Although the African Great Lakes are important regulators for the East African climate, their influence on atmospheric dynamics and the regional hydrological cycle remains poorly understood. This study aims to assess this impact by comparing a regional climate model simulation that resolves individual lakes and explicitly computes lake temperatures to a simulation without lakes. The Consortium for Small-Scale Modelling model in climate mode (COSMO-CLM) coupled to the Freshwater Lake model (FLake) and Community Land Model (CLM) is used to dynamically downscale a simulation from the African Coordinated Regional Downscaling Experiment (CORDEX-Africa) to 7-km grid spacing for the period of 1999–2008. Evaluation of the model reveals good performance compared to both in situ and satellite observations, especially for spatiotemporal variability of lake surface temperatures (0.68-K bias), and precipitation (−116 mm yr−1 or 8% bias). Model integrations indicate that the four major African Great Lakes almost double the annual precipitation amounts over their surface but hardly exert any influence on precipitation beyond their shores. Except for Lake Kivu, the largest lakes also cool the annual near-surface air by −0.6 to −0.9 K on average, this time with pronounced downwind influence. The lake-induced cooling happens during daytime, when the lakes absorb incoming solar radiation and inhibit upward turbulent heat transport. At night, when this heat is released, the lakes warm the near-surface air. Furthermore, Lake Victoria has a profound influence on atmospheric dynamics and stability, as it induces circular airflow with over-lake convective inhibition during daytime and the reversed pattern at night. Overall, this study shows the added value of resolving individual lakes and realistically representing lake surface temperatures for climate studies in this region.
The African Great Lakes (AGL) represent the largest reservoir of freshwater lakes in the tropics, including the second largest freshwater lake on earth in terms of surface area (Lake Victoria, 68 800 km2) and the second largest in terms of volumetric water storage (Lake Tanganyika, 17 800 km3). The AGL provide numerous ecosystem services to local communities, such as fishing resources, drinking water, and electrical power. Lake Victoria alone directly supports 200 000 fishermen operating from its shores and sustains the livelihood of more than 30 million people living at its coasts (Semazzi 2011).
Lakes also influence regional climate conditions, in particular in regions where they are abundant. Mediated by the fluxes of energy, moisture, and momentum, they significantly alter the surface energy and water balance in a particular region and therefore its climate. Such reciprocal relations are often studied using numerical models. At the global scale, general circulation model (GCM) studies demonstrate an average annual cooling effect of lakes but also highlight important seasonal and regional effects (Bonan 1995; Subin et al. 2012; Rooney and Bornemann 2013). Numerous studies also mention the increase in annual precipitation due to the presence of lakes (e.g., Coe and Bonan 1997). Regional climate model (RCM) studies have been performed at higher horizontal resolution (e.g., 50 km), thereby better emphasizing the lake–atmosphere interplay at the local scale (e.g., Hostetler et al. 1994; Hostetler and Giorgi 1995; Bates et al. 1995; Goyette et al. 2000; Anyah et al. 2006; Samuelsson et al. 2010; Gula and Peltier 2012; Lauwaet et al. 2012; Martynov et al. 2012; Gu et al. 2015; Notaro et al. 2013; Bennington et al. 2014; Williams et al. 2015).
Although the AGL are widely recognized as an important driver of the East African regional climate system, their impact on atmospheric dynamics and the regional hydrological cycle remains poorly understood. Studies on the interplay between climate and the AGL are therefore necessary to identify, among other information, the patterns of evaporation and precipitation induced by lake presence and the reasons behind their observed spatiotemporal variability. The need to enhance the understanding is urgent, as every year, possibly more than 5000 fishermen lose their lives over Lake Victoria alone, mostly due to hazardous weather conditions and associated water currents (Semazzi 2011).
The poor understanding of African lake–climate interactions is related to a number of challenges when applying a regional climate model to the AGL region: (i) First, a model should be applicable to tropical conditions. Among other requirements, such a setup must be able to reproduce tropical features, such as deep convection or mesoscale convective systems [and associated precipitation extremes; Lauwaet et al. (2009); Goyens et al. (2012)] and must include a land surface model (LSM) capable of simulating tropical vegetation. In recent climate model studies over tropical Africa, these challenges have received particular attention (e.g., Vizy and Cook 2012; Saeed et al. 2013; Crétat et al. 2014; Panitz et al. 2014; Dosio et al. 2015). (ii) Second, the horizontal model grid resolution should be sufficiently high to resolve individual lakes and strong orography of the region. Combined with the long integration period and high vertical extent necessary to model tropical climatic conditions, the need for computational resources rises. (iii) Third, a correct representation of lake surface temperatures within the atmospheric model is essential to investigate the two-way interactions between climate and lake processes over East Africa (Stepanenko et al. 2013). This can be particularly challenging for large, deep lakes given the role of vertical heat transport and three-dimensional water circulation within these lakes (Gu et al. 2015; Bennington et al. 2014). (iv) Last, the limited number of observational datasets available and issues regarding the quality of available products over highly mountainous, tropical domains (e.g., Dinku et al. 2008) complicate a comprehensive model evaluation over this region. Consequently, up to now only a few RCM studies explicitly examine aspects of lake–atmosphere interactions over a specific African Great Lake (Savijärvi and Järvenoja 2000; Song et al. 2004; Anyah and Semazzi 2004; Anyah et al. 2006; Argent et al. 2015; Sun et al. 2014a,b; Williams et al. 2015), while none investigates these interactions for the whole region.
Hence, the main goals of this study are to assess the regional climate impact of the AGL and understand the physical mechanisms underlying these impacts. This is achieved by addressing the aforementioned challenges in the following ways: (i) The regional Consortium for Small-Scale Modelling model in climate mode (COSMO-CLM; Rockel et al. 2008) is applied in its tropical setup [configuration following Panitz et al. (2014)]. In their comprehensive evaluation of the COSMO-CLM evaluation simulation conducted in the framework of the African Coordinated Regional Downscaling Experiment (CORDEX-Africa; Giorgi et al. 2009), Panitz et al. (2014) demonstrate the ability of COSMO-CLM to reproduce the overall features of the African climate (when downscaling ERA-Interim) and show that, relative to several gridded observational products, it even outperforms ERA-Interim in terms of the mean annual precipitation cycles and interannual variability over major parts of the domain. Among the considered regional biases, the authors point to an underestimation of the precipitation intensity by 2 mm day−1 in East Africa, which is ascribed to an underestimation of the latent heat flux and overestimation of the sensible heat flux. To address this issue, here we employ COSMO-CLM coupled to a state-of-the-art LSM, the Community Land Model, version 3.5 (CLM3.5; Davin and Seneviratne 2012). For various applications, this version already showed improved skills compared to the standard version (see section 2a). (ii) Model integrations are performed for a period of 13 years on a horizontal resolution high enough for large to medium-sized lakes to become resolved features (0.0625°; ~7 km). (iii) The one-dimensional Freshwater Lake model (FLake), recently coupled to COSMO-CLM (Mironov et al. 2010), is used to compute lake water surface temperatures. (iv) Special attention is paid to model evaluation, in particular by using remote sensing products and comprehensive in situ observational datasets to assess the model’s ability to reproduce observed spatial and temporal lake surface temperature variability, turbulent fluxes, and precipitation. Adopting one or more of the above challenges may improve the representation of the regional climate, as demonstrated in many recent studies (e.g., Mironov et al. 2010; Nikulin et al. 2012; Akkermans 2013; Laprise et al. 2013; Akkermans et al. 2014; Crétat et al. 2014; Dosio et al. 2015; Williams et al. 2015).
