1 Introduction

Lake Malawi (also known as Lake Nyasa in Tanzania and Lago Niassa in Mozambique) is the ninth largest and third deepest fresh water lake on Earth (Bootsma and Hecky 1993). It holds nearly 7% of the Earth’s available surface freshwaters (Branchu et al. 2005) and is home to a great diversity of fish species (Guildford et al. 1999; Banda et al. 2001). Lake Malawi is located in the East African Rift Valley and is bordered by Malawi, Mozambique and Tanzania. Its water balance is dominated by rainfall over the lake and evaporation, with a small contribution of river inflow and discharge (Spigel and Coulter 1996). The climate of the Lake Malawi’s region is characterized by a strong seasonal cycle marked by one rainy season during the austral summer, which results from the region’s location at the southern edge of the seasonal migration of the inter-tropical convergence zone (ITCZ; Eccles 1974; Brown et al. 2000). This seasonality is reflected in the lake’s mixing regime, as the rainy season (November to March) is characterized by enhanced water column stratification due to relatively warm summer temperatures and weak winds (Eccles 1974).

Lake Malawi is approximately 570 km long and 75 km wide, and therefore a high horizontal resolution (grid spacing of the order of ten Kilometre (10 km) or less) is needed in order to capture the lake’s effects on the surrounding region. This resolution is well beyond that of current global climate models (GCMs), but can be reached with the use of regional climate models (RCMs). In particular, RCMs coupled to one-dimensional lake models have been extensively used in a wide variety of studies over different lake basins, including model performance assessments (e.g. Bates et al. 1993, 1995; Small et al. 1999; Martynov et al. 2012; Turuncoglu et al. 2013; Williams et al. 2015; Sun et al. 2015) and the study of lake-effects on the surrounding climate (Bonan 1995; Anyah and Semazzi 2004; Anyah et al. 2006; Martynov et al. 2012; Notaro et al. 2013; Vavrus et al. 2013; Nicholls and Toumi 2014; Thiery et al. 2015; Nyamweya et al. 2016; among others). These studies focused on lakes in Europe (e.g. Caspian Sea Basin), North America (e.g. Great Lakes) and Central-eastern Africa (e.g. Lake Victoria, Lake Tanganyika) and have shown that the presence of large lakes can indeed modify regional climates, in particular through the recycling of lake water evaporation and the intensification of overlying storms. Despite the importance of this local forcing, to our knowledge no studies have explicitly applied coupled RCM/lake models to Lake Malawi towards the assessment of the complex lake–atmosphere interactions occurring in the region.

As a first step towards the production of climate change scenarios over the Lake Malawi area, in this paper we implement a double nested coupled RCM/lake model for the region and investigate the model performance in reproducing local climate and lake variables, along with the lake-effect on precipitation, hydrologic components and atmospheric circulation. Specifically, we employ the International Centre for Theoretical Physics (ICTP) regional model RegCM4 (Giorgi et al. 2012) interactively coupled to a one-dimensional lake model (Hostetler et al. 1993; Small et al. 1999) to perform two continuous 17 years simulations over the Lake Malawi region. In order to achieve a high horizontal resolution, we run our coupled model in a double one-way nested mode in which a 10 km grid spacing domain covering the Lake Malawi area is nested within a larger 25 km grid spacing domain covering the entire Southern Africa region, in turn driven by initial and lateral boundary conditions (LBCs) from an atmospheric general circulation model (AGCM). In order to investigate lake effects we compare simulations in which the lake model is either included or entirely removed and replaced by a grassland surface type. We analyze both climate and lake variables by comparison with available station and satellite observation datasets.

Section 2 presents the method, model configuration and experiment design, including the satellite and gauge based observation datasets used for model validation. Results and discussions are presented in Sect. 3, with focus on mean seasonal and annual cycle of 2 m surface air temperature, precipitation and lake surface temperature. The effects of Lake Malawi on precipitation, surface hydrologic components and atmospheric circulations are then investigated, while Sect. 4 presents our main conclusions. We finally point out that the present experiments were conducted as part of the project Socioeconomic Consequences of Climate Change in Sub-Equatorial Africa (SOCOCA, http://www.mn.uio.no/geo/english/research/projects/sococa/), aimed at studying the impacts of climate change on southern Africa and the Lake Malawi basin.

