Today there is very little annual variability in temperature at Lake Victoria (Nicholson, 1996). Precipitation, however, varies considerably and is controlled by the location of the Intertropical Convergence Zone (ITCZ), which crosses Lake Victoria twice a year, creating rainy seasons in March and in October (Nicholson, 1996). Precipitation is also controlled by the movement of the Congo Air Boundary (CAB), the strength of the Indian and Atlantic monsoons, and topography (Berke et al., 2012). MAP across the catchment varies from 1,400 to 1,800 mm year⁻¹ (Yin and Nicholson, 1998; Sutcliffe and Parks, 1999). Because MAP is almost equal to average evaporation (1,460 mm year⁻¹), with local precipitation derived primarily from the lake itself, lake level responds directly to changes in rainfall (Broecker et al., 1998; Yin and Nicholson, 1998; Milly, 1999; Bootsma and Hecky, 2003). The historic vegetation from the Lake Victoria region is evergreen bushland, thicket, and forest habitats (Andrews, 1973; White, 1983). In the Lambwe Valley, between Rusinga Island and Karungu, the large mammalian herbivores historically included bushbuck (Tragelaphus scriptus), African buffalo (Syncerus caffer), Bohor reedbuck (Redunca redunca), waterbuck (Kobus ellipsiprymnus), duiker (Sylvicapra grimmia), topi (Damaliscus lunatus), impala (Aepyceros melampus), hartebeest (Alcelaphus buselaphus), roan antelope (Hippotragus equinus), oribi (Ourebia ourebi), giraffe (Giraffa camelopardalis), black rhinoceros (Diceros bicornis), bushpig (Potamochoerus larvatus), and elephant (Loxodonta africana) (Allsopp and Baldry, 1972; Kimanzi, 2011).

The Late Pleistocene geology, fossils, and MSA artifacts from Rusinga and Mfangano Islands have been the focus of research since 2009 (Tryon et al., 2010, 2012, 2014; Faith et al., 2011, 2012, 2014; Van Plantinga, 2011; Jenkins et al., 2017; Blegen et al., in press). More recently, this research has expanded to include the deposits around the region known as Karungu ~40 km to the south near the town of Sori (Figure 1; Beverly et al., 2015a,b; Blegen et al., 2015; Faith et al., 2015). Abundant fossils and MSA artifacts have long been reported from Karungu and Rusinga and Mfangano Islands (Pickford, 1984; Owen, unpublished manuscript), but until recently only limited geological mapping and reconstructions of the paleoclimate or paleoenvironment have been conducted. Karungu and Rusinga and Mfangano Islands were originally mapped by Oswald (1914), Kent (1942), Van Couvering (1972), Whitworth (1961), and Pickford (1984) and have been more intensively mapped by Beverly et al. (2015a,b), Blegen et al. (2015, 2016), and Tryon et al. (2010, 2012, 2014).

Artifacts and fossils were surface collected from all units and excavations have been conducted at Wakondo, Nyamita, and Aringo (Figures 1, 2). Locations of in situ fossils and MSA artifacts are indicated in the stratigraphy (Figure 2). Artifacts from excavations at Nyamita and Wakondo on Rusinga Island in particular appear associated with formerly stable land surfaces, including weakly developed paleosols (Jenkins et al., 2017; Blegen et al., in press). Artifact density and the presence of non-local raw material types from excavated areas and surface collections are generally consistent with production by highly mobile populations of foragers (Tryon et al., 2010; Faith et al., 2015; Garrett et al., 2015; Blegen et al., in press).

