strong predictor variables do not always exist, and 3) long time series of relevant data are required to form the relationships (Jones et al. 2004; Leung et al. 2003).
2. Data and methods
a. Study area
Tanzania (Fig. 1), one of the East African countries, is located in the Southern Hemisphere. Its climate consists of two rainfall regimes; a bimodal one with rainfall seasons in March–May (MAM) and October–December (OND), and a unimodal one with a rainfall season in December–April (DJFMA). These rainfall regimes are controlled by the northward and southward migration of the intertropical convergence zone (ITCZ). The bimodal rainfall regime is observed over the areas located in the northern parts of the country while the unimodal rainfall regime is more typical for the southern, central, and western parts of the country, coinciding with the austral summer (December–February). The remaining months, from June to September and January to February (for bimodal areas) are relatively dry.
It is widely established that rainfall over East Africa varies in space and time (Camberlin and Philippon 2002; Indeje et al. 2000). The variability has been linked to the presence of local features (Indeje et al. 2000), variations in the dominant large atmospheric circulation (Camberlin and Okoola 2003; Mapande and Reason 2005), and global warming (Schreck and Semazzi 2004).
Some circulation oscillations have been linked to the variation of East African seasonal rainfall, ENSO (Indeje et al. 2000; Schreck and Semazzi 2004; Kijazi and Reason 2005; Ogallo 1988), the zonal circulation in the equatorial Indian Ocean (Black et al. 2003; Kabanda and Jury 1999; Marchant et al. 2007; Saji et al. 1999), the Madden–Julian oscillation (MJO) (Mpeta and Jury 2001; Mutai and Ward 2000), and the quasi-biennial oscillation (QBO) (Indeje and Semazzi 2000; Ng’ongolo and Smyshlyaev 2010). ENSO alone is found to account for almost 50% of the East African rainfall variance (Ogallo 1988).
b. Data and choice of predictors
Predictand data used for the ESD model development are daily rainfall observation records for several stations in Tanzania (Fig. 1) obtained from Tanzania Meteorological Agency (TMA). The time period available for this dataset differs from station to station with some stations going back to the 1940s and few to the 1960s. However, for this work we used data from 1979 to 2012, which matches the MERRA data period. Within this period, 99.9% of the daily data are available for analysis. However, quality control of the data including inhomogeneity that could be caused by any change to station or its operation was not performed. Predictor variables used for correlation analysis and hence ESD model development are the reanalysis of monthly mean surface and atmospheric large-scale circulation fields that were taken from the NASA’s Modern-Era Retrospective Analysis for Research and Applications (MERRA), available online at http://disc.sci.gsfc.nasa.gov/daac-bin/DataHoldings.pl. The MERRA time period covers the modern era of remotely sensed data, from 1979 to 2012. A special focus of the atmospheric assimilation has been on the hydrological cycle. The reanalysis is available at a spatial resolution of 0.5° latitude × 0.67° longitude (Rienecker et al. 2011).
As predictor variables for downscaling future rainfall projections we have used GCM monthly mean rainfall simulations for the period 2006–2100. These simulations are obtained from phase 5 of the Coupled Model Intercomparison Project (CMIP5) climate change experiments (Taylor et al. 2012). This dataset is available online at http://pcmdi9.llnl.gov/esgf-web-fe/. Projection simulations forced with two representative concentration pathways, RCP8.5 and RCP4.5 (Moss et al. 2010), were used. These long-term integrations are initiated from multicentury preindustrial control integration (Taylor et al. 2012). RCP8.5 is consistent with a high emissions scenario and RCP4.5 is consistent with a midrange mitigation emissions scenario. The GCMs used in this study are listed in Table 1. The set of GCMs includes models with different spatial resolutions and degrees of complexity. Generally, the spatial resolution of CMIP5 GCMs roughly ranges from 0.5° to 4° for the atmosphere component and from 0.2° to 2° for the ocean component (Taylor et al. 2012). In our selection, we have included GCMs of variable resolutions from very high to low.