Within this paper, the results of the RCM control simulation are extensively evaluated. This simulation is subsequently compared to a no-lake simulation, where the lakes are replaced by representative land pixels. This approach makes it possible to quantify the impact on the regional climate, in particular on near-surface temperature and precipitation. The impact of the lakes is then investigated in more detail by applying a surface energy balance decomposition technique and studying the dynamical changes induced by the lakes.
2. Model, data, and methodology
To investigate the influence of the AGL on the climate, we use the three-dimensional, nonhydrostatic regional climate model, COSMO-CLM (Rockel et al. 2008), version 4.8. The model development is a joint effort of the Consortium for Small-Scale Modelling and the Climate Limited-Area Modelling Community (CLM-Community). Its dynamical core solves the fully compressible hydrothermodynamical equations for three-dimensional Cartesian wind components; pressure perturbation; temperature; specific humidity; cloud water content and optionally cloud ice content; turbulent kinetic energy; and rain, snow, and graupel specific water content (Doms 2011). Physical parameterizations include a description of subgrid-scale turbulence (Raschendorfer 2001), grid-scale clouds and precipitation (Doms et al. 2011), shallow and moist convection (Tiedtke 1989), and -two-stream radiative transfer (Ritter and Geleyn 1992). A detailed description of the model system dynamics, numeric, and physical parameterizations can be found in the model documentation (e.g., Doms 2011; Doms et al. 2011; http://www.cosmo-model.org).
Additionally, the native COSMO-CLM Multi-Layered land surface model (TERRA-ML; Grasselt et al. 2008) is replaced by the more comprehensive Community Land Model (CLM; Oleson et al. 2004, 2008), version 3.5. CLM3.5 is maintained at the National Center for Atmospheric Research (NCAR), and given its explicit treatment of the photosynthetic process and its control on the stomatal conductance (big-leaf approach), it can be considered a third-generation LSM (Sellers et al. 1997). Moreover, an individual land pixel can be composed of multiple land units due to the nested tile approach, which allows representation of multiple soil columns and represents biomes as a combination of different plant functional types. An extensive description and evaluation of CLM3.5 in offline mode is presented in Oleson et al. (2004, 2008) and Stöckli et al. (2008).
The model system COSMO-CLM coupled to CLM3.5 (hereinafter referred to as COSMO-CLM2) is described in detail by Davin and Seneviratne (2012) and references therein. The added value of COSMO-CLM2, relative to the standard setup, was recently demonstrated over Europe (Davin et al. 2011; Davin and Seneviratne 2012; Lorenz et al. 2012) and central Africa (Akkermans 2013). A better partitioning of the turbulent fluxes and associated improvement of climate characteristics, such as near-surface temperature, precipitation, and cloud cover, were found in all cases, supporting the usage of this model for studying biogeophysical impacts of altered land surface characteristics (e.g., Davin et al. 2014; Lejeune et al. 2015; Jacobs et al. 2015; Vanden Broucke et al. 2015). Based on an offline comparison of both LSMs at four central African flux tower sites, Akkermans et al. (2012) attributed this enhanced skill mainly to a more realistic representation of the leaf area index, surface albedo, and root depth in CLM3.5 compared to TERRA-ML. Furthermore, Akkermans et al. (2014) extensively evaluated COSMO-CLM2 over the larger central African domain and found a close correspondence between the observed and modeled near-surface temperature, column precipitable water, and surface net longwave radiation. For precipitation, cloud cover, top of the atmosphere outgoing longwave radiation, and various solar radiation quantities, the model adequately reproduced observed values, with some notable differences potentially ascribed to an underestimation of the optical thickness of convective clouds by COSMO-CLM2.
As a one-dimensional lake model embedded in COSMO-CLM2, FLake computes the evolution of a lake column temperature profile and the integral energy budgets of its different layers (Mironov et al. 2010). The model considers two layers in the water column: a mixed layer at the top and a thermocline down to the lake bottom. While the mixed layer is assumed to have a uniform temperature , the temperature-depth curve in the thermocline is parameterized through the concept of self-similarity or assumed shape (Kitaigorodskii and Miropolskii 1970). This shape is only dependent on the temperature at its extremities and a shape factor, which describes the curve through a fourth-order polynomial (Mironov 2008). In the absence of snow and ice cover and when neglecting bottom sediments, the prognostic variables are the mixed layer depth , the bottom and water column average temperatures ( and , respectively), and the shape factor with respect to the temperature profile in the thermocline . Given their control on turbulence in the surface mixed layer (Read et al. 2012), is predicted taking convective and mechanical mixing into account, whereas volumetric heating is computed from the net shortwave radiation penetration through the lake surface and absorption at depth according to the Beer–Lambert law. To this end, the default downward light attenuation coefficient of 3 m−1 is employed. A recent global bathymetry dataset (Kourzeneva 2010; Kourzeneva et al. 2012) is used to specify lake depth of each lake pixel, thereby assuming a maximum lake depth of 50 m. A detailed description of FLake is provided in Mironov (2008) and Mironov et al. (2010).
Although FLake is a relatively simple model, its skill proves comparable to other, more comprehensive one-dimensional lake models for various lake types and climatic conditions (e.g., Perroud and Goyette 2009; Stepanenko et al. 2010, 2013, 2014). Moreover, FLake was extensively tested in offline mode over several AGL and judged an adequate tool to compute lake surface temperatures in comparison to other one-dimensional lake models, even though lake bottom temperatures may be sensitive to changes in certain external parameters and driving variables, such as wind speed (Thiery et al. 2014a,b). This is the first study to consider lake–atmosphere interactions over the AGL using FLake interactively coupled to an RCM. Previous RCM studies focusing on Lake Victoria either used prescribed lake surface temperatures (e.g., Anyah and Semazzi 2004; Argent et al. 2015; Sun et al. 2014a; Williams et al. 2015), the Hostetler lake model (e.g., Song et al. 2004), or the Princeton Ocean Model (POM; e.g., Song et al. 2004; Anyah and Semazzi 2006; Sun et al. 2014b).
c. Model setup
To obtain information on small-scale variability, RCMs dynamically downscale lateral boundary conditions at a coarser resolution to a regional finer-scale model grid (Laprise et al. 2008). Here, we use the COSMO-CLM CORDEX-Africa evaluation simulation (Panitz et al. 2014) as lateral boundary conditions. The CORDEX-Africa evaluation simulation represents a dynamical downscaling to a 50-km (0.44°) horizontal grid spacing of the ERA-Interim, made available by the European Center for Medium-Range Weather Forecast (ECMWF; Dee et al. 2011) for 1979–2014 at 79-km (T255) resolution. Given the large difference in resolution between the latter and the target configuration, a direct downscaling of ERA-Interim is not advisable in this case.