2 Model description, experimental design and observational datasets

2.1 Model description

In this study we use the latest version of the regional climate modeling system RegCM4 developed at the Abdus Salam International Centre for Theoretical Physics (ICTP), which is described by Giorgi et al. (2012). RegCM4 is a compressible, primitive equation, sigma vertical coordinate model with dynamics based on the hydrostatic version of the National Centre for Atmospheric Research/Pennsylvania State University’s mesoscale meteorological model version 5 (NCAR/PSU’s MM5; Grell et al. 1994; see also Giorgi et al. 1993a, b). The model includes multiple options of physics parameterizations. Based on previous studies (e.g. Sylla et al. 2012; Mariotti et al. 2014; Diallo et al. 2015, 2016; Tall et al. 2017; N’Datchoh et al. 2017), as well as a series of preliminary tests over the selected domain, the following options are used here: a modified radiative package of the NCAR Community Climate Model version 3 (CCM3; Kiehl et al. 1996), the nonlocal vertical diffusion scheme of Holstag et al. (1990) to represent boundary layer processes, the sub-grid explicit cloud scheme of Pal et al. (2000), the ocean flux scheme of Zeng et al. (1998), the Biosphere–Atmosphere Transfer Scheme (BATS; Dickinson et al. 1993) for the representation of the land surface processes and the cumulus convection scheme of Grell et al. (1994) (with the closure of Fritsch and Chappell 1980; Frich et al. 2002) to represent convective precipitation.

RegCM4 is interactively coupled to the one-dimensional (1D), energy-balance lake model described by Hosteler et al. (1993) and Small et al. (1999). The lake model is run at each grid point of Lake Malawi with a vertical resolution of 1 m (dz = 1 m) and the number of layers is determined by the lake bathymetry at that grid point location. The model allows for vertical energy transfer within each lake layer through eddy and molecular diffusion and vertical convective mixing, but it does not allow for exchanges across lake points in the horizontal. RegCM4 provides as input to the lake model all surface energy and mass fluxes, along with precipitation. The lake model computes the vertical water temperature profile and the lake surface temperature (LST) at each lake point, which is then passed back to the RegCM4. The coupled system RegCM4/lake model has been recently assessed over the Caspian Sea Basin (Turuncoglu et al. 2013) and the Great lakes of the United States (Notaro et al. 2013 and references therein).

2.2 Experimental design

We use a double one-way nesting approach in which an intermediate resolution configuration of RegCM4 (25 km grid spacing) is first run to provide meteorological boundary conditions for a high resolution inner domain. The intermediate resolution domain and simulation is from Diallo et al. (2015), and covers the entire southern Africa region as shown in Fig. 1a. This simulation is driven at the lateral boundaries by meteorological fields from the National Centre for Atmospheric Research (NCAR) Community Atmospheric Model version 4 (CAM4; Gent et al. 2011) run at a horizontal resolution of 0.9° × 1.25° (Lat × Lon). Sea surface temperatures (SSTs) are prescribed from a corresponding simulation conducted with the Community Climate System Model version 4 (CCSM4; Gent et al. 2011), and both CAM4 and the 25 km mother-domain RegCM4 run do not include Lake Malawi. The simulation covers a 17-year period nominally referred to as 1992–2008, where the first 2 years are discarded to allow spin-up of land-surface fields. The results are thus only analyzed for the 15-year period 1994–2008. The reader is referred to Diallo et al. (2015) for more details on the intermediate resolution simulation set up.

Fig. 1
figure 1

Domain, topography (in meter) for: a 25-km RegCM4 (mother domain; RegCM4_R25), b 10-km RegCM4 (nested domain) and c 1D lake model bathymetry (in meter) used in 10-km RegCM4 (RegCM4_R10L). White box in a denotes the position of the nested 10-km domain. The red box in b indicates the ‘‘Malawian region’’ (26°E–41°E and 19°S–8°S) and Lake Malawi (white box, 14°S–09°S and 33.75°E–35.5°E) averaging regions used in the analysis

Full size image

Figure 1b shows the double nested inner domain, which includes Malawi and its neighboring countries at a grid spacing of 10 km. Two simulations are completed for this double nested domain. In the first experiment (hereafter referred to as RegCM4_R10L), Lake Malawi is represented by the one-dimensional lake model interactively coupled to RegCM4 at each lake point of the domain grid. The lake temperatures are initialized following the mean temperature profile estimate of Vollmer et al. (2002) and, as mentioned, the lake points do not interact horizontally with each other. The lake bathymetry is shown in Fig. 1c. In the second double-nested experiment (hereafter referred to as RegCM4_R10NL), Lake Malawi is removed and the lake grid cells are replaced by grassland surfaces. This land cover type is selected as it represents the most common across the surrounding region.