2) Predictor selection
Potential predictor variables were selected by referring previous works on ESD and empirical seasonal forecasting done in Africa. These are atmospheric large-scale fields such as 850-hPa height (gpm85) (Landman et al. 2012) and sea level pressure (SLP) (Huebener and Kerschgens 2007; Landman and Tennant 2000; Zhao et al. 2005). Other fields are large-scale precipitation (Pr) (Diro et al. 2012; Greene et al. 2012) and large-scale temperature (T2m) (Ambrosino et al. 2011; Greene et al. 2012). We also included horizontal components of wind (V10m and U10m) and specific humidity (shum10m) at 10-m level.
According to Kalnay et al. (1996) and Kistler et al. (2001), the reliability of the output variables from the reanalysis can be classified in categories A to D, depending on the relative influence of the observational data input and the model forecast on the variable during data assimilation. With the exception of SLP, which is classified as a class A variable, the rest of the selected variables (i.e., gpm85, V10m/U10m, T2m, and shum10m) are considered as class B variables and Pr as a class C variable. However, because of the scarcity of data in the tropics (where this study is undertaken) the relative influence of model forecast during data assimilation in all variables is considered to be large as compared to that of observational data input. For this reason, all variables were given equal weight in the predictor selection.
Spatial correlation analysis was performed between several predictors derived from the MERRA dataset and the predictand (station observations) to identify predictors with strong links to the observed precipitation. Monthly large-scale variables (SLP, Pr, gpm85, V10m/U10m, T2m, and shum10m) were tested for their significance in describing their relationship with the local monthly precipitation values in Tanzania. This was done by considering different domains, rainy seasons, and station locations. Pr and shum10m stand out as showing significant links with the stations’ observed rainfall. Since both parameters represent moisture sources, we decided to use precipitation fields with the advantages that 1) they represent the synoptic-scale processes in the atmosphere including changes in air humidity and movements of air (Linderson et al. 2004) and 2) they carry climate change signals due to anthropogenic activities (Benestad et al. 2008; Chen et al. 2006; Wilby et al. 1998) and they represent other processes not explained by the atmospheric circulation. Finally all downscaling models were developed using the same predictor (Pr) regardless of the season, type of statistics, and location of the station.
The ESD models used in this study are built in the following steps: First, the long-term mean for each dataset [i.e., MERRA (1979–2012) and GCMs (2006–2100)] is subtracted from that data on each grid point, and the two datasets are then combined along the time axis (concatenated) to form a single dataset covering the period 1979–2100 (see, e.g., Barnett 1999; Benestad 2001a). A bilinear spatial interpolation is then used to convert the GCM spatial grid to be in the same grid as that of MERRA. Second, a standard empirical orthogonal function (EOF) analysis (Lorenz 1956) is applied to the anomalies of the combined dataset with the common grid (which is now 0.5° × 0.67°). The eigenvectors (EOFs) produced from the EOF analysis represent patterns of variability that MERRA and GCMs data share in common. The EOF analysis applied to the combined dataset is referred to as common EOF analysis (Barnett 1999; Benestad 2001a) and its use ensures that the same large-scale spatial climatic patterns that are associated with the observed local climate anomalies are used to infer local climate changes related to the changes in the regional climate (Benestad 2002). Third, using a stepwise screening process based on the Akaike information criterion (AIC) (Akaike 1974), a multiple linear regression analyses between the predictand (i.e., observed precipitation) and the principal components (PCs) of the 20 leading EOFs of the combined dataset (MERRA + GCMs) is performed to decide the number of PCs to be included as predictors in the final models. During stepwise screening, a model that minimizes loss of information in simulating the current climate (i.e., with minimum AIC value) is retained for downscaling process. A leave-one-out cross-validation was then performed to assess the prediction skill of the models. This assessment was repeated for a test case where the station data had been replaced by random numbers for which there should be no skill. This assessment suggested that the models were not overfit for a number of stations and that the higher number of EOFs contained details related to some of the stations (see Fig. 2). The leading 20 EOFs are used because it allows more regional details in predictor fields to be used in the downscaling models especially for the MAM and DJFMA seasons that do not show a pronounced link with large-scale circulations. Before the model calibration, the best-fit linear trend is subtracted from each grid point in the observed predictor values (MERRA) and from the predictand (station observations), as the presence of a linear trend may introduce systematic biases to the model calibration (Benestad 2001b). However, the trends are included when the models make predictions. The downscaling models are calibrated with the part of the combined PCs that represents the actual observations (i.e., MERRA), and for future projections the downscaling is generated using part of the combined PCs that represents the GCMs simulations. More details of common EOF analysis and downscaling procedures can be found in Benestad (2007, 2001a, 2002).