Tephra deposits can be correlated between sites on the basis of geochemical compositional similarity and have been dated by multiple radiometric methods (Tryon et al., 2010; Beverly et al., 2015b; Blegen et al., 2015, 2016, in press). Four named tuffs are used to correlate the sections between Rusinga Island and Karungu: Wakondo, Nyamita, Nyamsingula, and Menengai. The Wakondo Tuff is >68 ± 5 ka based on overlying optically stimulated luminescence (OSL) dates and <94.0 ± 3.3 based on the U-series age of the underlying tufa deposits (Beverly et al., 2015b; Blegen et al., 2015). Additional data on the tufas are presented in Beverly et al. (2015b). The Nyamita Tuff is constrained by OSL dates above and below the tuff that indicate deposition at ~49 ka (Blegen et al., 2015). The Nyamita Tuff is thick, distinct, and found at almost all Pleistocene sites in this region, making it an ideal stratigraphic datum (Figure 2). The Menengai Tuff has been ⁴⁰Ar/³⁹Ar dated to 35.62 ± 0.26 ka at its eruptive source, Menengai Caldera in the Central Kenya Rift (Blegen et al., 2016), and the Nyamsingula Tuff is dated to between 36 and 49 ka based on the stratigraphic relationships with the underlying Nyamita Tuff and overlying Menengai Tuff (see Blegen et al., 2015, 2016, in press for an extensive discussion of tephra correlations and dates).

Bulk geochemistry from 7 paleosols previously published in Beverly et al. (2015a) and new data from 6 additional paleosols were used to estimate MAP using the following proxies: CALMAG (Nordt and Driese, 2010) for Vertisols only, and CIA-K (Sheldon et al., 2002) and PPM_1.0 (Stinchcomb et al., 2016) for all other paleosol types. CALMAG was only applied to those paleosols identified as paleo-Vertisols based on these vertic features: pedogenic slickensides, wedge peds, rhizocretions, and micromorphological features indicative of shrink-swell processes (i.e., b-fabrics and septarian and circumgranular cracks). The CALMAG estimates for MAP range from 557 to 911 mm year⁻¹ with an average of 750 mm year⁻¹ (Table 3). The CIA-K estimates for MAP range from 552 to 1,000 mm year⁻¹ with an average of 800 mm year⁻¹. The PPM_1.0 estimates for MAP range from 747 to 1,283 mm year⁻¹ with an average of 1,010 mm year⁻¹. The CIA-K and PPM_1.0 proxies were applicable for all identified paleosols because none are identified as paleo-Aridisols or Oxisols, they have no significant topographic relief, and they have no evidence indicative of a highly waterlogged soil. The Nyamita Tuff is used as a division between the upper and lower paleosols because it is dated to ~49 ka (OSL), and because it is lithologically distinct and a visible stratigraphic marker on the landscape that is present at most sites.

The stratigraphy at Karungu and Rusinga and Mfangano Islands (Figure 2) is dominated by paleosols that, based on various dating methods (Blegen et al., 2015, 2016) and paleosol features, likely represents most of the ~58 kyr interval between 94 and 36 ka. Paleosol 3 is dated to between 94 and 49 ka and is identified across the region at Kisaaka, Aringo, Aoch Nyasaya, Wakondo, and Nyamita. The very well-developed pedogenic slickensides and pedogenic carbonates indicate that this paleo-Vertisol was subjected to prolonged pedogenesis and was a stable land surface regionally for a long period of time. The paleo-Inceptisols above the Nyamita Tuff are dated to between 49 and 36 ka and pedogenic features identified in these paleosols support the interpretation that they were deposited over a shorter duration of time. The paleo-Inceptisols from Kisaaka and Aringo contain up to 3–5% illuviated clay. Using the age-estimate model for clay coating accumulation by Ufnar (2007), 3–5% illuviated clay indicates that pedogenesis occurred for ~3–7 kyr per paleosol (Beverly et al., 2015a). This may be an overestimation for an East African monsoonal climate, which may weather more quickly than subtropical soils from southeastern Mississippi used to develop this chronofunction (Ufnar, 2007), but it does indicate a period of weathering long enough for the soil to come into equilibrium with climate. The three paleosols preserved in the late Pleistocene deposits along the northeastern margin of Lake Victoria are spread across an area of ~1,000 km² along the northeastern margins of Lake Victoria and thus can be inferred to represent regional climate. Based on their distribution and our age estimates for pedogenesis, these three paleosols likely represent a time-averaged, but relatively complete, record of ~58 kyr of climate in the Lake Victoria region between 94 and 36 ka.

