Elemental geochemistry of bulk paleosol samples was characterized using two methods: (1) untreated, whole-sample assays using fused-bead ICP-AES and ICP-MS analysis for major and trace elements, respectively, at ALS Global (alsglobal.com); and (2) whole-sample assays of decarbonated aliquots of the same samples using X-ray fluorescence (XRF) for major element analysis at Baylor University. Samples sent to ALS Global were also analyzed for loss on ignition (LOI) at 1,000°C. Samples analyzed in-house at Baylor University were powdered and homogenized, decarbonated using 10% HCl, and neutralized using repeated rinsing with de-ionized water. The decarbonated samples were then re-ground, homogenized, fused into glass beads, and analyzed for major element concentrations using a Rigaku ZSK Primus II X-ray fluorescence analyzer. Total and organic carbon content were measured on aliquots from each horizon using an element analyzer (EA) before and after in situ decarbonation, respectively. Inorganic carbon content was calculated as the difference between the two measurements.

Thin sections of oriented samples from each horizon were prepared by Spectrum Petrographics (http://www.petrography.com) and characterized using a combination of methods from Brewer (1964) and Bullock et al. (1985). Carbonates were analyzed in thin section to identify mineral phases. Nodules and rhizoliths were micro-drilled on paired billets and hand samples after cross-comparison with thin sections. Carbonate powders were analyzed for inorganic carbon (δ¹³C) and oxygen (δ¹⁸O) stable isotope composition at Baylor University using a Thermo-Electron dual-inlet gas-source stable isotope ratio mass spectrometer. The delta notation is used in reporting stable isotope compositions (e.g., δ¹⁸O = [R_sample / R_standard − 1] x 1000), where R is the ratio of the heavy to light isotope (e.g., ¹⁸O/¹⁶O). All isotope values are reported relative to the Vienna Pee Dee Belemnite (VPDB) standard with analytical precision within 0.03%0 and 0.1%0 for δ¹³C and δ¹⁸O, respectively.

Clay mineralogy of parent material and weathered soil horizons was analyzed using oriented clay slides on a Siemens D5000 X-ray diffractometer that uses Cu Kα radiation and operates at 40 Kv and 30 mA. The clay-sized fraction (<2 μm equivalent diameter) was separated via centrifugation following the methods of Poppe et al. (2001) and oriented clay slides were prepared following the methods of Moore and Reynolds (1997). To differentiate possible commonly-occurring clay minerals, four treatments were performed on clay slides (Moore and Reynolds, 1997): (1) samples were Mg²⁺-saturated using a Mg salt solution, (2) Mg²⁺-saturated samples were glycolated overnight in an airtight chamber, (3) duplicate samples were K⁺-saturated using KCl, and (4) a subset of K⁺-saturated samples were heated for at least 2 h at 550°C. All clay samples were analyzed from 2 to 30° 2Θ with a step size of 0.02° 2Θ and dwell time of 1 s. Clay minerals were identified following the methods of Moore and Reynolds (1997).

Bulk geochemistry of paleosol horizons was compared using mass-balance calculations (Brimhall and Dietrich, 1987; Chadwick et al., 1990; Brimhall et al., 1991), without including bulk density due to the effects of compaction on paleosol samples (Driese et al., 2000, 2016; Driese, 2004). Translocation (τ, or “tau”) of elements was calculated relative to parent material, where: τ = {[(C_{j, w} /C_{j, p})]/[(C_{i, w} /C_{i, p})] − 1} × 100, in which the concentrations (C) of elements (j) in weathered horizons (w) are assessed relative to immobile elements (i) in the parent material (p). Profiles of tau values therefore can be used to show gains (positive values) or losses (negative values) as a result of pedogenic processes or sedimentary additions, assessed as a function of starting parent material composition. The composition of C and BC horizons was averaged to account for minor biases due to grain size-dependent compositional changes and/or incipient weathering of individual C or BC horizons.

