Effects of livestock and wildlife grazing intensity on soil carbon dioxide flux in the savanna grassland of Kenya

Although grazing is the primary land use in the savanna lowland of southern Kenya, the effects of grazing on soil carbon dioxide flux (RS) remain unclear. A 12-month study was conducted from January to December 2020 on the effects of six grazing intensities sites (overgrazed (OG), heavily grazed (HG), moderately grazed (MG), moderately to lightly grazed (M-LG), lightly grazed (LG) and no grazing (NG)) on RS on. A camera trap was used to monitor the total number of animals at each site, indicating the grazing intensity. Weekly measurements of RS were taken using static greenhouse gas chambers along with simultaneous measurements of soil temperature (TS) and volumetric soil water content (WS) (depth of 5 cm). Mean RS at HG, MG, M-LG and LG sites was approximately 15–25% higher than at NG and OG sites (p < 0.001). Mean WS increased with decrease in grazing especially in the dry season, while TS increased with increase in grazing. We observed bimodal temporal variation in RS and WS due to two wet seasons in the year. Thus, variation in RS across the study period followed the changes in WS rather than those in TS. Mean values of RS in the wet seasons were significantly higher (> 45%) than those in the dry seasons, and WS accounted for 71% of the temporal variability in RS (p < 0.05). In addition, the enhanced vegetation index (EVI, interpreted as a proxy for vegetation cover) explained 60% of the variance of RS, and WS and EVI together explained 75%. EVI showed a negative relationship (p < 0.05) with animal intensity, indicating that more grazing reduced vegetation cover and, consequently, soil organic carbon and biomass. Soil bulk density was lower at less grazed sites. While RS variability was unaffected by total nitrogen content, pH, and texture, correspondence analysis demonstrated that the main factors influencing RS dynamics across the year under different grazing intensities were WS and vegetation cover. Our results contribute to closing the existing knowledge gap regarding the effects of grazing intensity on RS in East Africa savannas. Therefore, this information is of great importance in understanding carbon cycling in savanna grassland, as well as the identification of the potential consequences of increasing land pressure caused by rising livestock numbers, and will assist in the development of climate-smart livestock management in East Africa.