d. ESD model implementation
Prior to statistical model development and hence downscaling, small-scale variability in the large-scale precipitation field was removed. A low-pass filter (boxcar smoother) was applied to the monthly large-scale precipitation field (reanalysis and GCMs). The width of the boxcar was 5 points (grid boxes) along the x and y axes and 3/5 points (months) along the time axis for the bimodal/unimodal climate regimes, respectively. The motive behind spatial smoothing is to reduce noise and capture the large-scale features. Temporal smoothing on the other side removes the short-term fluctuations hence emphasizes the long time fluctuations.
Predictors for model calibration are PCs of the 20 leading EOFs that represent the actual observation part of the combined dataset (i.e., MERRA). These 20 leading PCs undergo stepwise screening process in a regression method to select potential predictors that will be included in the final downscaling model. For future projections, predictors are PCs of the EOFs that were retained from the stepwise screening process but that represent the GCM simulations part of the combined dataset. Predictands are parameters that were calculated from the daily rainfall but aggregated in to seasonal basis for an individual local station and for each rainy season (i.e., MAM, OND, and DJFMA). The examined parameters are as follows: 1) frequency of wet-day occurrence (), henceforth wet-day frequency [this parameter explains how frequently it rains in a season; it is calculated as the fraction of number of wet days () to the total number of days () in a season and it is measured in fraction (of counts)]; 2) seasonal total rainfall (), henceforth total rainfall, the sum of rainfall amounts for all wet days in a season (measured in mm); and 3) wet-day mean (), which is calculated as the total rainfall () divided by total number of wet days () in a season; the wet-day mean is equivalent to daily rainfall intensity (measured in mm day−1). Wet days are described as number of data points with rainfall 1 mm day−1 while all days include days with valid recordings (i.e., rainfall 0).
Changes in these parameters were calculated by comparing future 30-yr climate periods (near term = 2010–39; midcentury = 2040–69; and end of century = 2070–99) to a baseline climate period (1979–2012), both extracted from the downscaled scenario time series. Each downscaled value was obtained from individual run using a GCM that is either forced with RCP4.5 or RCP8.5. The outputs from each run were combined together (tabulated) and further statistical analysis (e.g., mean, median, percentiles) was performed. The downscaling analysis and graphics processing are performed in the R environment (http://www.R-project.org/) by using the empirical statistical downscaling packages clim.pact (Benestad 2004) and ggplot2 (Wickham 2009). These packages are open source and free.
Since rainfall regimes across Tanzania vary significantly in space and time and since projections also vary widely between seasons and regions, results are presented for each season but for both emission scenarios combined for ease of discussion. Throughout the discussion the ESD models will be referred to simply as models and global climate models will be referred to as GCMs.
a. ESD models skill score using reanalysis data
The skill of the models discussed here is the coefficient of determination (abbreviated R2). This tells how much of the variance of the dependent variable (predictand) is explained by the regression line (expressed as percentage). The skill of the model may be influenced by several factors such as a weak relationship between predictor(s) and predictand, poor quality of the data (station observation and reanalysis), the design of the empirical model used, the size of the predictor domain, and many others.
In this study, as in previous statistical downscaling studies (Benestad 2002), the predictor domain size was found to influence the skill of the model. When a larger domain was used, the model skill was generally found to be lower than when a small domain was used. Only in a few cases the opposite was detected, for some stations and particular seasons. According to Benestad (2002) it is tricky to find the optimum downscaling domain, as the domain needs not to be too large or too small. As opposed to our finding, the study of Yang et al. (2013) found that different predictor domains had less influence on the projected monthly rainfall over Taiwan. For our downscaling we have chosen the domain 6°N–18°S, 15°–60°E.