We stress that this experimental set up was used for consistency with the modeling stream in the SOCOCA project. Alternatively, global analyses of observations could have been used to drive the model, however Diallo et al. (2015) showed that the intermediate resolution simulation exhibited a good performance in reproducing the climatology of this region, and in this paper we mainly focus on climatological responses in the model domain rather that individual specific years or events. Also, in this paper we focus on the double nested domain, since we are mostly interested in lake effects, and we refer readers to Diallo et al. (2015) for the analysis of the intermediate resolution simulation (which does not include the lake).

2.3 Observational datasets

The model variables are assessed against multiple gauge and satellite based observation datasets in order to account for uncertainties in observations over the African continent (Sylla et al. 2013; Mariotti et al. 2014; Diallo et al. 2013, 2014; Fotso-Nguemo et al. 2017): (1) the monthly mean 2 m-temperature and precipitation dataset at 0.5° × 0.5° resolution from the University of Delaware (UDEL; available from 1900 to 2010; Legates and Willmott 1990); (2) The 0.5° resolution monthly means of precipitation and temperature (version 3.24) from the Climatic Research Unit of the University of East Anglia (CRU TS3.24; http://badc.nerc.ac.uk/browse/badc/cru/data/cru_ts/cru_ts_3.20/data/; available from 1901 to 2015; Harris et al. 2014); (3) the monthly mean combined satellite-gauge precipitation data product at 2.5 × 2.5° horizontal resolution from the Global Precipitation Climatology Project (GPCP Version 2.2; Huffman et al. 2007; Adler et al. 2003; available from 1979 to present); and (4) the monthly mean satellite-derived rainfall estimates at 0.25° × 0.25° resolution from the Tropical Rainfall Measuring Mission (TRMM VB42V47, available from 1998 to present; Huffman et al. 2001, 2007). In this regard, we note that differences between gauge based and satellite based observations, although significant, do not qualitatively change the assessment of the model performance (e.g. Diallo et al. 2012, 2016).

The latest release of the ARC-Lake project (version 1.1.2; MacCallum and Merchand 2011, 2012) satellite observation dataset is used for the validation of lake surface temperature (LST). The ARC-Lake v1.1 data product covers the period from 1st June 1995 to 31st December 2009 with a spatial resolution of 0.05°, and a detailed description of the ARC-Lake dataset can be found in http://www.geos.ed.ac.uk/arclake.

The lake effects on different surface variables (latent heat flux, 2 m-temperature, sensible heat flux, and precipitation) are analyzed for the monsoon onset phase, October through December (OND), the monsoon receding phase, January through March (JFM) as well as the dry seasons April through June (AMJ) and July through September (JAS), when the lake effects are less important. To match the higher resolution of the RegCM4 simulations and to facilitate inter-comparison, all validation datasets are mapped onto the model grid using a bilinear interpolation procedure. The model performance is assessed using three statistical verification metrics: (1) the mean bias (MB), which indicates if the simulations overestimate/underestimate the mean magnitudes of the observed values; (2) the root mean square error (RMSE), representing the overall accuracy of the model simulations with respect to the observations; and (3) the pattern correlation coefficient (PCC, i.e. spatial correlation). These metrics are computed over the region covering most of the interior land domain, or “Malawian region” (Malawi, Mozambique and Tanzania), defined as 26°E–41°E and 19°S–8°S (red box in Fig. 1b).