The elemental composition of paleosol B horizon samples was used to estimate mean annual precipitation (MAP) and temperature (MAT) using a suite of pedotransfer functions (Sheldon et al., 2002; Gallagher and Sheldon, 2013; Table 1). The chemical index of alteration minus potassium (CIA-K), which tracks feldspar hydrolysis and base cation leaching (Sheldon et al., 2002) was used to estimate MAP. The salinization index (NaK; Sheldon et al., 2002) and the paleosol weathering index (PWI; Gallagher and Sheldon, 2013), which track thermodynamic controls on silicate dissolution, were used to estimate MAT. The paleosol-paleoclimate model (PPM_1.0) (Stinchcomb et al., 2016) was also used to co-predict MAP and MAT. PPM_1.0 uses eleven elemental oxides as input variables (Fe₂O₃, Al₂O₃, SiO₂, TiO₂, ZrO₂, CaO, MgO, K₂O, Na₂O, MnO, and P₂O₅) to simultaneously model both MAP and MAT using a thin-plate spline drawn through sample scores generated by a partial least squares regression analysis of modern soil B horizon samples. PPM_1.0 calculations were run in SAS statistical software (SAS Institute, Inc., v. 9.2.). Climate results are discussed in terms of climate zones after (Bull, 1991), which include MAP zones of arid (<250 mm), semiarid (250–500 mm), subhumid (500–1,000 mm), and humid (1,000–2,000 mm), and MAT zones of frigid (0–8°C), mesic (8–15°C), thermic (15–22.5°C), and hyperthermic (>22°C).

We frame our vegetation reconstruction in terms of photosynthetic pathway, which is possible as a result of two broad photosynthetic mechanisms that differentially fractionate atmospheric CO₂, resulting in plant-respired soil CO₂ with differing carbon isotope signatures (δ¹³C). Modern plants that use the Calvin cycle (C₃ pathway) include most trees, shrubs, and cool-season grasses, and have δ¹³C signatures between −20 to −33%0, with a mean of −26.7%0 (O'Leary, 1988; Cerling et al., 1997). Water stress is the primary factor that drives the enrichment of C₃ plant δ¹³C signatures above ca. −24 to −25%0 (Ehleringer and Cooper, 1988). The Hatch-Slack (C₄) pathway, which is utilized mostly by modern tropical and warm-season grasses and sedges, is a series of anatomical and biochemical adaptations that allows for closure of stomata without photorespiration (Sage, 2004). Modern C₄ plants have δ¹³C values between −9 to −11%0, with a mean value of −12.5%0 (O'Leary, 1988; Cerling et al., 1997). Carbon isotope values of vegetation are recorded in the isotopic composition of pedogenic carbonates that form in equilibrium with plant-respired soil CO₂ at depth in soils of moderate to high productivity (Cerling et al., 1989).

Endmember δ¹³C compositions for C₃, water-stressed C₃ (wsC₃), and C₄ vegetation were modeled using nominal fractionation factors for each photosynthetic pathway/plant type and corrected for an average atmospheric CO₂ composition of −5.8%0 from benthic foraminifera for the interval 17.07–17.5 Ma (Tipple et al., 2010). Fractionation factors between plants and CO₂ (ε_plant−CO2) reported by Passey et al. (2002) were used, where ε_C3−CO2 = −19.6%0, ε_wsC3−CO2 = −16.7%0, and ε_C4−CO2 = −4.7%0. A net fractionation of +15%0 between plant-respired soil CO₂ and pedogenic carbonate (Fox et al., 2012, after Cerling and Quade, 1993) yielded endmember compositions of −10.4%0, −7.5%0, and +4.5%0 for carbonates forming under C₃, wsC₃, and C₄ vegetation cover, respectively.

NG15 is bounded by trough cross-stratified fluvial channel sandstones (NG14 and NG16). Across ca. 150 m of lateral outcrop, NG15 varies from thinly-bedded, tan fluvial sandstones to red, calcareous sandy siltstones. The NG15-A and B profiles were described where NG15 occurs as sandy siltstone. A total of five superposed paleosols were identified, most of which were underlain by an alluvial parent material (C or BC) horizon (Figure 3). The parent material horizons were more coarse-grained than overlying subsurface horizons, and in some cases retained primary horizontal stratification. Bedding contacts at the base of parent material horizons are inclined from NG15-A to NG15-B with upwards of 10 cm relief.