The choice of number of the leading EOFs used in the regression analysis also showed a significant impact on the final downscaling models’ skill. When eight leading EOFs were passed through the stepwise-screening process in the regression analysis, the models’ skill was generally weaker than when 15 or 20 leading EOFs were used. This was found for all rainfall parameters and for all seasons. It is most noticeable in the DJFMA season and for values (Fig. 3). In fact, for one station (Dodoma) there is no model selected for when the leading eight EOFs are used, implying that there was no correlation between these eight spatial fields and the station precipitation. This could be because local rather than large-scale processes, not captured by the first eight leading EOFs, are dominating in this case. Based on the above considerations and the cross-validated test results we decided to use 20 leading EOFs as an initial pool to be used by a stepwise screening.
b. The overall skill score of the models under the common EOF strategy
The algorithms in our EOF analyses are constrained to patterns (EOFs) that are common in both sets of data and their corresponding PCs that are used for model calibration. The skills of the models differ with respect to the particular GCM used (Fig. 4) as each GCM uses specific assumptions and formulations of processes in addition to having different resolution. As a result, GCMs may yield different responses to the same external forcings (anthropogenic and natural) and also represent internal climate variability differently. Generally, the highest skill score (between 70% and 95%) is shown for and for OND; exceptions are Dar es Salaam (Dsalaam in Fig. 4) and Mwanza () whose skill ranges between 50% and 75%. Skill of 50% and up to 90% is depicted for the MAM and DJFMA seasons for and . In particular Same and Kigoma have lower skill for , and Mwanza has lower skill for . Wet-day mean () showed a wide range of skills from very low (25%) to very high (~80%) for all seasons for specific locations.
The high skill obtained in the OND season indicates that the interannual variance for and in the calibration period (1979–2012) is well explained by the fitted models (see, e.g., Fig. 5). The model (fit in the figure) has well reproduced the observed total rainfall ( = 85%) and the wet-day frequency ( = 94%). Year-to-year variability and the maximum values recorded are well represented in the model. Analysis of residuals of regression for the period 1979–2012 (not shown) that were computed using least squares regression analysis depicted a slight linear trend (although not statistically significant) for OND. This was found for all stations and all GCMs used for, but only for three stations (Arusha, Bukoba, and Mwanza) and for some GCMs for . In accordance with Hanssen-Bauer et al. (2003), such a trend may be due to the fact that models do not satisfactorily reproduce the observed fluctuations in these rainfall parameters.
In some stations, seasons and rainfall parameters models have shown to have very low prediction skills (see, e.g., Fig. 2), but still this cannot be associated with the model overfit; rather, other factors such as inability of MERRA and GCMs to simulate large-scale features are responsible. In this regard, we can generally say that the goodness of the model (skill) and its prediction skill as shown in Fig. 2 is specific to location, season, and to some extent parameter. However, even with the high skill of the downscaling models identified here, we are left with uncertainty in the downscaling of the future due to uncertainties in the GCM calculated precipitation.
c. Downscaled scenarios
Results presented here are downscaled multimodel ensemble scenarios for specific locations and seasons. The multimodel mean and the spread in projections from individual models are presented. The reason for this approach is that large wetting or drying may occur to some locations even if the multimodel mean has lost that signal. Moreover, the multimodel mean alone may be misleading especially when used for decision making (e.g., for adaptation purposes where the range of outcomes needs to be taken into account). The multimodel ensemble mean projection for individual stations and seasons in this study is still considered as the most likely outcome. Results are given in Figs. 6–8 for the three rainfall parameters for individual stations, seasons, and periods, including the variability between models (whisker plots). Their corresponding magnitudes of change are expressed in percentages. Figure 9 gives a picture of the spatial distribution of the multimodel mean projections in all three parameters across the country.