3 Results and discussions

3.1 Model performance: mean climatology and seasonal cycle

Figure 2 shows the mean seasonal 2 m surface air temperature for OND, JFM, AMJ and JAS averaged for the period 1994–2008 from CRU, UDEL, and RegCM4_R10L along with the corresponding biases (defined as the difference of simulated minus observed data). In all seasons, observations show maximum temperatures (>25 °C) over the flatter coastal Mozambique region adjacent to the Indian Ocean and over a tilted band encompassing southwest Mozambique and southern Zambia. Lower temperatures are found in the northwest sector of the region and towards central Zimbabwe. Slight differences between the two observation data are evident. For instance UDEL is cooler than CRU in JFM, AMJ and JAS over north Zambia, whereas in both OND and JFM UDEL is warmer over Lake Malawi (Fig. 2b, g). RegCM4_R10L reproduces reasonably well the observed spatial temperature patterns in the wet seasons, with biases mostly lower than 2° throughout the domain and 1.5° in the Lake Malawi region. In the dry seasons the bias is positive over the northwestern regions of the analysis domain, but less than 1.5° in the southern ones.

Fig. 2
figure 2

Mean October–December (OND, a–c), January–March (JFM, f–h); April–June (AMJ, k–m) and July–September (JAS, p–r) 2 m surface air temperature (unit: °C) climatology from observations (CRU and UDEL) and RegCM4_R10L along with the corresponding biases [(d, e, i, j, n, o, s, t); unit: °C] with respect to both observations for the period 1994–2008

Full size image

Figure 3 shows the mean precipitation for OND, JFM, AMJ and JAS averaged for the period 1994–2008 from CRU, UDEL and RegCM4_R10L along with the related biases, while the same field for TRMM and GPCP are shown in Supplementary Figure S1. The observed average precipitation during OND is greater than 2 mm/day over most of the domain, with maxima over the north-western and southeastern regions, which are in fact more pronounced in GPCP and TRMM (3–8 mm/day) than CRU and UDEL (see Supplementary Figure S1a, b). In JFM the precipitation band moves to the north and exhibits three rainfall peaks of 7–10 mm/day over central Tanzania, toward the Mozambique channels and across northern Zambia. In AMJ and JAS, the dry seasons over the Malawian region, the mean rainfall over most of the domain does not exceed 1 mm/day. The four observation datasets show generally similar patterns, although some discrepancies are found in the spatial location and magnitude of the maxima.

Fig. 3
figure 3

Mean October–December (OND, a–c), January–March (JFM, f–h); April–June (AMJ, k–m) and July–September (JAS, p–r) precipitation (unit: mm day−1) climatology from observations (CRU and UDEL) and RegCM4_R10L along with the corresponding biases [(d, e, i, j, n, o, s, t); unit: mm day−1] with respect to both observations for the period 1994–2008

Full size image

RegCM4_R10L is generally able to reproduce the main characteristics of the precipitation patterns; however it tends to overestimate precipitation, particularly in the southern sector during OND (Fig. 3d, e), while it underestimates precipitation over northern Mozambique and southern Tanzania, in particular during JFM (Fig. 3i, j) and AMJ (Fig. 3n, o). Note that similar bias distributions are found when comparing RegCM4_R10L against either the TRMM or the GPCP satellite-based datasets; although the magnitude and location of the bias patterns are slightly different (see Supplementary Figure S1). This bias pattern appears to reflect a southward shift of the ITCZ compared to observations, and is similar to that found in the driving AGCM and 25 km resolution RegCM4 simulation (Diallo et al. 2015), which indicates that it may be at least partially related to the lateral boundary forcing. However, the convection scheme used within this hydrostatic model may also play a role, as we are approaching resolutions for which the hydrostatic assumption becomes less valid.

To provide a more quantitative assessment for the overall performance of RegCM4_R10L, we computed the MB and PCC for the OND, JFM, AMJ and JAS mean seasonal and annual 2 m-temperature and precipitation over the inner domain land areas of over the “Malawian region” (red box of Fig. 1b). The results are summarized in Table 1. The temperature MB with respect to CRU ranges from 0.40 to 1.32 °C while it is 0.78 to 1.47 °C compared to UDEL. The temperature PCC is 0.86 or higher against both observation datasets, with lower correlation coefficients and warmer biases during the dry season (AMJ & JAS; Fig. 2n, o, s, t). Table 1 shows that the model tends to underestimate precipitation (negative MB), especially in JFM, while the PCC is high for the individual seasons (0.72–0.88), but lower for the annual average (0.52–0.58).