Detailed descriptions of paleosol macro- and micromorphology are reported in Table A1, with key observations presented in Figure 3. Horizonation is largely similar between profiles NG15-A and B. The uppermost paleosol is truncated from above and grades laterally from a C1-C2 to a Bw-BC horizon sequence between the profiles, respectively. The second paleosol consists of a Btk-BC horizon sequence with lenses of sand and reworked, hard, gray carbonate nodules variably coated in MnO₂ or Fe₂O₃. The Btk horizon has well-developed prismatic ped structure, clay-filled root traces, clay coatings on peds, and vertically-oriented hairline root traces (Figure 4). With the exception of a thin, truncated Bk horizon that underlies the second paleosol in NG15-A, the rest of the succession consists of alternating couplets of Bw-C horizons, each of which are bounded by erosional contacts. The Bw horizons in the lower portion of the profile have root traces that tend to be <2 mm wide and occur as drab haloes, calcareous rhizoliths, or clay-filled pores.

Pedality—a measure of ped structure development—and reaction to HCl were characterized for profile NG15-B using USDA-NRCS methods (Figure 3; Schoeneberger, 2002). The B horizons tend to have more well-developed pedality than parent material horizons. Effervescence was very strong in both B horizons and parent material horizons.

Thin sections from all horizons of the uppermost 4 paleosols in NG15-B were studied for a variety of pedogenic and sedimentary properties (Figure 5). Surveys of clay, carbonate, and Fe features were converted to qualitative indices with values corresponding to features being (0) absent, (1) few, (2) common, (3) abundant, or (4) omnipresent (Figure 3). Pedogenic clay fabric includes grain coatings (granostriated), pore linings (porostriated), and matrix impregnations (argillans). The clay index closely tracks pedality and is highest in B horizons, particularly the 2Btk horizon.

Carbonates exhibit a range of morphologies in profile NG15-B. Pedogenic micrite is the dominant phase in the 2Btk horizon, though minor amounts of sparite coat septarian shrinkage cracks in some cases. Hard, gray carbonate nodules tend to be micritic with minor quantities of sparite in septarian cracks. Diagenetic calcite spar is pervasive in parent material horizons and cross-cuts pedogenic features, such as fecal pellets and clay fabric. Sparite crystals are nucleated on detrital silicate grains and pedogenic features, with growth axes radiating displacively into the paleosol matrix. The carbonate index includes both pedogenic and diagenetic phases, and therefore tracks increases in pedogenic carbonate content in the Btk horizon, as well as diagenetic sparite that occurs in C horizons.

Fe features primarily occur as concretions in B horizons. Fe concretions are nucleated on weathered detrital grains and microvertebrate bone chips, which typically also have clay coatings (Figures 5L,M). The Fe index is highest in B horizons and closely tracks the clay index and pedality.

Microcharcoal (10–500 μm wide) was observed in most thin sections (Figures 5B,C,H,I). Grains of microcharcoal were differentiated from detrital silicates and heavy minerals by their opacity under plane- and cross-polarized light, tattered and irregular edges, and lack of cleavage planes. The concentration of microcharcoal grains was not estimated because, in many instances, grains occurred in a large range of sizes, likely as a result of fragmentation.

Elemental geochemistry, loss-on-ignition (LOI), and measured organic and inorganic carbon content for the NG15-B profile is reported in Table A2. Untreated samples were compared with samples that were decarbonated before elemental analysis (Figure 6).

Samples from the five uppermost horizons of NG15-B profile yielded similar X-ray diffraction results (Figure 8). Primary peaks on each diffractogram of K⁺-saturated samples occur at 10.15, 5.02, and 3.34 Å, which are associated with illite/mica 001, 002, and 003 peaks, respectively. Quartz can also contribute to the peak at 3.34 Å, but because a secondary peak at 4.26 Å is absent, quartz is either absent in the clay-sized fraction or contributing very little to the illite 003 peak. Peaks associated with smectite (17.5 Å) and kaolinite (7.18 and 5.38 Å) were not observed. The lack of change with Mg²⁺ saturation, glycolation, or heat treatments further supports an interpretation of exclusively illitic clay mineralogy. The calcite peak corresponding to a d-spacing of 3.03 Å is present in all samples, and is sharpest in the Btk horizon. Peaks occurring at 4.51 and 3.52 Å may indicate minor quantities of detrital palygorskite and kaolinite, respectively (e.g., Myers et al., 2011), though no secondary peaks associated with either mineral is present.