Introduction
Savanna grasslands cover ca. 50% of the African continent (Scholes et al., 1997;Castaldi et al., 2006). Covering such a huge area, they have the potential to affect atmospheric carbon dioxide (CO 2 ) concentrations, net ecosystem productivity and the carbon balance at local, regional and global scales (Scholes et al., 1997;Castaldi et al., 2006;Raich and Schlesinger, 1992). Soil respiration also referred to as the soil CO 2 flux (R S ) including root and microbial respiration is the single largest soil-atmosphere carbon flux from savanna grasslands (Makhado and Scholes, 2012). Key environmental factors that control R S in the savanna grasslands include precipitation, soil moisture (W S ) and soil temperature (T S ) (Davidson et al., 1998;Lloyd and Taylor, 1994;Merbold et al., 2009), as well as vegetation and substrate availability (Wan and Luo, 2003). Nevertheless, land use types and management practices, such as grazing can modify these factors and further alter R S and carbon cycles (Wang and Fang, 2009). Thus, accurate assessment of R S and its controlling factors are crucial for understanding carbon cycling in savanna (Bond-Lamberty andThomson, 2010, 2018).
Grazing by both livestock and wildlife is an important component of savanna grasslands (Abdalla et al., 2018). Each animal species differs in terms of its nutritional requirements, jaw anatomy, and grazing habits (Olff and Mark, 1998). Additionally, animal grazing dynamics are influenced by forage and water availability, which varies depending on productivity (Begzsuren et al., 2004). Thus, animal grazing, feeding preferences, and trampling may alter grass cover, vegetation composition, and microclimate, in turn altering R S and its controlling factors (Patton et al., 2007).
Grasses rather than large shrubs or trees dominate grassland (Grace et al., 2006). Thus grazing may reduce vegetation cover especially the grass cover, thereby reducing above-and belowground carbon and nutrient allocation to plants, root respiration and production, and soil organic matter, depending on the intensity (Abdalla et al., 2018;Carbone et al., 2008;Wang et al., 2017). A reduction in vegetation cover exposes soil to direct sunlight, which increases soil surface temperatures and evapotranspiration, and reduces W S , and further alters metabolic activity of microbes and plant roots (Carbone et al., 2011;Bond-Lamberty andThomson, 2010, 2018;Grace. et al., 2006). Moreover, trampling of soil increases soil bulk density (BD), which reduces water infiltration and water availability for root and soil microbes (Patton et al., 2007;Tollner et al., 1990) and diffusion of gases in/out of the soil (Li et al., 2008;Patton et al., 2007;Tollner et al., 1990). All these factors may have a direct or indirect effect on R S (Cao et al., 2004). Moreover, dung and urine deposition by animals provide easily degradable carbon and nitrogen resources, which can stimulate plant and microbial growth, further affecting R S . (Janssens et al., 2001;Davidson and Ivan, 2006).
The essential mechanisms driving the responses of R S to grazing are still not fully understood, as contradictory results have been reported from a range of studies (Cao et al., 2004;Wang et al., 2020). While some have reported an increase in R S with increased grazing in China (Dong et al., 2000), in the northern great plains and south-central North Dakota grasslands of the USA (Frank, 2002;Patton et al., 2007), others have reported a decrease in R S with decreased grazing in a humid savanna in Kenya (Ondier et al., 2020), the Kalahari desert in southern Africa (Thomas, 2012), alpine meadows and semiarid grasslands in China (Cao et al., 2004;Wang et al., 2020) and managed pastures in Brazil (Brito et al., 2015). Yet others found no effect of grazing on R S in a semi-arid savanna in South Africa (Munjonji et al., 2020), as well as a grassland and an alpine meadow in China (Ling-Hao et al., 2000;Cui et al., 2014).
The inconsistencies between these results suggest that the effect of grazing on R S is complex and might be ecosystem-dependent. The reason may be that different grasslands respond differently to grazing depending on the vegetation type, the animal type, and the feeding preferences, all of which depend on the history, management practices, and levels of grazing utilization of the grasslands (Cao et al., 2004;Frank et al., 2002;Jia et al., 2007;Patton et al., 2007). In addition, the effects of grazing are combined with and are affected by other biotic and abiotic factors that vary with climate zones (Cao et al., 2004;Frank et al., 2002;Jia et al., 2007). The majority of these studies are from temperate grasslands that have soil rich in organic carbon and abundant grass growth (Dong et al., 2000;Frank et al., 2002;Cao et al., 2004;Patton et al., 2007), but such studies from savanna landscapes in Africa remain scarce (Grace et al., 2006), even less is reported for savanna grasslands in the Sub-Saharan Africa (Oduor et al., 2018).
Savanna grasslands unlike temperate grasslands, have low soil carbon (Grace et al., 2006) and are characterized by unpredictable rainfall and variations in vegetation growth, which combined with grazing leads to highly dynamic environments (Grace et al., 2006;Broadl et al., 2018). Rainfall and W S are major determinants of primary productivity (Grace et al., 2006). The availability of WS increases microbial activity and decomposition of soil organic matter, which stimulates grass growth and root growth, thereby stimulating R S (Carbone et al., 2011;Orchard and Cook, 1983). While active roots contribute directly to R S , dead roots and root exudates provide carbon as a source of energy and nutrients for soil microbial biomass (Tufekcioglu et al., 2001). Previous studies in savanna grasslands have shown that W S and vegetation status significantly influences R S variations, but T S has a marginal effect on R S because of its marginal intra-annual variation in the tropics (Grace et al., 2006;Ardö et al., 2008;Broadl et al., 2018). The effect of grazing on soil and vegetation, however, is one of the critical factors that alters productivity, microclimates, and eventually R S in savanna grassland.
In the savanna grasslands of Sub-Saharan Africa, grazing intensity has increased due to an increase in demand for livestock products as a result of population growth (Osborne et al., 2018). This has seen an increase in the number of grazing livestock over the years (Osborne et al., 2018), and as a result, overgrazing due to overstocking is a major cause of soil and vegetation degradation (Abdalla et al., 2018;Patton et al., 2007). Overgrazing reduces ground cover and thus exposes the soil to more radiation, leading to higher evapotranspiration and erosion (Abdalla et al., 2018). Furthermore it leads to less inputs of organic matter as litter, nutrients, and biotic activity, which consequently affects R S (Abdalla et al., 2018).
In Kenya, savanna grasslands cover more than 80% of the landscape and contain approximately 50% of the total livestock population, primarily within pastoral communities. Similarly they are home to nearly 80% of the total wildlife population (Kinyenze et al., 2019). Given the vast area covered by this ecosystem, any effects of grazing on R S are likely to influence the local, regional, and global carbon cycles (Scholes et al., 1997;Castaldi et al., 2006;Raich and Schlesinger, 1992). To the best of our knowledge, only a few studies have been conducted in Kenya and in East Africa on the response of R S on restored communal grazing lands (Oduor et al., 2018) and the interactive effects of grazing, rainfall and drought on R S (Ondier et al., 2020;Munjonji et al., 2020). Nonetheless, no such study has been conducted on the effects of different levels of grazing on R S in the savanna grasslands.
To contribute to closing this knowledge gap, we measured R S under five different levels of grazing in the savanna grasslands of southern Kenya. In addition, a control site with no grazing was established in a fenced enclosure meant for restoration purposes. This study aimed at (i) evaluating the effects of different grazing intensities on R S , (ii) assessing the temporal variations in R S at each grazing site for a full year and (iii) identifying key environmental variables driving the response of R S to variations in grazing. We hypothesized that grazing exclosure and lighter grazing sites would have higher R S rates than heavily grazed and overgrazed sites due to the effects of grazing on vegetation cover and W S . Furthermore, we hypothesized that W S pulses, such as occurring during the rainy seasons, are an important driver of R S , with T S being of lower priority due to its low seasonal variation in the tropics.