Table 1 Mean bias (MB, units : °C or mm day−1), pattern correlation coefficient (PCC) of October–November–December (OND), January–February–March (JFM), April–May–June (AMJ), July–August–September (JAS) and annual mean 2 m-temperature and precipitation during 1994–2008 in RegCM4R_10L simulation compared to observations (CRU and UDEL) over the “Malawian region” (26°E–41°E and 19°S–8°S)
Full size table

Figure 4 shows the mean annual cycle of surface air temperature and precipitation averaged over the inner domain land areas in the model and in the different observational datasets, while Table 2 provides the corresponding temporal MB, RMSE and PCC. The model exhibits a warm bias of 1–2 °C during the dry season months, April through July, while during the rest of the year, and in particular the wet season, the bias is less than 1 °C (Fig. 4a). Averaged over all months, the MB is positive, but less than 1 °C. For precipitation, all the observation datasets show a marked seasonal cycle with a rainy season between November and March and close to no precipitation between July and September. The model reproduces this seasonal cycle feature, but underestimates precipitation amounts during the peak rainy months of December–January. As shown in Fig. 3, this is mostly due to an underestimation over the southern Tanzania region. Both for temperature and precipitation the model reproduces closely the observed seasonal cycle, with PCC exceeding 0.9 (see Table 2).

Fig. 4
figure 4

Mean annual cycle over the “Malawian region” (26°E–41°E and 19°S–8°S) of 2 m surface air temperature (a; unit: °C) from UDEL, CRU and RegCM4_R10L and precipitation (b; unit: mm/day) from UDEL, CRU, GPCP, TRMM and RegCM4_R10L. 2 m-surface temperature and precipitation are averaged over 1994–2008 for observations (CRU, UDEL and GPCP) and RegCM4_R10L. TRMM is averaged for the period 1998–2008

Full size image
Table 2 Mean bias (MB, units: oC or mm day−1), root-mean-square error (RMSE, unit: standard deviation) and pattern correlation coefficient (PCC) of mean annual cycle 2 m-temperature and precipitation for the period 1994–2008 in RegCM4R_10L simulation against observations (CRU and UDEL) averaged over the “Malawian region” (26°E–41°E and 19°S–8°S)
Full size table

In summary, despite the presence of some biases in the model, most noticeably a warm bias in the dry season and a precipitation underestimation over the northeastern portion of the domain in the wet season, the model shows a reasonably good performance in simulating the climatology of 2 m surface temperature and precipitation over the region, in line with previous coarser resolution simulations (Sylla et al. 2012; Ogwang et al. 2014; Li et al. 2015; Diallo et al. 2015; Moalafhi et al. 2016). In the next section we turn our attention to the simulation of lake variables and lake effects by the coupled modeling system.

3.2 Lake simulation and lake effects

Here we first provide an assessment of the model simulation of lake surface temperatures and then we analyze the simulated lake effects through a comparison between lake and no-lake experiments.

3.2.1 Simulation of lake surface temperature (LST)

The LST is one of the most important parameters affecting the interaction between the atmosphere and the lake. Figure 5 displays the spatial distribution of seasonal mean OND, JFM, AMJ and JAS LST from the ARC-lake satellite observations and the RegCM4_R10L simulation averaged over the period 1996–2008, while Fig. 6 shows the corresponding LST biases (RegCM4_R10L minus ARC-lake). Both in JFM and OND the observations show maximum lake temperatures along the western shores, up to about 30 °C and a gradient towards the east, where minimum temperatures are found over the eastern coastal areas of about 25 °C. This temperature gradient is associated with the presence of an upwelling circulation across the lake (Vollmer et al. 2002; Chavula et al. 2009), and is confirmed by in situ water temperature data (Hamblin et al. 2003; Woltering et al. 2011; MacCallum and Merchand 2012). Conversely, during the dry seasons (AMJ and JAS) the temperature gradient is reversed, with maximum LSTs (up to 24 °C in AMJ and up to 27 °C in JAS) located along the northeastern regions of the lake and decreasing towards the southwest. By and large, the model reproduces these LST patterns and their seasonal evolution, indicating that the LST is mostly driven by the atmospheric forcing. In the model, the west-east gradient is overestimated in OND and JFM, resulting in a positive temperature bias of up to a few degrees over the western shores and a larger negative bias (up to 4 °C) over the eastern ones (Fig. 6). This is evidently due to the lack of horizontal circulations and mixing, particularly in the east–west direction. In the dry seasons the bias is more uniform across the lake surface, being positive in AMJ and negative in JAS (in both cases between 0.5 and 1.5 °C in magnitude). The positive AMJ LST bias may be related to the warm bias observed throughout the domain (Fig. 3), while the negative JAS bias may be due to excessive vertical mixing. Despite these discrepancies with the observed data, however, overall the evolution of the spatial patterns seasonal LSTs is reproduced by the model, with biases of relatively small magnitude.