Two approaches were used to estimate MAP and MAT values: pedotransfer functions using weathering index proxies for MAP or MAT, and a multivariate spline model that uses 11 elemental oxides as input values to co-predict MAP and MAT (Figure 9). The CIA-K proxy for MAP yields arid to semiarid values (<500 mm) for untreated samples and mostly subhumid values (500–1,000 mm) for decarbonated samples. MAP values calculated using PPM_1.0 range from arid to humid for untreated samples, and, interestingly, range from arid to subhumid for treated samples—substantially lower than CIA-K predicted MAP. Decarbonation resulted in no detectable change in MAT values predicted by NaK and PWI, which were within the frigid (<8°C) and mesic zones (8–15°C), respectively. The PPM_1.0 model yielded mesic to thermic MAT values (ca. 8–18°C) for untreated samples and frigid to mesic values (ca. 5–14°C) for treated samples.

Pedogenic carbonate δ¹³C and δ¹⁸O results are presented in Figure 10; raw values are reported in Table A4. Values for δ¹³C span approximately 9%0, ranging from −7.3 ± 0.1 to −1.9 ± 0.2%0 VPDB, and the majority of samples are more positive than the calculated endmember of water-stressed C₃ vegetation. Previously analyzed carbonate nodules from units 19 and 20 by Driese et al. (2016) fall within the range of samples presented in this study, with the exception of a single nodule close to the C₃ endmember composition. Samples of lacustrine oncoid carbonate reported by Driese et al. (2016) fall in the middle of the isotope space and mirror the overall negative correlation between δ¹³C and δ¹⁸O. This suggests that lacustrine carbonates carried an isotope signal that sampled a mixture of local vegetation, similar to Pleistocene riverine tufa from Karungu and Rusinga Island (Beverly et al., 2015). Mean δ¹⁸O values reported in this study range from −4.6%0 to −8.0%0, which are similar to pedogenic carbonate values ranging between ca. −5%0 to −8%0 reported from middle Miocene sites at Fort Ternan and West Turkana in Kenya (Cerling, 1992), and thus likely did not form from highly evaporitic waters.

Measured δ¹³C values show limited correlation with carbonate morphology. Micritic, hairline rhizoliths from Bw horizons in the NG15-B profile have the highest δ¹³C values (up to −1.9%0). Mixed samples containing both rhizoliths and nodules are variable across the sample space. The composition of hard, gray nodules interpreted to be detrital is not unique; those occurring the 4Bw horizon have relatively high δ¹³C and low δ¹⁸O values, whereas those found in the 2Btk horizon have lower δ¹³C and higher δ¹⁸O values.

Using the calculated C₃ and C₄ endmembers of carbonate compositions, 21–57% of the standing biomass associated with NG15 was C₄ vegetation. If a water-stressed C₃ endmember is used, 2–47% C₄ biomass is required to account for the observed δ¹³C values, which on the upper end at least is a considerable and possibly surprising amount of C₄ biomass for an interval of this age.

Faunal composition of the excavated NG15 site is given in Table 2. NISP for the excavation grid at NG15 was 415, including surface collections within the grid boundaries. Mammals are the dominant taxon in the excavated grid, with 81.1% of the total abundance (83.6% when including the indeterminate Mammalia). Aquatic taxa are present, but not very common, and include crocodiles (1.7%), turtles (1.2%), and fish (1.2%). In contrast, crocodiles and turtles collectively represent almost a third of the total NISP recovered through surface collections at Ngira as a whole.