MAP estimates range from 557 to 1,283 with an average of 750 ± 108, 800 ± 182, and 1,010 ± 228 mm year⁻¹ for the CALMAG, CIA-K, and PPM_1.0 proxies, respectively (Table 3). Although there is variability between each proxy, they are within standard error margins, and all estimates are significantly less than modern MAP in the Lake Victoria basin. Calculating precipitation over the Lake Victoria Basin is difficult due to the lack of consistent station coverage and interruptions in recording periods (Yin and Nicholson, 1998), and thus reported precipitation varies wildly. The most commonly reported MAP ranges from 1,400 to 1,800 mm year⁻¹ (Yin and Nicholson, 1998; Sutcliffe and Parks, 1999). Assuming a MAP of 1,400 mm year⁻¹, this suggests that average precipitation during the Late Pleistocene was 54, 57, or 72% of modern precipitation (CALMAG, CIA-K, and PPM_1.0 proxies, respectively). Thus, we suggest that the Lake Victoria region was likely significantly drier for this entire interval, but acknowledge that sub-millennial scale trends may not be resolvable. This agrees with isotopic data from fossil fauna, which document a >20% reduction in MAP relative to modern conditions (Garrett et al., 2015).

Given our evidence for drier conditions and previous water budget modeling for the lake, we hypothesize that Lake Victoria had a much smaller surface area than modern from 94 to 36 ka. Previous water budget models (Broecker et al., 1998; Milly, 1999) indicate that Lake Victoria would be reduced to <20% of current surface area with a MAP between 557 and 1,283 mm year⁻¹. A Late Pleistocene dessication surface of unknown duration is also identified in Lake Victoria at ~80 ka using sedimentation rates and seismic stratigraphy (Stager and Johnson, 2008), which supports our interpretation. Additionally, this dry interval between 94 and 36 ka in the Lake Victoria region also partially overlaps with the megadrought identified in Lakes Malawi, Tanganyika, and Bosumtwi between 135 and 70 ka (Scholz et al., 2007) with a revised age-model indicating Lake Malawi was significantly lower between 146 and 115 ka and again between 107 and 74 ka and has remained deep since ~70 ka (Lyons et al., 2015; Ivory et al., 2016).

Zonal shifts in the ITCZ are often used to explain precipitation changes in tropical Africa; however, it is unlikely that the ITCZ could be shifted so far south to prevent its bi-annual crossing of Lake Victoria. Widespread climatic variability likely peaked between 145 and 60 ka (Laskar et al., 2004; Scholz et al., 2007) due to eccentricity-enhanced precession, and climate modeling (Clement et al., 2004) indicates that when eccentricity is high, precessional forcing is greater at low latitudes compared to high latitudes. This orbital mechanism explains the aridity for Lakes Malawi, Tanganyika, Botsumtwi, and Victoria during the megadrought, but after 70 ka, all these lakes began to refill, with the exception of Lake Victoria (based on MAP estimates and previous water budget modeling). The refilling of these lakes is attributed to stable precipitation conditions as eccentricity decreased (Scholz et al., 2007). The paleosol-based paleoprecipitation reconstruction indicates that Lake Victoria region remained dry until 36 ka, which is similar to the cooler and drier conditions in the Horn of Africa between 50 and 75 ka (Tierney et al., 2017) and to the record of Lake Naivasha (Trauth et al., 2003), ~250 km to the east. Because eccentricity decreased from 70 to 36 ka, the cause for extended aridity from ~70 to 36 ka in the Lake Victoria region must be due to factors other than eccentricity-enhanced precession. Lake Naivasha level was low-to-intermediate between 105 and 60 ka and dry between 60 and ~35 ka (Trauth et al., 2003). Low levels in Lake Naivasha have been attributed to low sea surface temperatures (SSTs) in the Indian Ocean that weakened the monsoon (Bard et al., 1997; Trauth et al., 2003), and the desiccation of Lake Victoria between 15 and 18 ka has similarly been attributed to low SSTs (Stager et al., 2011). This suggests that Lake Victoria may have been antiphased with the southern tropics between 70 and 36 ka as has been demonstrated in other records (Partridge et al., 1997; Gasse, 2000; Trauth et al., 2003; Castañeda et al., 2007).