Fig. 5
figure 5

Mean October–December (OND, a–b), January–March (JFM, c–d), April–June (AMJ, e–f) and July–September (JAS, c–d) lake surface temperature (LST, unit: °C) for the period 1996–2008 from ARC-lake observations (a, c, e, g) and RegCM4_R10L (b, d, f, h)

Full size image
Fig. 6
figure 6

Mean October–December (OND, a), January–March (JFM, b), April–June (AMJ, c) and July–September (JAS, d) lake surface temperature (LST, unit: °C) bias for the period 1996–2008 with respect to ARC-lake observations

Full size image

The biases of Fig. 6 are reflected in the annual cycle of lake temperature averaged over the entire lake area (Fig. 7). During the summer rainy season, November through February, the model reproduces the observed values, while during the rest of the year we can see an offset of the phase of the seasonal cycle, with minimum temperature found in July for the observations and August for the model. This can also probably be ascribed to the lack of horizontal mixing in the lake model. The amplitude of the lake wide surface temperature cycle shows however a good agreement with observations, with a high correlation (~0.96).

Fig. 7
figure 7

Mean annual cycle of lake surface temperature (unit: °C) averaged over the lake region (14°S–09°S and 33.75°E–35.5°E) for the period 1996–2008 from ARC-lake observations and RegCM4_R10L

Full size image

In summary, despite the lack of horizontal circulations, which likely contribute to the spatial distribution of the biases observed in Figs. 6 and 7, the model appears to show lake surface temperatures which are mostly in line with observations, particularly in the warm and wet season, which is the focus of this paper.

3.2.2 Influence of Lake Malawi on regional climate

We now turn our attention on the comparison between simulations with Lake Malawi included (LAKE, i.e. Control; RegCM4_R10L) and removed (NOLAKE; RegCM4_R10NL) in order to isolate the lake effects on the surrounding climate. Figure 8 displays the mean OND, JFM, AMJ and JAS difference in mean surface air temperature (upper panels; a–d) and surface sensible heat flux (SHF; bottom panels; e–h) between RegCM4_R10L and RegCM4_R10NL (LAKE–NOLAKE hereafter), while Fig. 9 displays the corresponding changes in surface latent heat flux (LHF, a–d) and precipitation (e–h). Note here, that dots indicate regions where the differences are statistically significant at the 90% confidence level based on a two-tailed student’s t-test. The corresponding averages over the lake area (white box in Fig. 1b) grid point only are summarized in Tables 3 and 4. The lake produces a statistically significant local cooling effect of up to about 0.6 °C in JFM and 0.80 °C in AMJ, while exceeding 1.30 °C in OND (Fig. 8a–d; Table 3). This cooling spreads throughout the region in all seasons, but with magnitudes generally less than 0.5 °C away from the lake’s vicinity. The changes of SHF have a similar spatial structure as the temperature field, with a substantial decrease over the lake ranging from 16.40 to 27.84 W/m2 (Table 3). Over most areas of SHF decrease (increase), the 2 m-temperature is found to decrease (increase), mostly due to increased evaporative cooling (see Fig. 10). In fact, a statistically significant increase of LHF occurs over the lake (23.80–42.07 W/m2; Table 3), while the increase in surface LHF over the surrounding land regions is significant only in small areas.