The COSMO-CLM2 model system is used to generate two climate simulations. First, a control simulation (CTL) is conducted where the lakes are included in the model integration. For this purpose, FLake is employed to compute the water temperatures for the AGL. The second simulation [the no-lake simulation (NOL)] is identical to the control experiment, but each lake pixel is replaced by a representative land pixel selected from all land pixels within a distance of 50 km. The external parameters and initial conditions of replaced land pixels consequently show similar values and spatial patterns as the surrounding land; for instance, the area change of various plant functional types (PFTs) within the domain is 0.35% on average and less than 3% for each individual PFT.
Both experiments are conducted at a horizontal resolution of 0.0625° (~7 km), which is finer than previous multiyear RCM experiments over this region (e.g., Sun et al. 1999; Song et al. 2004; Anyah and Semazzi 2004; Anyah et al. 2006; Williams et al. 2015; with resolutions ranging from 20 to 60 km), using 50 vertical levels and a time step of 60 s. The model domain encompasses the central part of the East African rift (Fig. 1) and, therewith, includes most of the AGL. Since enlarging the model domain by 50 grid points (3.125°) in each direction reduced the predictive skill of a 1-yr test simulation and more than doubled the total computation time, this option was not considered. The simulations cover the period 1996–2008 (13 yr), of which the first three years are considered as spin-up time and excluded from the analysis. Inspection of predicted lake water surface temperatures shows that three years is enough to spin up the model: lake water surface temperatures evolve to normal climatological values within the first year already. Finally, on each side, 10 grid points were excluded from the analysis. Computation of the Davies relaxation function (Davies 1983) in the east–west direction showed that direct influence of lateral boundary conditions vanishes beyond this zone. The presence of two north–south oriented mountain ranges within the model domain are expected to further reduce the imprint of the parent model on the COSMO-CLM2 predictions.
The configuration used in this study is based on the tropical setup of COSMO-CLM described by Panitz et al. (2014) and used for the CORDEX-Africa simulations (Panitz et al. 2014; Dosio et al. 2015). In the tropical configuration, the model top is set to 30 km, and the lower height of the damping layer is increased from 11 to 18 km. In addition, we introduced the following modifications: (i) First, CLM3.5 replaces the default LSM in COSMO-CLM. (ii) Second, FLake is used instead of the Hostetler-based lake model embedded within CLM3.5 (Bonan et al. 2002), as the latter exhibits a cold lake temperature bias over the AGL (Thiery et al. 2014b). As recommended by Thiery et al. (2014a), the lake temperature profile is initialized with stratified conditions, an approach also adopted by Hernández-Díaz et al. (2013). (iii) Last, certain model settings (e.g., time step, number of vertical layers, and width of the relaxation layer) were adapted to be consistent with the higher horizontal grid resolution.
d. Evaluation datasets
The performance of the CTL simulation is assessed by comparing model output to observational products for various quantities (see Table 1 for their technical specifications). For precipitation, we use gridded datasets from the following sources: the Global Precipitation Climatology Project (GPCP; Huffman et al. 2001), the Global Precipitation Climatology Centre (GPCC; Rudolf et al. 2011), the University of Delaware (UDEL; Legates and Willmott 1990), the Climate Prediction Center Morphing Technique (CMORPH; Joyce et al. 2004), and satellite-derived precipitation rates from the Tropical Rainfall Measurement Mission (TRMM) products 3B42 and 2B31 (Kummerow et al. 2000). The TRMM 2B31 product is a combined rainfall-profile product derived from the Precipitation Radar and TRMM Microwave Imager (Bookhagen and Strecker 2008). Since the standard TRMM precipitation product (3B42) and CMORPH demonstrate poor performance over complex African terrain (Dinku et al. 2008), the TRMM 2B31 product is preferred over 3B42, given the higher spatial resolution and ability to reproduce orographically induced precipitation patterns over complex terrain (Bookhagen and Burbank 2006; Bookhagen and Strecker 2008; Burbank et al. 2012). The high-spatial-resolution TRMM 2B31 1999–2008 data were reprojected from their original 4.3–5-km resolution to the 0.0625° grid using a bilinear interpolation algorithm to account for projection and resolution inhomogeneities (Bookhagen and Strecker 2008). The precipitation rates (mm h−1) from each satellite overpass (~16 day−1) were subsequently rescaled to monthly mean precipitation values using the number of measurements within each grid cell, after which a convolution filter with a 5 × 5 equal weights window was applied. Note that no further calibration of the data was performed. Finally, a dataset of in situ daily accumulated precipitation measurements was compiled. A similar approach was followed for observed monthly and daily mean near-surface temperature observations; however, measurements of this type are very sparse in the region.
Particular attention is paid to the model’s ability to reproduce observed lake surface water temperatures (LSWT), using both in situ and satellite measurements as a reference. To this end, the Along-Track Scanning Radiometers (ATSR) Reprocessing for Climate: LSWT and ice cover (ARC-Lake) project provides high-resolution, spatially explicit global LSWT observations (MacCallum and Merchant 2012). The lake temperature evaluation is complemented using 419 conductivity–temperature–depth (CTD) casts collected during daytime at three multiyear monitoring sites: Ishungu (Lake Kivu), Kigoma (Lake Tanganyika’s northern basin) and Mpulungu (Lake Tanganyika’s southern basin) (Plisnier et al. 2009; Darchambeau et al. 2014). Vertical piecewise cubic Hermite interpolation (increment 0.1 m; De Boor 1978) was applied to each cast, after which the near-surface water temperature (0.5 m) was extracted and used as a reference. This depth is assumed to be the depth closest to the surface with a reliable measurement from the CTD cast.
The net surface shortwave (SWnet) and longwave (LWnet) radiation, latent heat flux (LHF), and sensible heat flux (SHF) were compared to satellite-derived values available from the Global Energy and Water Cycle Experiment Surface Radiation Budget (GEWEX-SRB; Stackhouse et al. 2011), the benchmark synthesis of available diagnostic evapotranspiration estimates (LandFlux-EVAL; Mueller et al. 2013), and the global upscaling of observations from eddy-covariance flux towers using the Flux Network Multi-Tree Ensembles (FLUXNET-MTE; Jung et al. 2010), respectively. The MTE approach represents a machine learning algorithm integrating in situ flux measurements with satellite observations and surface meteorological data to produce global gridded evapotranspiration estimates and associated uncertainties (Jung et al. 2009). Finally, the cloud cover fraction (CCF) is evaluated using the International Satellite Cloud Climatology Project (ISCCP) D2 dataset (Rossow and Schiffer 1999).
All gridded datasets used are available for the entire analysis period, except for CMORPH (2003–08), LandFlux-EVAL (1999–2005), and GEWEX-SRB (1999–2007). Aside from ARC-Lake, their horizontal resolution is usually much coarser than the CTL simulation (Table 1). No height correction, gap filling, or other modifications were applied to the various datasets. The COSMO-CLM2 rotated model grid is considered as the reference grid to which all other gridded products are remapped using bilinear interpolation. Evaluation with in situ data is performed by comparing the observed time series to the model grid box encompassing the measurement site.
e. Surface energy balance decomposition
The energy balance at the surface–atmosphere interface is given by
where is the surface emissivity, is the Stefan–Boltzmann constant (5.67 × 10−8 W m−2 K−4), Ts is the surface temperature, is the surface albedo with respect to shortwave radiation, SWin is the incoming shortwave radiation, LWin is the incoming longwave radiation, LHF and SHF are the turbulent fluxes of latent heat and sensible heat, and G is the combined subsurface storage and conductive heat fluxes.