1) Projection for 2010–39
The spread in projection is depicted for all seasons and for all stations (Fig. 6) as depicted in the box-and-whisker plots. The largest spread in projection is depicted in OND (notice the difference in scales between the seasons in the figure) where individual models project more than 100% increase, namely Dar es Salam, Tanga, and Morogoro. Dar es Salaam is the only location with wetter condition as depicted by all models. This is accompanied by an increase in . The multimodel mean projection shows wetter conditions () for Dar es Salaam, Same, Morogoro, Arusha, Musoma, and Bukoba during OND and they are accompanied by an increase in only (Musoma and Arusha), only (Bukoba), or both (Same, Morogoro, and Dar es Salaam)
Ensemble projections of are not uniform for MAM even for stations situated in the same climatic zone (e.g., the northeastern highland, represented by Arusha and Same). Almost all models project a decreasing for Arusha, with a robust signal of drier conditions which is backed up by the projected decrease in . A majority of models indicate increasing rainfall for Same, Dar es Salaam, and Morogoro.
In the one rainy season (DJFMA) region there are many stations that are projected to get lower total precipitation in the ensemble mean. For most stations the wet-day mean increases, resulting in lower wet-day frequency at all stations. About 50 percentile of models indicate increasing rainfall () for Mtwara and Tabora.
2) Projection for 2040–69
The change for this period (Fig. 7) also shows a large scatter in projected , again in particular during OND. A majority of models point to wetter conditions as does the multimodel mean projection for all stations, both in terms of total precipitation and frequency, but still with an increase in wet-day mean (except for Arusha). In extreme cases, individual models are suggesting up to more than 100% increase of total rainfall for, for example, Dar es Salaam and Morogoro, for the near-term projection.
In the MAM projections there is less, but still substantial, spread in . A majority of models point toward drier conditions for most stations (Arusha, Bukoba, Mwanza, Musoma, and Tanga), accompanied by decrease of both and . Otherwise increase in is depicted for Dar es Salaam, Morogoro, and Same, and it is accompanied by an increase in . As for near-term projections, a robust signal toward lower total precipitation () is indicated for Arusha, along with decreasing .
A majority of models show drier conditions () for DJFMA for most stations, but the signal of change is weak as compared to other seasons (between 0% and 10%). This change is directly supported by the decrease in . However, about half of the models project wetter () conditions for Tabora and Mtwara accompanied by increasing .
3) Projection for 2070–99
The change for this period is shown in (Fig. 8). Again, the largest changes are found during OND. In this period there is a clear and robust signal toward wetter conditions () for most stations, except Morogoro and Tanga where about 50 percentile is below zero. In the ensemble mean, the increase in the total precipitation in this period is due to an increase in both and . Although there is large scatter in the model projections for MAM, Dar es Salaam and Morogoro and to some extent Same show a clear signal of increasing total rainfall, which is backed up by the increase of (Dar es Salaam and Morogoro) and (Same) . While the projection of over two stations (Mwanza and Bukoba) in the lake zone seems to be uncertain (i.e., boxes cross the zero line) the other station (Musoma) shows a clear signal toward a dry condition, mostly due to lower rainfall intensity ().
For DJFMA, a clear signal of decreasing (represented by all models) is depicted at Songea, a station in the southern part of the country. This is due to lower frequency () and intensity (). A majority of models show drier conditions () for almost all stations, except the western part of the country (Kigoma and Tabora) and Mtwara (south) where the model projections spread through the zero line. However, the signal of change in the multimodel-mean projection is weak as compared to other seasons (between 0% and 20%). The projected dry condition () is attributed to a decrease in in all stations and accompanied by either increase or decrease in at individual stations.
For all three periods, a significance test for projected changes in rainfall parameters was performed. It showed that some downscaling models provided a significant change (at 5% significant level; t test) especially for the periods 2040–69 and 2070–99 while some did not show any significance. It varied between the GCMs used, RCPs, stations, and downscaled parameters. Therefore, the multimodel projection presented here represents a combination of changes that are statistically significant and those that are not.