Fig. 8
figure 8

Mean October–December (OND: a, d), January–March (JFM: b, f), April–June (AMJ: c, g) and July–September (JAS: d, h) 2 m-temperature (top panel, unit: °C) and surface Sensible Heat Flux (bottom panel, unit: W/m2) differences between LAKE and NOLAKE (RegCM4_R10L-RegCM4_R10NL). The dotted areas denote differences which are statistically significant at a significance level of 90% of student’s t-test value

Full size image
Fig. 9
figure 9

Mean October–December (OND: a, d), January–March (JFM: b, f), April–June (AMJ: c, g) and July–September (JAS: d, h) surface Latent Heat Flux (top panel, unit: W/m2) and precipitation differences (bottom panel, unit: mm/day) between LAKE and NOLAKE (RegCM4_R10L-RegCM4_R10NL). The dotted areas denote differences which are statistically significant at a significance level of 90% of student’s t-test value

Full size image
Table 3 Simulated lake contribution (LAKE minus NOLAKE) seasonal mean 2 m-temperature (units: °C), surface sensible heat flux (SHF, units: W m−2) and surface latent heat flux (LHF, units: W m−2) for the period 1994–2008 over the lake area (14°S–09°S and 33.75°E–35.5°E)
Full size table
Table 4 Simulated Lake contribution (LAKE minus NOLAKE) seasonal mean precipitation in absolute value (units: mm day−1) and relative value (unit: %) for the period 1994–2008 over the lake area (14°S–09°S and 33.75°E–35.5°E)
Full size table
Fig. 10
figure 10

Monthly mean annual cycle differences between LAKE and NOLAKE (RegCM4_R10L-RegCM4_R10NL) of precipitation (unit: mm/day) and evaporation (unit: mm/day), averaged over the Lake Malawi region (14°S–09°S and 33.75°E–35.5°E)

Full size image

Of particular interest is the precipitation response to the lake-effect. In all seasons, and especially the rainy ones, we find a dipolar structure over the lake, with an enhancement of rainfall over the northern lake areas, by about 0.5–1.5 mm/day, and a more mixed signal over the southern ones. In addition, we find an area of suppressed precipitation by the lake over northern Zambia, close to the northern boundary of the analysis area (which is not however the northern boundary of the domain, see Fig. 1). The largest rises in precipitation occur roughly in the same areas where the largest 2 m surface temperature declines are found, and as expected the lake precipitation effect is much smaller during the two dry seasons (Fig. 9g, h). Areas of both positive and negative precipitation differences are then found throughout the land regions of the analysis area, but they are not always statistically significant, particularly during the dry season. The increase in precipitation over the Lake Malawi region is consistent with the LHF increase in the region (Fig. 9a–d). The seasonal impact of the lake on precipitation is presented in Table 4, which shows that the lake’s contribution to the precipitation falling over the lake area is positive in each season, with large contributions in OND (23%) and JFM (19%). Note that the large magnitude of relative precipitation change (in %) found for both AMJ and JAS seasons (Table 4) is mostly a reflection of the small amounts of precipitation in the no-lake experiment.

We also note in Fig. 9 the presence of some small lakes (Lake Mweru, Lake Rukwa and southern part of Lake Tanganyika) near the upper boundary of the domain which actually cause a decrease of precipitation. This is an artifact of how the LST are obtained over these small lake areas, where we do not run the coupled lake model. Following the standard RegCM4 procedure for inland water bodies, the LST there are interpolated from the closest ocean grid points, which are characterized by relatively cold values (see also Fig. 8). These excessively cold LST inhibit convection, causing a decrease in precipitation over and around the lakes. This also results in a decrease of soil evaporation and increase in sensible heat flux over the land areas around the lakes. The comparison of the precipitation effect of these small lakes with that induced by Lake Malawi, which is of opposite sign, thus highlights the importance of the lake–atmosphere coupling.

Figure 10 presents the annual cycle of the difference in monthly mean precipitation and evaporation (LAKE–NOLAKE; i.e. RegCM4_R10L-RegCM4_R10NL) averaged across the lake area (white box in Fig. 1b) grid points. Both precipitation and evaporation increase throughout the whole year in response to the lake forcing, however with distinctly different annual cycles. The lake effect precipitation follows the precipitation cycle and peaks in January, while the evaporation increase peaks in the dry months, May through October, when solar insolation is maximum because of lack of clouds and the overlying air is drier. Furthermore, over the lake, the water vapor mixing ratio (not shown) increases by up to 2 g/kg, due to enhanced evaporation. This increase in evaporation, however, is not sufficient to cause detectable rain over the lake during the dry season, when large scale dynamics dominates (e.g. Anyah and Semazzi 2004; Anyah et al. 2006; Notaro et al. 2013; Nicholls and Toumi 2014; Thiery et al. 2015).