Only 152 of the 347 mammal specimens are sufficiently well-preserved to be identified at the genus level (Table 3). By comparison, the surface collections made at Ngira include over 3,000 fossils, but only 482 belong to small mammals (i.e., body size of the rodent M. pentadactylus or smaller) identifiable to the genus level. The taxonomic composition of the NG15 excavation includes 10 genera, and surface collections from throughout Ngira yielded 18 genera of comparable size ranges. Both samples show seven genera in common, whereas three genera are unique to the NG15 excavation sample and 11 to the surface collections (Table 3). Accordingly, the Jaccard similarity coefficient and the Sørensen–Dice index are very low (33.3 and 50% respectively; see Lyman, 2008), which indicate that the overall surface collection is quite different from the NG15 excavation in their taxonomic composition.

The small mammal community at Ngira is dominated by artiodactyls (especially the genus Dorcatherium) in the overall surface collection sample. In contrast, the NG15 excavation fauna is dominated by rodents, especially the genus Paraphiomys. Lagomorphs are equally well represented in both samples, whereas macroscelids (especially the genus Myohyrax) are more abundant in the surface collection sample. The NISP of each genus in common between surface collected and excavated samples at NG15 differ significantly in terms of taxonomic abundance (chi-square = 220.27; p < 0.001) and are not correlated (Spearman's rho = 0.36; p > 0.10), which suggests that the fauna recovered by the NG15 excavation may represent a separate ecosystem, or at least—taking into account time averaging for the surface collections—a specific ecosystem within an environmental mosaic that varied spatially and temporally at Ngira. Notably absent from NG15 are carnivorous mammals and primates, which are also rare in the surface and historical collections from Ngira.

Notable taxonomic differences were also observed within the NG15 profile (Table 4). The depth interval from 5–50 cm, which includes the C1, C2, and 2Btk horizons, is dominated by Paraphiomys and lacks remains of Dorcatherium and Pedetidae. The depth interval from 50 to 80 cm yields a moderate abundance of Dorcatherium and Pedetidae, but lacks the rodent Kenyamys. Although the taxonomic abundance of each depth interval is not significantly different (chi-square = 9.43; p > 0.32), the taxonomic composition does differ between the upper and lower portions of the profile (Jaccard similarity coefficient of 33.3%; Sørensen–Dice index of 50%).

The stratigraphic interval of NG15 transitions laterally from small fluvial channels (Driese et al., 2016) to pedogenically modified floodplain siltstones. Microtopography observed across the ca. 3 m described in the current study includes dipping sedimentary and paleosol horizon contacts that are overlain by coarser sediment, reworked carbonate nodules, and microvertebrate skeletal elements. We interpret these features to be the result of cut-and-fill episodes that were subsequently overprinted by pedogenic alteration.

The mineral assemblage reflects granitic basement sources in the surrounding uplands (Smith and Mosley, 1993), and lacks evidence of drainage connection to the Kisingiri volcanic complex to the northwest. The detrital grains observed in thin section consist of dominantly poly- and mono-crystalline quartz, with accessory orthoclase, microcline, and plagioclase feldspars. This mineral assemblage is likely derived from a granitic source rather than the mafic-alkaline volcanic source at Kisingiri, due primarily to the lack of tuffaceous material, volcanic rock fragments, and paucity of mafic minerals observed in thin section. Amphiboles and pyroxenes are rare in NG15 thin sections and are highly weathered. The Fe released by decomposition of these mafic minerals was conserved as concretions in B horizons.

The clay mineral assemblage in NG15 is composed of illite, and volcanic grains are notably absent. Illite is most likely a detrital rather than diagenetic mineral for two reasons. First, unaltered smectite was observed in the lowest unit at Ngira (NG01) and contains no illite (Driese et al., 2016). If the illite observed in NG15 were an alteration product of smectite, one would expect that NG01 would also have been illitized. The unconformity above NG01 coincides with a change in sediment composition and geomorphology, which accounts for the lack of detrital illite in NG01. Second, mass-balance calculations after removal of calcite from NG15 revealed that relatively little elemental variation was present between paleosol horizons. Significant production of pedogenic clay in B horizons would have been evident as increases in Al₂O₃ relative to SiO₂ (Retallack, 2008), which was not observed. Therefore, we interpret micromorphologic clay fabric and macroscopic clay coatings to be reworking of detrital clay.