The causes for the surface temperature response to the presence of lakes are investigated using a surface energy balance decomposition method developed by Juang et al. (2007) and further modified by Luyssaert et al. (2014) and Akkermans et al. (2014). The net impact of the AGL on surface temperature is decomposed a posteriori and attributed to direct contributions of modified biogeophysical processes (such as surface reflection and evapotranspiration) and indirect contributions due to atmospheric feedbacks (such as cloud radiative feedbacks or surface layer stability changes).
The net change in outgoing longwave radiation is given by : that is, the first-order derivative of the left hand side in Eq. (1) with respect to Ts. Note that here, the term surface temperature (Ts) denotes the surface radiative temperature at which the soil–vegetation–lake complex emits longwave radiation to the atmosphere. From this, Ts can be derived by solving the total derivative of the Eq. (1) and solving for Ts:
with SHF encompassing the SHF changes induced by the modified aerodynamic resistance and the modified temperature gradient. Most terms in Eq. (2) can be calculated from the difference between mean quantities in the control and no-lake simulation (CTL minus NOL), except for , which is recomputed from the model equations.
a. Model evaluation
1) Precipitation and air temperature
Observed precipitation patterns are reasonably reproduced by COSMO-CLM2, in terms of spatial patterns (Fig. 2) as well as the amplitude and phase of the annual cycle (Fig. 3). The model captures the enhanced precipitation in the northwestern part of the domain, associated with strong orographically induced convection in a region where tropical rainforest is the dominant land cover type (Figs. 2a–g). However, the disagreement between the various observational products regarding the magnitude of precipitation over this part of the domain is large (with differences locally exceeding 3900 mm yr−1; Fig. 2h), and the bias is either positive or negative, depending on the benchmark product considered. Generally, the limited number of rain gauges within the model domain and the numerous data gaps in the available time series call for cautious use of gridded precipitation datasets based on in situ measurements, but also satellite-based or mixed-source products contain important uncertainties over the mountainous regions of tropical Africa (e.g., Dinku et al. 2008; Nikulin et al. 2012; Endris et al. 2013; Sylla et al. 2013). Furthermore, COSMO-CLM2 is able to reproduce the higher precipitation observed by TRMM 3B42 and TRMM 2B31 over some AGL, especially during the wet seasons [March–May (MAM) and October–December (OND), see below]. In particular, the strong precipitation signal over Lake Kivu is reproduced, while the spatial pattern of precipitation over Lake Victoria is reproduced but underestimated relative to both TRMM products. The dry bias over Lake Victoria relative to TRMM 3B42 and 2B31 amounts to −493 mm yr−1 (−24%) and −691 mm yr−1 (−30%), respectively, and is centered over the western half of the lake, as the modeled precipitation maximum is shifted towards the northeast. Relative to GPCP and CMORPH, in contrast, COSMO-CLM2 overestimates precipitation over Lake Victoria by 217 mm yr−1 (16%) and 250 mm yr−1 (19%), respectively. Again, the maximum difference between the gridded observational products is large (1131 mm yr−1 averaged over Lake Victoria; Fig. 2h), and largely exceeds the biases between COSMO-CLM2 and individual products. Such a large spread calls for further research to quantify the uncertainties of observed precipitation in the region. Furthermore, over the central plateau, where the observational products agree fairly well, a dry bias is apparent. This is confirmed by the comparison to in situ precipitation records, located mostly on the central plateau (Figs. 4a,b). Over 75% of the domain, the absolute precipitation bias against individual products remains below 185 mm yr−1, corresponding to 13% of the average precipitation observed over the region (observational ensemble mean).
Seasonal precipitation variability in the AGL region is primarily governed by the north–south migration of the intertropical convergence zone (ITCZ) across the region (Nicholson 1996; Williams et al. 2015). Monthly mean precipitation subsequently indicates a bimodal distribution with a distinct dry season during June–August (JJA) and two wet seasons: MAM (long rains; northward ITCZ migration) and OND (short rains; southward ITCZ migration), separated by the slight decrease in rainfall during January–February (JF) (little dry season; Fig. 3). The annual cycle of precipitation is reasonably captured by COSMO-CLM2; averaged over the entire domain, fair correspondence exists between the observations and CTL, although precipitation is underestimated in April and during the main dry season. Multiyear monthly accumulated, domain-averaged precipitation biases against individual observational products are smaller than 90 mm month−1 during 90% of all months. Furthermore, monthly mean, domain-averaged modelled precipitation values are positively correlated with corresponding observations [significant (p < 0.001) Spearman rank correlation of 0.72, 0.85, 0.84, 0.80, and 0.84 with respect to TRMM 3B42, GPCC, GPCP, UDEL, and CMORPH, respectively].
Although very sparse, aggregated daily and monthly mean near-surface temperature observations in the region (T2m) suggest the ability of COSMO-CLM2 to reproduce spatial differences in T2m (Fig. 4c). In addition, their annual cycle is well captured by COSMO-CLM2, but the magnitude of the seasonality is slightly exaggerated by the model (not shown).
2) Lake temperature
COSMO-CLM2 succeeds in reproducing the observed differences in LSWT among four major AGL but demonstrates a slight warm bias for each of them (0.68 K on average; Figs. 5a–h; Table 2) and overestimates the amplitude of the seasonal LSWT cycle (Figs. 5i–l). Spatial LSWT patterns correspond to observations especially over Lake Victoria and Lake Tanganyika, with statistically significant spatial rank correlations of 0.68 and 0.61, respectively, and to a lesser extend over the other AGL. Over Lake Victoria, COSMO-CLM2 suggests a positive temperature gradient from southwest to northeast, with a maximum temperature difference of 1 K. Although this modeled pattern is in contradiction with previous studies (e.g., Song et al. 2004; Anyah et al. 2006; Sun et al. 2014b,a), it matches with the spatial pattern derived from remote sensing (Figs. 5a,e), as well as with recent observations based on water temperature measurements at 48 locations during two cruises in February and August 2000 (MacIntyre et al. 2014). Unlike the observations, COSMO-CLM2 furthermore predicts slightly higher lake surface temperatures in coastal zones and embayments as a result of the lower lake depths in these sectors.