Two competing effects contribute to the lake-effect precipitation. On the one hand the surface cooling by the lake induces a local increase in sea level pressure (see supplementary Figure S2a, b) and thus a thermal stabilization, while on the other hand the increased evaporation enhances the moisture content of the lower levels and thus the potential for convection associated with increased equivalent potential temperature. Figure 11 shows that the lake causes an increase in convective precipitation, i.e. that the increase in convective instability due to higher moisture amounts prevails (see also Supplementary Figure S3a, b). In turn, increased local convection yields enhanced low-level convergence over the lake and surrounding regions (Fig. 12), which further enhances the lake-effect precipitation.

Fig. 11
figure 11

Mean October–December (OND: left panel) and January–March (JFM: right panel) convective precipitation differences (unit: mm/day) between LAKE and NOLAKE (RegCM4_R10L-RegCM4_R10NL). The dotted areas denote differences which are statistically significant at a significance level of 90% of student’s t-test value

Full size image
Fig. 12
figure 12

Mean October–December (DJF: left panel) and January–March (JFM: right panel) vertically integrated moisture flux from the surface (1000 hPa) to 850 hPa (streamlines, unit: m/s) and its convergence (shaded, unit: mm/day) differences between LAKE and NOLAKE (RegCM4_R10L-RegCM4_R10NL)

Full size image

Figure 13 shows, as a function of latitude, the vertical cross-section of differences in mean vertical velocity (omega) averaged from 33.75°E to 35.5°E for the OND and JFM seasons (Fig. 13a, b), along with the corresponding precipitation differences (Fig. 13c, d). The vertical velocity field shows that there is a net rising motion extending from the surface to the middle troposphere (900–600 hPa) at 14°S–10°S during both seasons, which is in agreement with the atmospheric instability increases above the lake. This anomalous rising motion indicates that convection is enhanced as a result of the presence of the lake and feedbacks with increased low-level convergence.

Fig. 13
figure 13

Vertical cross-section as a function of latitude of October–December (OND; left panels) and January–March (JFM; right panels) vertical motion (unit: mm/s, top panels) and precipitation (unit: mm/day) differences between LAKE and NOLAKE for the period 1994–2008. The dotted areas denote differences which are statistically significant at a significance level of 90% of student’s t-test value. The black shaded areas indicate grid points under the topography

Full size image

4 Summary and conclusions

In this paper we present a study of the effects of Lake Malawi on the climate of the surrounding region using a double nested high resolution (grid spacing of 10 km) regional climate model coupled to a one-dimensional spatially distributed lake model run at each grid point of the lake. The possible mechanisms on how lake–atmosphere interactions affect the south-eastern Africa regional climate, particularly Malawian regions are also investigated. We first show that the regional climate model is able to reproduce the basic surface climate characteristics of the region, with biases in line with previous applications of this model. Comparison with satellite observations shows that the coupled lake model realistically reproduces the geographical pattern distribution of lake surface temperature, although it overestimates the observed east–west lake surface temperature gradient and it shifts the minimum in average lake temperature by about one month. Both these shortcomings can be likely attributed to the lack of horizontal mixing in the model (Martynov et al. 2012; Notaro et al. 2013; Turuncoglu et al. 2013). On the other hand, despite the absence of horizontal dynamics, the lake surface temperature biases are small in the lake interior, and during the wet season the average lake temperatures are reproduced remarkably well.

Lake effects are assessed by inter-comparing simulations with the lake and with lake points replaced by grass surfaces. The presence of Lake Malawi induces a decrease of surface temperature and an enhancement of low level water vapor mixing ratio caused by an increase of evaporation. This causes an increase in convective instability, rising motions and moisture convergence, which finally result in enhanced precipitation. Overall, the annual average precipitation amount over the lake increases by approximately 15% due to the lake, with most of this increase occurring over the northern portion of the lake and highlighting the non-uniform response to the presence of the lake.

Thus our study shows that the presence of Lake Malawi has a significant effect on the local climate characteristics, although in our experiments this effect is not felt strongly in regions far from the lake. Our coupled atmosphere–lake modeling system appears to reproduce reasonably well the evolution of the lake surface temperatures, despite the use of a simple one dimensional distributed model. Therefore we plan to apply this system to the simulation of future climate over the region under forcing from increased greenhouse gas (GHG) concentrations.