Forbes et al. (2004) previously described the stratigraphy and selected paleosols at Ngira. Their KU-3 paleosol is almost certainly the same interval as NG15, as their description includes clay-lined columnar peds in a Bt horizon underlain by thin couplets of siltstone and sandstone (see Figure 3 in Forbes et al., 2004). The ped structure in the Btk horizon is actually prismatic, forming as a result of clay translocation and stabilization by herbaceous roots, rather than columnar peds, which form as a result of clay flocculation in sodic soils (Brady and Weil, 2008). Forbes et al. (2004) calculated constitutive mass-balance for samples equivalent to the Btk horizon of NG15 relative to a granitic parent material and found losses of 80–90% for SiO₂, Na₂O, and K₂O. This result led to the interpretation that the paleosols formed under subhumid to humid conditions. However, the parent material for the NG15 paleosol was fine-grained alluvium—originally sourced from granitic basement—and had significant additions of CaO associated with diagenetic calcite (Figure 6). Pre-weathering and reworking of alluvium before deposition at NG15 led to the overestimation of mass loss calculated by Forbes et al. (2004). After removing carbonate-associated CaO, we found that only moderate in situ weathering occurred and the climate was more likely dry subhumid.

The NG15 paleosols show evidence of variable drainage conditions. The 2Btk horizon contains prismatic ped structure, macro- and micro-scopic illuvial clay coatings, pedogenic carbonate, and oxidized color, all of which form during well-drained conditions (Birkeland, 1999). In contrast, isolated zones of Fe depletion are also present in both hand sample and thin section and are the result of saturated conditions. Hematite concretions are present in all B horizons and require intervals of water saturation to reduce and mobilize Fe from the surrounding soil matrix, and subsequent intervals of well-drained conditions to oxidize and stabilize Fe²⁺ to Fe³⁺ (Vepraskas, 2001). Taken together, these features indicate that the paleosols in NG15 were episodically waterlogged, but the predominance of oxidized features is consistent with well-drained conditions for relatively longer durations.

Mean annual precipitation estimations using CIA-K are bounded by underestimates associated with untreated samples containing artificially high CaO, and overestimates from decarbonated samples containing artificially low CaO due to the removal of small quantities of pedogenic carbonate. The best estimate for paleo-MAP is likely subhumid (500–1,000 mm), which falls between the values calculated for untreated (<500 mm) and decarbonated samples (800–1,200 mm). Previous MAP estimates from paleosol B horizons in NG19 and NG20, which occur ~5 m stratigraphically above NG15 and showed no evidence of diagenetic calcite, also yielded subhumid MAP estimates (Driese et al., 2016). Results from the PPM_1.0 model are variable with high error that encompasses all values calculated using CIA-K, and therefore offer less practical utility than the CIA-K estimates. The presence of illuvial clay and pedogenic carbonate is an indication of seasonally-variable precipitation.

Estimations of MAT using NaK, PWI, and PPM_1.0 are all clearly too low given the equatorial setting of the locality and presence of taxa that currently inhabit thermic to hyperthermic climate zones (e.g., crocodilians and turtles). Instead we suggest that MAT was probably considerably warmer and similar to fossil leaf estimates from Rusinga Island of 23–35°C (Michel et al., 2014). Paleosol studies frequently find the results from elemental proxies for MAT are below reasonable values, primarily because MAT is much less important in controlling soil weathering compared to other variables, such as MAP (e.g., Stinchcomb et al., 2016).