The model’s ability to reproduce lake surface temperatures is also apparent from the comparison between modeled and observed lake water temperatures at 0.5 m for three multiyear monitoring sites in Lake Kivu and Lake Tanganyika (Fig. 6). FLake reasonably predicts the annual mean 0.5-m lake temperature (with a warm bias of 1.23°, 0.71°, and 0.60°C at Ishungu, Kigoma, and Mpulungu, respectively) and reproduces the phase of the annual cycle but again exaggerates the amplitude of the seasonal temperature variations. Despite this overestimation, FLake captures the increasing seasonal cycle’s amplitude with distance from the equator (Fig. 6). As this distance grows, the dry season becomes more pronounced, leading to a stronger evaporative-driven cooling of the lakes’ mixed layer (Thiery et al. 2014a). Furthermore, interannual (near) surface lake temperature variability is present but can be considered small compared to the seasonal cycle. Observed anomalies include the high temperatures during the 2005 long rains at Kigoma relative to the preceding months and the low temperatures during the 2007/08 stratified season at Ishungu (Figs. 6a,b). While the model reproduces the former warm anomaly at Kigoma, it does not predict the cold anomaly at Ishungu.
The overestimation of the annual LSWT cycle is primarily due to the underestimation of the mixed layer depth when stratified conditions are imposed in FLake (Thiery et al. 2014a). For a relatively shallow mixed layer, heat is concentrated in the top of the water column rather than being distributed over a larger volume. This, in turn, causes a stronger temperature response to seasonal changes in the lake’s heat content and hence an overestimation of the annual LSWT cycle (Figs. 5i–l). Another potential cause for the remaining biases might be the role of three-dimensional water circulation within the larger lakes. For the relatively shallow Lake Victoria, Song et al. (2004) argue that an RCM coupled to a one-dimensional lake model reproduces first-order spatial lake surface temperature patterns but that only a three-dimensional lake model predicts a southwest–northeast heat transport within the lake. In the deep Lake Tanganyika, actual lake hydrodynamics are still debated (e.g., Naithani et al. 2003; Verburg et al. 2011), but here variable wind forcing (Naithani et al. 2002), differential cooling (Verburg and Antenucci 2010; Verburg et al. 2011), water exchange across the thermocline (Gourgue et al. 2007), and internal Kelvin waves (Naithani and Deleersnijder 2004) may also eventually influence LSWT patterns. Further research could help to discriminate the influence of lake hydrodynamics on lake–atmosphere exchanges for this region.
3) Surface energy balance and clouds
The ability of COSMO-CLM2 to represent the surface climate is investigated by evaluating the surface energy fluxes. Good agreement is noted for the surface radiative fluxes, SWnet and LWnet, both in terms of spatial patterns and the domain-averaged annual cycle, indicating, among other things, the skill of CLM3.5 to simulate spatial and temporal albedo patterns over land (Figs. 7a,b,f,g,k,l). Modeled over-lake variations in SWnet (Fig. 7f) are due to spatial cloud cover variability, given the constant water surface albedo employed in FLake (0.07). Furthermore, very good agreement exists between the model and observations for cloud cover fraction (+4% bias; Figs. 7e,j,o; Table 2). The seasonal cycles of LWnet and CCF are strongly linked to each other and indicate more longwave radiative energy loss during the main dry season, when lower cloud cover fraction reduces the amount of incoming longwave radiation. During the short rains and the little dry season, a slight overestimation of CCF by ~8% induces a negative LWnet bias of −5 W m−2 during these months. Lower performance is noted for the turbulent fluxes, LHF and SHF (−22 W m−2 and +10 W m−2 bias, respectively; Figs. 7c,d,h,i,m,n; Table 2), with generally an underestimation of LHF and an overestimation of SHF. Possibly, this overestimation of the Bowen ratio is caused by limited soil moisture availability, but given the lack of observational evidence, this is impossible to verify. Overall, domain-averaged monthly mean values for SWnet, LWnet, LHF, SHF, and CCF are within the observational uncertainty range during 10, 11, 7, 4, and 11 months of the year, respectively.
4) Comparison to ERA-Interim and CORDEX
To put the performance of the COSMO-CLM2 CTL simulation into perspective, the same evaluation procedure was repeated for ERA-Interim and the COSMO-CLM CORDEX-Africa evaluation simulation (herein referred to as CORDEX; Panitz et al. 2014), each, respectively, representative of a state-of-the-art global reanalysis product and a continental-scale RCM simulation. Results indicate that COSMO-CLM2 largely outperforms ERA-Interim for all considered variables, except for SHF and CCF, where both products depict similar skill (Table 2), and with the seasonal cycles of CCF mostly within the margins of observation uncertainty. Especially for precipitation and lake surface temperatures, COSMO-CLM2 demonstrates remarkably lower biases and spatial root-mean-square errors (RMSE) compared to ERA-Interim. Relative to the CORDEX simulation, the enhanced skill of COSMO-CLM2 is also evident for P, SWnet, and LSWT, whereas LWnet, LHF, SHF, and CCF predictions demonstrate similar skill. Several effects are responsible for this enhanced skill. First, relatively few observations are assimilated into ERA-Interim over central Africa, leaving the model to run more freely over this region compared to other parts of the globe. Second, the use of a lake model clearly improves model performance relative to other common techniques, such as the sea surface temperature interpolation employed in ERA-Interim and CORDEX (Balsamo et al. 2012), for instance by reducing the absolute LSWT error over Lake Tanganyika by 86% and 65% relative to ERA-Interim and CORDEX, respectively (Table 2). Third, COSMO-CLM2 benefits from the higher resolution relative to ERA-Interim and CORDEX (7 vs 79 and 50 km, respectively), allowing for more finescale circulation and associated precipitation patterns to develop over this complex terrain. Finally, the use of the Community Land Model improves the representation of the land surface albedo and other land surface characteristics relative to the default LSM of COSMO-CLM (Akkermans 2013).
In addition, the evaluation procedure was repeated for the NOL simulation, wherein lakes have been replaced by representative land pixels (see section 2c). The added value of the CTL simulation relative to NOL is apparent especially for precipitation and (lake) surface temperature (not shown). For precipitation, the enhanced skill of CTL can be attributed to the underestimation of the moisture input into the atmosphere in NOL (in the absence of lake evaporation), whereas an underestimation of surface thermal inertia deteriorates surface temperature variability in NOL.
In general, the evaluation results demonstrate that the near-surface climate is well represented by COSMO-CLM2, the resolution is sufficient to capture effects of lakes and local orography on precipitation and other atmospheric variables, and the spatial and temporal patterns of the lake surface temperatures as well as differences between individual lakes are well represented by FLake. We therefore conclude that COSMO-CLM2 is an appropriate tool to investigate the impact of the AGL on the regional climate.
b. Impact of the African Great Lakes on the regional climate
Generally, the AGL clearly cool over-lake near-surface air (Table 3; Fig. 8), with average values of −0.67, −0.58, and −0.86 K for Lake Victoria, Lake Tanganyika and Lake Albert, respectively. Locally, the cooling exceeds −1.5 K over each of these lakes, and −2 K over Lake Victoria. An interesting exception to this general pattern is Lake Kivu, where the lake presence induces an average warming of over-lake T2m by +0.75 K relative to the NOL simulation. In section 3c, we further investigate causes of the temperature impact induced by the AGL.