The tephrostratigraphy and our detailed analyses of the paleosols between tephra allows for the reconstruction of the landscape in the northeastern Lake Victoria Basin as a series of paleo-CZs. Below the Nyamita Tuff, the oldest paleo-CZ (P3) is dominated by well-developed paleo-Vertisols. The parallel, cross, and granostriated b-fabrics in the micromorphology and well-developed slickensides all indicate rainfall seasonality. Bulk geochemical pedotransfer functions for reconstructing Vertisol properties such as salinity, sodicity, pH, and cation exchange capacity were applied on the Kisaaka Vertisols and all indicate fertile soil with saline-sodic conditions that would have affected plant size or been limited to salt tolerant species (Beverly et al., 2015a). In modern African savanna ecosystems, vegetation is greatly affected by both vegetation and soil texture (sand vs. clay), and a simple abiotic model created by Johnson and Tothill (1985) can be used to predict the type of vegetation in paleosols (Wynn, 2000; Beverly et al., 2015a). Based on MAP estimates and a dominantly clay grain size, this lower paleo-CZ was a seasonally dry but fertile soil affected by high salinity and sodicity. Vegetation would have ranged from savanna grassland to savanna parkland (Figure 3A).

In the paleo-CZ overlying the Nyamita Tuff (P1 and P2), the pedotransfer functions for reconstructing soil properties were not applied because these paleosols were identified as paleo-Inceptisols, and this proxy is only applicable to paleo-Vertisols. However, the abundance of illuviated clay (3–5%) indicates seasonal precipitation where moderate rainfall falls on a dry soil. Similar ranges of MAP, but higher amounts of volcaniclastic or tuffaceous material in the upper paleo-CZ suggest a seasonally dry, savanna parkland to savanna woodland similar to that currently found in the broader Serengeti Ecosystem (Figure 3A).

These reconstructions of the upper and lower paleo-CZ using paleosol features, grain size, and bulk geochemical proxies for MAP also agree with vegetation reconstructions using pedogenic carbonates, soil organic matter, and tooth enamel from Rusinga and Mfangano Islands, which indicate the local occurrence of a woodland to grassy woodland surrounded by an expansive C₄ grassland (Faith et al., 2015; Garrett et al., 2015; Tryon et al., 2016). The fossils from this region are dominated by a diverse population of grazing herbivores and arid-adapted ungulates such as oryx (Oryx beisa) and Grevy's zebra (Equus grevyi), which are found far outside their modern range of arid eastern and northeastern Africa (Faith et al., 2015).

All of these data indicate that an expansion of arid-adapted fauna and Serengeti-like grasslands likely occurred during the Late Pleistocene with the Serengeti as a reasonable modern analog (Figure 3B). This contrasts with the historic evergreen bushland, thicket, and forest habitats in the Lake Victoria region (Figure 3C; White, 1983). Vegetation models (Cowling et al., 2008) and phylogeographic observations derived from savannah ungulates (Lorenzen et al., 2012) predict that more open vegetation would have facilitated dispersals of mammal species, including humans, across equatorial east Africa, and the reduction in size of Lake Victoria, which would have removed a significant geographic barrier, is likely to have further enhanced dispersal potential. Importantly, many of the mammalian taxa that co-occur in fossil assemblages from Rusinga, Mfangano, and Karungu are allopatric today and historically, forming what Faith et al. (2015, 2016) refer to as “non-analog supercommunities,” comprised of taxa for which Lake Victoria currently serves as a major barrier. This includes, for example, the co-occurrence of white rhinoceros (Ceratotherium simum; historically northwest of the lake), Grevy's zebra (northeast of the lake), and southern reedbuck (Redunca arundinum: south of the lake). From this perspective, the absence of the lake and expansion of more open vegetation likely facilitated range expansions for these taxa. The fossil fauna co-occur with Middle Stone Age artifacts. The implication is that MSA-equipped humans also dispersed into the center of the basin with lake-level retreat. Reducing the effects of Lake Victoria as a biogeographic barrier for the movement of fauna would have encouraged, although not dictated, the dispersal of humans along the Nilotic corridor as an avenue for the movement of modern humans out of and within Africa between 94 and 36 ka (Tryon et al., 2016).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.