We have found spatial variability in model skill for all seasons and rainfall parameters investigated. The variability is likely due to the fact that different areas have different local factors that modulate the rainfall patterns. The skill scores are in general quite high, but weak skill scores found in some cases may suggest that local processes play an important role. Ideally we would expect the skill of the models in a particular station to be equal for all GCMs used, since models are developed by using observations (reanalysis vs station observations), which are the same for all common EOF setups. However, the difference and spread depicted is probably a reason that the AIC selected variable numbers of predictors and different common EOFs for individual models whose corresponding PCs represent different fluctuations. For demonstration, Table 2 shows the coefficients of EOFs selected for two pairs of neighboring stations situated in different homogeneous climatological regions, the northeastern highlands (Arusha and Same) and northern coast (Tanga and Dar es Salaam). We show results for these stations as they showed a remarkable difference in model skill scores and projections for MAM despite being in the same homogeneous climatological zone. From Table 2 we can see that a few EOFs are selected by AIC that were common for neighboring stations, but the rest are independent (i.e., they vary between the stations). Even with similar EOFs selected, the coefficient values that describe the contribution of each EOF in explaining variability in the local precipitation are also different in magnitude and sign. Another feature depicted in the table is that some models are developed with a large number EOFs and some with a minimal number.
The number and type of EOFs picked by AIC in the regression analysis for a particular station may partly explain the differences in the depicted variable models skill scores, and the weights (regression models coefficients) may explain the remaining part of the variable skill together with the spread in GCM projections. The analysis of the time evolution (PCs) of the spatial modes (EOFs) (Fig. 10) used in the regression demonstrated that 1) the PCs of individual EOFs differs depending on the GCM used (see Fig. 10, first and third panels in both columns) and 2) models consist of PCs that do not share a common trend (see Fig. 10, second and fourth panels in both columns). The PC part of the GCMs’ projections underestimates the amplitude and the periodicity of the time evolution of these spatial modes (especially for lower-order EOFs) as compared to the observed period (Fig. 10, first and third panels in the left column). However, the PCs for higher-order EOF (Fig. 10, first and third panels in the right column) are more like noisy fields and their evolution indicates comparable amplitude and the periodicity between the observed (reanalysis) and the GCM part. These noticeable behaviors in the PCs suggest that the spread in projections based on individual GCMs will even change between various projection periods.
We can therefore conclude that the spread and differences in projections can be explained by three factors that arose from the downscaling processes: 1) the number and type of EOFs picked by AIC in the regression analysis, 2) the weights of the regression coefficients, and 3) the trends in the individual PCs. In addition to these factors, the climate system noise and the uncertainty in emission scenarios are expected to cascade to the regional climate projection even with perfect GCMs used for downscaling (Mitchell and Hulme 1999). However, there is no clear evidence that the differences/spread in skill score can explain the differences/spread in projections. As an example, the OND season, which showed a remarkable spread in projections in all three periods, for in particular, has very small differences in skill scores, but the opposite is true for DJFMA.
Further analysis of GCMs’ area mean precipitation (Fig. 11 and Table 3) shows that there is no clear evidence that difference in GCMs simulations of rainfall over the downscaling domain can cause a noticeable spread in the rainfall projection, a spread that is especially large for OND for some stations. Table 3 can also be used to explain why there is no remarkable difference in projected changes between the three periods analyzed; the annual mean rainfall simulated by individual GCM over the downscaling domain is nearly the same for different periods of projection.
Regional climates are very complex due to processes that vary significantly with location and therefore respond differently to changes in global- to regional-scale influences (Christensen et al. 2013). Small-scale physical features within the country are expected to enforce a complex pattern of projected rainfall in different climatological zones even without the interaction with large-scale phenomena. Hilly areas with steep slopes, for example, enhance observed mean rainfall, especially during wet seasons (Oettli and Camberlin 2005). In Lake Victoria basin, on the other hand, the climate (rain) is induced by the lake/land breeze and the surrounding topography, which results in uneven distribution of rain over this area, with the western part of the basin receiving more rainfall than the eastern part (Anyah et al. 2006). But with the increased influx from the large-scale flow, a different story unfolds, especially for the eastern part. This phenomenon is consistent with the results from the current study during the period 2070–99 (MAM and OND seasons). There is a remarkable difference between projection of for Musoma and Mwanza (stations situated to the east and south of the lake respectively) as compared to a counterpart station in the western part of the Lake Victoria basin (Bukoba).