Besides cooling the over-lake surfaces, the AGL also have a pronounced downwind influence on T2m, as the cooler temperatures are advected by the daytime lake breeze into the adjacent land. This is especially visible north of Lake Victoria and around Lake Tanganyika and results in a domain-averaged cooling of −0.17 K. In contrast, the central part of the Albertine Rift mountains [e.g., Rwenzori (0.0°–0.8°N, 29.7°–30.2°E) and Virunga (1.3°–1.6°S, 29.1°–29.7°E) mountain ranges] block all near-surface atmospheric circulation and thereby constrain the influence of Lake Albert, Lake Edward, and Lake Kivu to their direct surroundings. Likewise, the steep topography surrounding Lake Tanganyika [e.g., Itombwe (2.5°–4.5°S, 28.5°–29.1°E), Kibira (2.2°–3.8°S, 29.1°–29.6°E), and Mahale (6.0°–6.5°S, 29.7°–30.1°E) Mountains] limits its cooling influence to the lower inland (Figs. 1, 8). The central plateau south of Lake Victoria is also not influenced by the lake presence, given the predominantly southeasterly flow in this part of the domain. These elements highlight that distance from the lake is not the only decisive factor determining the lake influence on T2m but that topography and atmospheric circulation also play a major role.
The lake influence on T2m is characterized by a clear seasonal pattern, with stronger cooling in the months following the main dry season (August–October: −1.4 K over the AGL vs −0.3 K during the remaining months; Figs. 10a,d), caused by seasonally variable land–water temperature contrasts. Near-surface water temperature variability in the AGL is primarily governed by LHF, for which the lake-induced change peaks during the main dry season (JJA; Figs. 10c,f) as near-surface relative humidity drops and induces the evaporative cooling of the lake surfaces (Savijärvi and Järvenoja 2000; MacIntyre 2012; Schmid and Wüest 2012; Thiery et al. 2014a). Land temperatures follow a similar seasonal pattern, but increase more rapidly at the end of the dry season because of their relatively low heat capacity. Hence, the largest land–lake temperature contrast and associated lake cooling effect is found from August to October. For similar reasons, the strongest AGL cooling occurs during daytime (0700–1600 UTC), because at night lakes have a warming effect (1800–0500 UTC; Figs. 10g,j).
As a result of the lakes’ presence, modeled precipitation is enhanced by 59 mm yr−1 (+5%) on average over the model domain, by 713 mm yr−1 (+79%) over all lake pixels, and by 732 mm yr−1 (+87%) over the four major AGL (Table 3; Fig. 9). The largest increase is noted over Lake Kivu (+1373 mm yr−1 or +145%), whereas the three other major AGL induce similar absolute increases over their surface [+706 mm yr−1 (+85%), +745 mm yr−1 (+82%), and +647 mm yr−1 (+112%) over Lake Victoria, Lake Tanganyika, and Lake Albert, respectively]. The smaller lakes within the model domain generate similar increases, except for Lake Edward, which does not produce any additional precipitation according to COSMO-CLM2.
In contrast to the influence of the lakes on T2m, the precipitation impact is very localized and mostly limited to the lake itself (Fig. 9c). Over Lake Victoria, lake-induced precipitation increases from southeast to northwest, demonstrating the nonuniform response of the atmospheric column to the presence of the lake, in line with Anyah and Semazzi (2004). Over Lake Tanganyika, the increase in precipitation over the central part is absent or weak compared to the northern and southern sectors. In this part of the lake, a narrow corridor induces relatively strong dry season winds (see also Savijärvi and Järvenoja 2000), which transport over-lake air masses westward into the Congo basin (see also Fig. 8a). The differential spatial impact patterns confirm that next to changes in evapotranspiration, temporal patterns and changes in circulation induced by the AGL also play a dominant role in redistributing precipitation (Datta 1981; Savijärvi 1997; Savijärvi and Järvenoja 2000; Song et al. 2004; Anyah and Semazzi 2004; Anyah et al. 2006; Anyah and Semazzi 2007, 2009; Williams et al. 2015). Our findings furthermore suggest that onshore pluviometer measurements will never fully capture precipitation effects induced by the AGL. The dynamical response to lake presence is further investigated in section 3d.
Precipitation over the AGL displays a distinct diurnal cycle, with precipitation occurring mostly at night and morning hours over the lakes (0000–0900 UTC). This is in general agreement with observations from a pluviometer representative of the central Lake Victoria (Datta 1981). For the surrounding land, in contrast, the precipitation maximum occurs around local noon (0700–1200 UTC; Figs. 10h,k). The nighttime precipitation maximum is enabled by the sustained evaporation from the lakes throughout the whole day. This is not the case for the land evapotranspiration, which displays a clear maximum during the local afternoon and slightly lags the precipitation peak (Figs. 10i,l). Overall, the AGL generate a total moisture input into the atmosphere of 222 km3 yr−1, which is 134 km3 yr−1 more than the equivalent land surface (+151%). This increase is more than sufficient to explain the enhanced over-lake precipitation production (+99 km3 yr−1), both on daily and subdaily time scales. On seasonal scales, the impact on precipitation is lowest during the main dry season (JJA: +7 mm month−1 over the AGL vs +77 mm month−1 during the remaining months).
c. Decomposing the lake-induced surface temperature change
An overview of the net changes in the main components of the surface energy balance induced by the lake presence, for all pixels and lake pixels only, is provided in Figs. 11a and 11c, respectively. The increase in net shortwave radiation (SWnet) is caused by the lower albedo of water relative to the surrounding land (−0.12) and reduced cloud cover during daytime. Furthermore, as lakes evaporate at the potential evaporation rate, they enhance evapotranspiration (LHF) relative to the surrounding land, while SHF is reduced (Table 3).
The seven factors contributing to a change in surface temperature are described in Eq. (2) and are expressed in Kelvin. The parameters and flux components responsible for these forcings are shown with their corresponding forcing in Figs. 11b,d. Note that the temperature imbalance is only 0.005 and 0.002 K for the all-pixels and lake-pixels case, respectively, demonstrating balance closure for the depicted components.
Lakes are darker and therefore reflect less incoming radiation, resulting in a positive contribution to Ts from changing albedo of +0.51 K over the whole domain and +4.91 K over the lakes (Figs. 11b,d: ). On the other hand, the contribution to Ts caused by the higher surface longwave emissivity of water relative to the surrounding vegetation is negligible (Figs. 11b,d: ). A strong increase in evaporation from the lakes contributes to a lower Ts by −1.30 K over the whole domain and up to −12.21 K over the lakes and forms the largest individual contribution to Ts (Figs. 11b,d: LHF). Note that, since the change in evapotranspiration is influenced by the latent heat of vaporization and air saturation (in turn, depending on temperature), it is impossible to distinguish between direct and indirect components in the case of LHF (Akkermans et al. 2014). The lower roughness length of the lake surface relative to vegetated land increases aerodynamic resistance and therefore slows down the turbulent heat transfer from the surface to the atmosphere. The upward SHF is also modified by a changing temperature gradient within the atmospheric surface layer. Together, both processes induce a warming of the surface by +0.81 K over the whole domain, and +8.64 K over the AGL, thereby presenting the strongest positive impact of Ts (Figs. 11b,d: SHF). Furthermore, since daytime cloud cover over the AGL is not significantly modified (+2%), the contribution of changing incoming shortwave radiation to Ts is negligible over the AGL (Figs. 11b,d: SWin). The same can be said for the contribution from changing incoming longwave radiation, a feedback depending both on cloud cover and atmospheric temperature (Figs. 11b,d: LWin). Finally, the impact of a changing subsurface heat flux represents a small positive contribution to Ts (Figs. 11b,d: G).