Previous investigations have often treated early Miocene deposits at Karungu and Rusinga Island, specifically the fauna of the Hiwegi Formation, as comparable (Nesbit Evans et al., 1981; Pickford, 1981). However, the results of this study and the work of Driese et al. (2016) call this assumption into question for a number of reasons. First, lithostratigraphic and paleopedologic analyses of the entire sequence indicate a variety of depositional environments (Driese et al., 2016). Additionally, our results, which document differences in the faunal assemblages between surface collections at Ngira and the excavation at NG15, also support the interpretation of multiple paleoenvironments within Ngira. Second, paleoenvironmental interpretations for the Hiwegi Formation on Rusinga Island indicate multiple environments ranging from more open shrubland to riparian woodland to closed-canopy tropical seasonal forest (Collinson et al., 2009; Conrad et al., 2013; Maxbauer et al., 2013; Michel et al., 2014). Third, the paleoenvironmental reconstructions for nearly every Critical Zone component reconstructed for NG15 are different from most recent environmental and climatic reconstructions of the Hiwegi Formation. While the stratigraphy at Ngira consists of interbedded fluvial and lacustrine strata (Driese et al., 2016), sedimentation in the Hiwegi Formation is a mixture of fluvial, palustrine, and primary and reworked volcaniclastic material (e.g., ash fallout and lahars; Bestland and Krull, 1999; Maxbauer et al., 2013; Michel et al., 2014). This indicates that there were substantial differences in physical sedimentary environments between Karungu and the Hiwegi Formation. We reconstruct the NG15 interval as an open canopy environment, ranging from an open riparian woodland to a wooded grassland biome. This is similar to, though perhaps more open than, environmental reconstructions for the Grit Member and lower Fossil Bed Member of the Hiwegi Formation, which are reconstructed as a patchwork of woodland and forested biomes within a riparian habitat in a warm and highly seasonal climate in which evaporitic conditions were recurrent (Collinson et al., 2009; Conrad et al., 2013; Maxbauer et al., 2013). Likewise, the interpretations for the NG15 interval are considerably different from the paleoenvironmental reconstruction for the Kibanga Member of the Hiwegi Formation, which is reconstructed to be a closed-canopy, tropical seasonal forest (Michel et al., 2014). Fourth, there are notable differences in the fauna recovered from Karungu and that of the Hiwegi Formation, particularly in the composition and abundance of small mammals and especially primates.

These differences between Karungu and Rusinga Island draw into question the reliability of wholesale comparisons of faunal composition and abundances between early Miocene sites before detailed geologic analyses are performed. For example, Pickford (1981) suggested that the Karungu fauna was most similar to that of the Hiwegi Formation from Rusinga Island (i.e., his faunal Set II). Perhaps the most conspicuous difference between Karungu and the Hiwegi Formation is the rarity of primate taxa at Karungu, in comparison to the relatively common primates found in the Hiwegi Formation. Coupled with evidence for variability in paleoenvironment and paleovegetation through the deposits at Karungu and through the Hiwegi Formation (e.g., this study; Collinson et al., 2009; Maxbauer et al., 2013; Michel et al., 2014; Driese et al., 2016), the differences in the faunal abundance and occurrences between Karungu and Rusinga Island are likely related to true faunal differences, possibly as a result of paleoenvironment differences.

Compared to the 34 genera (smaller or equal in size with M. pentadactylus) recorded from the Hiwegi formation at Rusinga Island (e.g., Pickford, 1986), the small mammal collections of both the NG15 excavation (10 genera) and the Ngira surface collections (18 genera) are less taxonomically diverse. This could be an artifact of sampling bias, but we have exhaustively collected the Ngira deposits at Karungu over the last several years and small mammals are uncommon and carnivore and primate taxa are very rare. Thus, we hypothesize that the difference in small mammal diversity and abundance between Karungu and the Hiwegi Formation on Rusinga Island reflect true differences in the fauna. When our results are coupled with recent paleoenvironmental reconstructions from Karungu and the Hiwegi Formation (e.g., Collinson et al., 2009; Conrad et al., 2013; Maxbauer et al., 2013; Michel et al., 2014; Driese et al., 2016), we conclude that a substantial heterogeneity of environments and climates existed in western Kenya in the early Miocene, and the paucity of primates at Karungu is not likely the result of biotic factors alone.