Anomalous low-level westerly flow (OND) and easterly flow (MAM) are projected at the end of twenty-first century (2041–60) (Cook and Vizy 2013). This suggests that even with strong influx of moisture from the Indian Ocean to the eastern part of the country (MAM), more moisture is expected to be transported to the Congo basin and leave the country dry. Moreover, more moisture influx from the Congo basin is expected during OND, when enhanced moisture convergence is expected over the country and hence more rain. Thus, changes in wind flows may partly explain projected wetter conditions for the OND season and decreased total rainfall for a majority of stations receiving rains during MAM as depicted by the majority of models in the current study, and in the period 2040–69 in particular.
A positive phase of the Indian Ocean zonal mode (IOZM) that is associated with excessive OND rains over East Africa is projected by the CMIP5 GCMs (Shongwe et al. 2011). This condition is consistent with the projected OND wetter conditions identified in this study. Global warming, on the other hand, has been directly linked to recent changes in mean rainfall (Giorgi et al. 2011) in relation to either change in or (intensity) (Benestad 2013). Under global warming, the increase in precipitation intensity is driven by the greater atmospheric moisture holding capacity and the increase (decrease) in mean precipitation is driven by a correspondingly higher (lower) increase in surface evaporation (Giorgi et al. 2011). The response of these rainfall parameters to the global warming explains the thermodynamic effects of the Earth–atmosphere interactions whose impact is considered to be global and to a lesser extent regional.
However, we should keep in mind that not all climatic zones will experience similar changes and thus the current study yields large spatial variability in the projected changes in the rainfall parameters. Thus there is no uniform projection for areas characterized by complex topographical features such as that of Tanzania (see also Taye et al. 2011).
This study compliments previous studies in climate analysis and projection in the region of study making use of different approaches, such as different downscaling methods (statistical or dynamical), predictors, observations (parameters and observation periods), time scale in analysis (e.g., seasonal or annual), and driving GCMs. For example, Akurut et al. (2014) found an increasing annual rainfall over the Lake Victoria basin for middle and end of this century, and these changes are attributed to an increase in and not . This is not the case as depicted in this study for different areas in the basin on a seasonal basis. In Musoma as an example, increase in for OND is attributed to increase in and not . In the same area, rainfall is projected to increase by up to 30% for some areas in OND whereas it is expected to decrease by up to 5% in MAM. The study of Shongwe et al. (2011) over East Africa found projected wetter conditions with more intense wet seasons (OND and MAM) and less severe droughts for the period of 2050–2200. Similar results are depicted in the current study; projected wetter conditions () are predicted due to both an increase in precipitation intensity () and an increased frequency of rainy days () at the end of the century (2070–99), for all stations receiving OND seasonal rains. For a station-specific comparison, the results of the current work are in agreement with that of Tumbo et al. (2010) for the Same station during MAM and OND seasons and the periods 2040–69 and 2070–99.
5. Summary and conclusions
We have investigated possible changes in the seasonal , , and using regression-based ESD. The analysis is performed for the three periods 2010–39, 2040–69, and 2070–99. Apart from the inherited uncertainty from the GCMs used in the downscaling processes, the design of the empirical models used in this study can be used to explain the differences/spread in models skills and projections in rainfall parameters. With the sample of GCMs used in this study, our analysis of the downscaled climate scenarios suggests that precipitation changes will vary across Tanzania. Spatial differences in projected climate will be experienced as it is portrayed by the observed climates (Camberlin and Philippon 2002).