As expected, the temperature impact of lake presence is much larger over the AGL compared to the rest of the domain (Table 3). However, due to the dynamical changes induced by the AGL and associated changes in near-surface temperature and moisture (see section 3d for a detailed description), some components are altered over land as well. For instance, the warming effect caused by the enhanced daytime incoming solar radiation (SWin) becomes slightly more important when considering the whole domain, while the influence of a change in incoming longwave radiation (LWin) switches sign (Figs. 11b,d).
Altogether, individual contributions to Ts result in surface temperature increases of 0.26 and 3.02 K over the whole domain and the AGL, respectively (Figs. 11b,d; Table 3). This seems to be in contradiction with the simulated change in T2m, which was shown to decrease on average (Figs. 8, 11b,d; Table 3). This apparent contradiction is, however, solved when comparing the forcing components separately during daytime (0900, 1200, 1500, and 1800 UTC) and nighttime (2100, 0000, 0300, and 0600 UTC; Fig. 12a). Note that the full surface energy balance decomposition method cannot be applied on subdaily time scales due to the lag effect of the heat storage component; hence, is computed from Eq. (2) in this case.
Although Ts and T2m depict the same sign on subdaily time scales, their magnitudes vary. During daytime, both Ts and T2m are lower in CTL relative to NOL, as a large fraction of the solar energy penetrates the lakes rather than being absorbed at the surface (Fig. 12a: G). At night, when this heat is released again from the mixed layer to the surface, the opposite effect is seen. The contribution of G to Ts is only partly offset by reduced SHF during daytime and enhanced LHF during nighttime in CTL relative to NOL (Fig. 12a: SHF and LHF). Furthermore, the lake influence on vertical turbulent heat transport explains the differences in magnitude between Ts and T2m: during daytime, the strong decrease in SHF over the lakes (shown by the strong positive contribution of SHF to Ts in Fig. 12a) effectively inhibits heat transfer away from the surface—hence, the more pronounced decrease in T2m relative to Ts. At night, the relative contribution to Ts vanishes as SHF drops over land while remaining more or less constant over lake surfaces (Fig. 12a), but again low SHF limits warming at 2 meters (see also Fig. 7i). Hence, when considering the whole day, the nighttime warming dominates Ts, whereas the daytime cooling dominates T2m (Figs. 11d, 12a).
Lake Kivu constitutes an interesting exception, as argued in sections 3b(1) and 3b(2), given the positive T2m anomaly and strong precipitation increase generated by its presence (Figs. 8, 9). The reasons for this behavior become apparent when applying the surface energy balance decomposition method to this lake alone. Over Lake Kivu, daytime Ts increases as SHF is strongly dampened during this time (+25.81 K contribution to Ts; Fig. 12b). Furthermore, the lake presence lowers daytime cloud cover and therefore enhances solar radiation input. Both effects are only partly offset by changes in other contributions to Ts (Fig. 12b: G, , and LHF). At night, the heat release from Lake Kivu is modeled to be more effective relative to the other AGL, further strengthening the nighttime warming influence. The different response of the various surface energy balance components is likely caused by the high altitude of Lake Kivu (1463 m MSL) relative to other AGL. In our simulation, the surface layer over Lake Kivu is predicted to be very unstable, with an average 3.4-K temperature difference between the lake surface and 2-m level (relative to 1.7, 1 .7, and 1.2 K over Lake Victoria, Lake Tanganyika, and Lake Albert, respectively). Overall, the absence of daytime Ts cooling and strong nighttime Ts warming in CTL relative to NOL causes Lake Kivu to display a general warming influence upon the near-surface air. In addition, the strong nighttime warming intensifies the dynamical response to lake presence (see also section 3d), leading to enhanced atmospheric column destabilization (Fig. 15b) and associated precipitation production.
d. Dynamical response to lake presence
The dynamical response of the atmosphere to lake presence is further investigated using a set of vertical cross sections along a transect covering the northern sector of Lake Victoria (Fig. 1). Based on the diurnal variability of lake impact on T2m and precipitation (Figs. 10g–l), we construct two composites of the 1999–2008 3-hourly mean model output, with a first composite averaging output during daytime (0900, 1200, 1500, and 1800 UTC; Fig. 13) and a second during nighttime (2100, 0000, 0300, and 0600 UTC; Fig. 14).
During daytime, the atmospheric temperature decrease induced by the lake extends well beyond its shores and up to ~2-km height above the surface (Fig. 13c). At 2 km MSL, annual average daytime temperature still decreases by around −0.14 K over the whole model domain and by −0.40 K over lake surfaces (both changes statistically significant at the 5% significance level). The pressure increase resulting from this cooling (in excess of +0.60 hPa; Fig. 13e) generates a lake breeze between ~1–2 km MSL across the lake shore towards the western rift valley (Fig. 13d). The lake breeze is offset by updrafts over land and with westerlies at ~3–5 km MSL and an over-lake, drying subsidence flow closing the secondary circulation cell over Lake Victoria’s western sector (Fig. 13d). Although the moisture buildup in the persistently unstable surface layer (Verburg and Antenucci 2010) tends to destabilize the atmospheric column, the clockwise circulation induces a stabilization of the atmospheric column, especially over the western sector of the lake, as reflected by an increased equivalent potential temperature above ~1.5 km MSL and extending up to ~6 km MSL (Fig. 13f):
where is the latent heat of evaporation (2.50 × 106 J kg−1), is the specific heat capacity of air at constant pressure (1005 J kg−1 K−1), is the specific humidity (kg kg−1), p0 is the standard reference pressure (1000 × 102 Pa), and p is the surface air pressure (Pa). The at lower levels enhances as the effect of increased humidity (Fig. 13d) overcompensates the effect of lower temperatures and higher air pressure (Figs. 13c,d). The stabilization of the air results in a decrease in upward convective mass flux density (i.e. the vertical mass transport per unit area at cloud-base height due to convection; Fig. 15a), less intense convection, and, consequently, less rainfall during daytime over the lake. In contrast, the lake presence enhances the convective mass flux density during daytime around the lake (Fig. 15a) and therefore daytime precipitation.
In the eastern lake sector, the vicinity of the Kenyan Rift Valley mountains (Fig. 13a) modifies the classic lake-breeze circulation as observed over the western sector. Here, the primary daytime circulation is one of orographically induced convection (Fig. 13b), associated with an anabatic wind advecting the over-lake moisture upslope up to ~3 km MSL, confirming the conclusion from Anyah et al. (2006) that the anabatic daytime circulation is enhanced by the lake breeze. The strengthening of the anabatic component induces near-surface divergence and associated downward secondary flow aloft from ~2 to 5 km (Fig. 13d). Consequently, the primary upward motion is already suppressed onshore, causing less intense stabilization over the eastern sector relative to the western part (Figs. 13d,f) and consequently a more limited decrease in convective mass flux density (Fig. 15a).