In general, early Miocene environments across East Africa are interpreted to have been dominated by tropical forest or woodland vegetation (Andrews and Van Couvering, 1975; Edwards et al., 2010). Stable isotope analyses of compound-specific biomarkers (Uno et al., 2016), pedogenic carbonate (Cerling, 1992), and fossil tooth enamel (Cerling et al., 1998) suggest that C₄ vegetation was not common across East Africa until ca.10 Ma, with a gradual rise of C₄-dominated grasslands starting in the late Miocene at ca. 8 Ma. Nevertheless, several lines of evidence suggest that a patchwork of environments existed in the early Miocene, at least in some parts of the region.

Sediment cores from the Niger River delta show evidence of grass pollen and charred grass cuticle appearing in low abundance at ca. 16 Ma, which is interpreted to signify the initiation of seasonally dry savannahs (Morley and Richards, 1993). A review of Cenozoic vegetation across East Africa by Bonnefille (2010) provides strong evidence for a patchwork of dry, seasonal woodlands to humid forest habitats across the tropical belts of the African continent as early as the late Eocene. Pedogenic carbonates from the Tugen Hills succession near Lake Baringo in Kenya have δ¹³C values consistent with mixed C₃ and C₄ vegetation from 15 Ma to Recent (Kingston et al., 1994). Contemporaneous mammals from the same strata in the Tugen Hills also have tooth enamel δ¹³C values that are higher than expected for a pure C₃ endmember, and in the absence of vegetation with extreme water stress, those mammals with high δ¹³C values likely consumed some portion of C₄ vegetation in their diets (Morgan et al., 1994). Based on the tendency of C₄ vegetation to exist in open canopy conditions in modern ecosystems, it is likely that the Tugen Hills succession contained at least a patchwork of woodland/shrubland to woody grassland environments.

Our observations supporting the presence of open canopy, dry woodland to woody grassland at Karungu are consistent with previous faunal analyses by Andrews et al. (1979), who compared the composition of modern faunas with that of early Miocene sites. These included a total of 15 species (including microvertebrates) (Andrews, 1974) from Karungu recovered from excavations and sieving from Oswald's (1914) Bed 16. According to our fieldwork, this stratigraphic interval corresponds to units 20–22 of Driese et al. (2016), 3–5 m above NG15. In Andrews' et al. (1979) analysis, the fauna from Karungu stood out because: (1) its ecological diversity patterns as well as its locomotor adaptations resemble those of non-forest faunas; (2) its taxonomic composition closely resembles that of woodland-bushland faunas (except for the lack of carnivorous species); (3) the smallest size categories predominate the assemblage; and (4) interpreted feeding patterns of the community overall were unlike those in any modern community. Interestingly, unlike the NG15 excavation, this faunal assemblage was dominated by Kenyalagomys and Myohyrax (74% of the total NISP; Andrews, 1974). But because both taxa have no living relatives in Africa today, Andrews et al. (1979) considered that this faunal assemblage may either represent a habitat type that is no longer represented in Africa, or was a limited sample from a larger community. Based on similarities in pedogenic carbonate δ¹³C values and paleoclimate estimates between NG15 and NG19-NG20, we hypothesize that the fauna recovered from NG20-NG22 also represent open canopy conditions. These environments may not be comparable to any known modern landscape due to the possibly ephemeral and unique nature of early C₄ savannahs or grasslands.

Finally, we posit a possible explanation for the lack of records of C₄ vegetation in recently studied marine cores. Uno et al. (2016) present a compound-specific isotope record from cores taken from the Somali Basin, Gulf of Aden, and Red Sea. Their data offer a large-scale picture of vegetation types in East Africa, but necessarily sample externally-drained continental basins. However, the Ngira locality was potentially an endorheic basin, given the rapid lake level fluctuations documented in the stratigraphy (Driese et al., 2016), and therefore is very unlikely to have contributed to the record of Uno et al. (2016). This finding underscores the fact that, while a major turnover from C₃ to C₄ vegetation occurred in the late Miocene, much work remains for documenting landscape heterogeneity earlier in the geologic history of East Africa.

Taken together, the observations of pedogenic carbonate δ¹³C signatures, paleosol morphology, microcharcoal, and microvertebrate fauna presented in this and previous studies support an interpretation of open canopy conditions within an overall diverse environmental mosaic in the early Miocene of western Kenya.

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.