Surface inhomogeneity in East Africa and Tanzania in particular makes the spatial structures of projected rainfall even more complex, especially in MAM. There is no uniform sign of change depicted by multimodel projection across different areas; therefore, drawing a conclusion of a uniform change for a region with heterogeneous surfaces is not possible. This demands the provision of location-specific climate analysis. In the near-term projection, Dar es Salaam has shown a unique wetter condition () for OND as depicted by all models. In end-of-century Morogoro (MAM), Bukoba and Same (OND), and Songea (DJFMA) also indicated a robust signal toward wetter (MAM and OND) and drier (DJFMA) conditions. In the remaining cases the likely future condition is determined by the majority of models pointing toward either wetter or drier conditions, although a few models point toward the opposite condition or do not show changes at all. The spread in projections was found to be larger in 2010–39 (e.g., in DJFMA and OND) and reduced slightly in 2040–69 and 2070–99, and it was nearly constant for MAM. A large spread in the period 2010–39 can be associated with the internal fluctuations in the climate system, which are expected to be more pronounced for the near-term projections than the effects of external forcing such as greenhouse gas increases (Collins et al. 2013).
In this study, there is no uniform conclusion regarding the magnitude of change in projected rainfall parameters (see, e.g., Fig. 9). The projected increase in total rainfall in OND, for example, is due to the increasing frequency of wet days () (in most stations and for all projection periods) and more intense precipitation events () (for some stations). For DJFMA, projected decreasing total precipitation is due to decreasing frequency of wet days (all stations and for all projection periods) and less intense precipitation events (only for some stations and for the periods 2040–69 and 2070–99 in particular). Although the western part of the country (Kigoma, Tabora, and Mbeya) is projected to be drier (), the multimodel ensemble indicates an increase in intense precipitation events. For MAM, with the exception of Dar es Salaam, Morogoro, and Bukoba, the increase (decrease) in total precipitation is accompanied by increasing (decreasing) intense precipitation events (). In the exceptions mentioned the increase (decrease) in total precipitation is accompanied by increasing (decreasing) frequency of wet days (). The magnitude of change also differs from area to area. As an example, in the northeastern highlands Same shows an increase of about 50% of OND precipitation toward the end of the century, while Arusha shows only an increase of about 20%. Likewise, Same has an increase of about 1% of MAM precipitation toward the end of the century, while Arusha shows a decrease of about 15%.
Increased total rainfall () as a result of increased frequency of wet days () alone can be important in maintaining soil moisture and hence good conditions for crops and pastures. From our results this condition is expected to be featured for some areas getting MAM rains (e.g., Dar es Salaam and Morogoro). When increased is a result of increasing or both and , growing conditions might be worsened due to short episodes of heavy rainfalls and related floods. Such extreme events are expected to cause a significant challenge to the agriculture sector, the economy, and the food security at the local level (United Republic of Tanzania 2014) and to other sectors sensitive to wet and dry extremes such as water resource management and construction industries (e.g., roads, dams for hydroelectric power). From our results this condition is expected to be featured largely for areas getting OND rains especially in the end-of-century time frame.
Decreasing total rainfall () as a result of decreasing frequency of wet days (), either alone or in combination with the wet-day mean (), implies a higher risk of long dry spells or drought. From our results such conditions are expected to be featured in the areas spanning one rainy season (DJFMA) and in some areas getting MAM rains (e.g., Bukoba and Tanga for all projection periods and Musoma and Mwanza in the near term and at midcentury). However, a few areas getting DJFMA season rains (e.g., Kigoma, Tabora, and Mbeya) may experience some episodes of heavy and damaging rains and as a result of projected increasing heavy rainfall events () during this season. The related impacts may include damage to crops, which in turn will affect the growth and harvests. This consequence may cause famine outbreaks and hence threaten the livelihoods of many.
This work was done under the support of the Climate Change Impacts Adaptation and Mitigation (CCIAM) project funded through the Norwegian Embassy in Tanzania. The authors wish to acknowledge the Tanzania Meteorological Agency (TMA) for providing rainfall data, the Global Modeling and Assimilation Office (GMAO), and the GES DISC for the dissemination of MERRA data. We also wish to acknowledge the World Climate Research Programme’s Working Group on Coupled Modelling, which is responsible for CMIP5, and we thank the climate modeling groups (listed in Table 1 of this paper) for producing and making available their model output. The authors also acknowledge use of the Ferret program for data processing in this paper. Ferret is a product of NOAA’s Pacific Marine Environmental Laboratory. Information is available at http://ferret.pmel.noaa.gov/Ferret/.