At night, the lake surface warms the boundary layer and generates a statistically significant pressure deficit of about −0.12 hPa at the surface (Figs. 14c,e). As a response, the secondary circulation over the western sector is effectively reversed, with the lake breeze being replaced by a weak land breeze from the colder land to the warmer lake surface and offset by a counterclockwise motion aloft (Fig. 14d). When advected over the water, the cold air is warmed and moistened by strong evaporation from the lake (LHF ~100 W m−2 higher than equivalent land points; Fig. 10l). While the near-surface air converges and induces upward motion, the warming signal dissipates more quickly in the vertical due to adiabatic decompression, and the enhanced moisture is efficiently transported aloft (Figs. 14c,d). At 1 km MSL, annual average nighttime specific humidity still increases by around +0.30 g kg−1 over the whole domain and by +1.66 g kg−1 over lake surfaces (both changes statistically significant at the 5% significance level). Also note that, again, this is the opposite of the daytime, when downward motion inhibits upward moisture transport building up in the boundary layer and rather transports it horizontally across the lake boundaries (Fig. 13d). At night, higher air temperatures, higher specific humidity, and lower air pressure reinforce each other in destabilizing the first ~1 km of the atmospheric column over Lake Victoria [see Eq. (3)], as reflected by the strong increase in there (~+6 K; Figs. 14c–f). This time, the increase of in lower levels is not accompanied by a notable decrease at higher levels caused by secondary circulation, pointing to the general destabilization of the nighttime over-lake atmospheric column. This, in turn, causes an increase in upward convective mass flux density (Fig. 15b), more intense convection, and associated precipitation during nighttime over the lake. Again, the sign is reversed around the lake, where the lake presence reduces convection and, therefore, precipitation.
At night, the difference between the western and eastern sectors is generally less pronounced, although some differences can be noted. Over the eastern sector, the nighttime land breeze enforces katabatic winds flowing down from the eastern mountains (Anyah et al. 2006) and inhibits upslope transport of the warm temperature anomaly (Figs. 14b–d). The enhanced katabatic wind generates near-surface divergence on the eastern mountain slope and associated secondary downward motion aloft from ~2 to 4 km MSL, displaying similar secondary circulation over land compared to daytime, although less pronounced (Fig. 14d).
The influence of the AGL on the regional climate was studied for the period of 1999–2008 with the regional climate model COSMO-CLM2. Efforts were made to enhance the realism of the simulations by accounting for (i) a model setup and land surface model suited for tropical conditions; (ii) a horizontal grid resolution effectively resolving individual lakes and complex topography, unprecedented for climate simulations in this region; and (iii) a state-of-the-art one-dimensional lake model capable of reproducing tropical LSWT.
A comprehensive evaluation of COSMO-CLM2, nested within the ERA-Interim-driven CORDEX-Africa simulation (COSMO-CLM), shows an adequate representation of precipitation and LSWT; the model reproduces the most important spatial patterns, including enhanced over-lake precipitation, absolute LSWT values, and LSWT gradients within and between lakes. The phase and amplitude of the mean annual cycle are also similar to observations, although the amplitude is overestimated for LSWT, and precipitation is underestimated during April and the main dry season. The remaining LSWT biases are ascribed to an underestimation of the mixed layer depth and the absence of three-dimensional circulation in the larger lakes. Furthermore, the mean annual cycles of net shortwave and longwave radiation at the surface, sensible and latent heat flux, and cloud cover are mostly simulated within the margins of observational uncertainty. Finally, we show that our simulation largely outperforms a state-of-the-art reanalysis product (ERA-Interim) for most of the considered variables, especially precipitation and LSWT, whereas the added value relative to a state-of-the-art, continent-scale RCM simulation (COSMO-CLM CORDEX-Africa evaluation simulation) is evident for precipitation, LSWT, and net surface shortwave radiation.
The AGL significantly reduce offshore near-surface air temperature by about −0.57 K, with maxima in excess of −1.50 K. The cooling effect is advected across the lake shores within dynamic and orographic constraints and is found to be strongest at the end of the main dry season, when the land surface warms faster than the water surface and the lake–land temperature contrast reaches a maximum. The four major AGL also enhance precipitation by +732 mm yr−1 (+87%) over their surface and even +1373 mm yr−1 (+145%) over Lake Kivu. All lakes together annually evaporate 222 km3 of water into the atmosphere. In contrast to the near-surface temperature impact, the precipitation change is highly restricted to the lake areas. Both for temperature and precipitation, the mean effect masks a pronounced diurnal pattern: the temperature signal exhibits strong cooling during daytime and moderate warming during nighttime over the lakes, whereas precipitation is enhanced mostly at night and during early morning.
Decomposition of the lake-induced surface temperature increase over the AGL shows that reduced albedo has a moderate warming influence (+5.91 K), while reduced sensible heating enhances surface temperatures even more (+8.64 K). For the most part, their effect is offset by the enhanced lake evaporation, responsible for −12.21 K cooling. The apparent contradiction between surface and near-surface temperature change is cleared by considering daytime and nighttime separately: daytime heat storage in the lake and reduced upward sensible heat flux dominate the near-surface air temperature change, whereas the nighttime warming determines the surface temperature signal.
Finally, analysis of the dynamical response using daytime and nighttime cross sections over Lake Victoria highlights the importance of circulation changes induced by the lake–land temperature contrast. During daytime, the lake breeze transports cold air across the lake borders and generates overland updrafts and over-lake subsidence. This secondary circulation stabilizes the atmosphere above ~1.5 km and therefore effectively suppresses convection from the unstable surface layer. At night, the thermal inertia of the lake surface generates a positive temperature anomaly and a pressure deficit and maintains the daytime evaporation rates, inputting large amounts of moisture into the boundary layer. These three effects together cause a strong destabilization of the lower atmosphere. As the land breeze and secondary circulation subsequently induce near-surface convergence and the lifting of these highly unstable air masses, strong convection is triggered, and precipitation is released over the lake. On the eastern shore, complex topography and associated gravity currents superimpose on this pattern, generating a somewhat different response, especially during daytime.
We would like to thank the CLM-Community (http://www.clm-community.eu) and, in particular, Tom Akkermans and Alessandro Dosio for the helpful discussions on the modelling with COSMO-CLM2. François Darchambeau, Pierre-Denis Plisnier, Meteo Rwanda, and Patrick Willems are thanked for supplying observational data, Thomas Antheunis for his thoughts on the surface energy balance decomposition technique, and Sonia Seneviratne for supporting this research. We sincerely thank three anonymous reviewers for their constructive remarks. Wim Thiery, Matthias Demuzere, and Stef Lhermitte are research fellows at the Research Foundation Flanders (FWO). This work was further supported by the Belgian Science Policy Office (BELSPO) through the research projects EAGLES, CLIMLAKE, and CLIMFISH. Parts of the station data were obtained from the FRIEND/NILE project of UNESCO and the Flanders in Trust Fund of the Flemish Government of Belgium. The computational resources and services used in this work were provided by the VSC (Flemish Supercomputer Center), funded by the Hercules Foundation and the Flemish Government department EWI.