Site characteristics
The experimental site was located in the lowlands of Taita-Taveta County in southern Kenya (38 • 8 ′ E, 3 • 28 ′ S, about 800-1000 m above mean sea level). The site is an expansive lowland with a slope of less than 5%. About 89% of county area is considered to be arid and semi-arid land (ASAL). The Tsavo East and Tsavo West national parks cover ca. 62% of the county area (CIDP, 2014), but private sanctuaries, such as the Taita Hills Wildlife Sanctuary (THWS) and LUMO Community Wildlife Sanctuary (an amalgamation of three ranches, namely Lualenyi, Mramba and Oza, hence the name LUMO), increase the land area under conservation. There are 28 ranches designated for livestock production in the county.
The local climate is tropical with two rainy seasons: a long rainy season between March and May (with a peak in April) and a short rainy season between October and December (with a peak in November). January and February are dry and hot months, while June to September are dry and cool months (Pellikka et al., 2018). The mean annual rainfall in these lowlands is about 500 mm and the mean annual air temperature is 23 • C, with mean daily minimum and maximum temperatures of 16.7 • C and 28.8 • C, respectively (Pellikka et al., 2018). The dominant soil type is rhodic ferralsol, which is characterized by very deep, acidic, dark red, weathered sandy clay soil (CIDP, 2014). For this study, grazing sites were located in both THWS and LUMO, which covers a wide range of grazing intensities.These sanctuaries consist of open savanna grassland, with areas of open woodland and shrublands with Acacia spp. as the main trees, having an extensive variety of herbaceous vegetation dominated by Chloris roxburghiana, Schmidtia bulbosa and Cenchrus ciliaris grass species. THWS (900-1100 m a.s.l.; covering an area of ca. 110 km 2 ) is a conservation area under private management for wildlife conservation only, and grazing is thus light to very light. LUMO (850-1050 m a.s.l; covering approximately 460 km 2 ), on the other hand, is a conservation area with some areas dedicated for livestock management (Amara et al., 2020). It is located between Tsavo West National Park and THWS and is a communally owned (GoK, 2013) for community livestock grazing and also wildlife. Large herds roam within LUMO, causing overgrazing, which is enhanced by illegal herders, who mainly enter the conservancy with their livestock in the dry season (CIDP, 2014). Conservation, on the other hand, has reduced the woody vegetation cover due to the large browser population, especially in the THWS (Amara et al., 2020). Our control site was located in a fenced reforestation project area located in THWS that was established in 2010. The electric fence and open soil around it prevents grazing livestock, other larger mammals and lizards, as well as fire, from entering this 13 hectare area (Fig. 1).

Experimental design
In November 2019, six sampling sites were established to represent different levels of grazing, and subsequent R S measurements were taken at these sites (see Table 1; Fig. 1). Because the differences in vegetation cover in the study area are primarily caused by grazing and browsing, the sites were selected using satellite and airborne remote sensing data (Amara et al., 2020). This was supported by field observations, the number of animals captured at each site by wildlife cameras, and consultation with the management from THWS and LUMO. At each site, four circular stainless steel chamber collars (21 cm in diameter and 10 cm in height) were randomly installed in January 2020. The collars were inserted into the soil to a depth of up to 5-6 cm one week prior to the first R S concentration measurements to avoid disturbance of the measurements due to collar installation. The collars were left in the ground for the entire study period (Søe et al., 2004). Collars that were interfered with or damaged by animals were replaced at least 24 h LG -moderately to lightly grazed, LG -lightly grazed and NG -no grazing (see Table 1 for details). Sentinel-2A satellite image, 20 December 2020, Sentinels Scientific Data HubCE4 (ESA, 2015). before the next gas sampling took place. Over the length of the study period, the heavy grazing (HG) site was affected by wildfire between 8 and 9 August 2020.

Measurements of soil CO 2 concentrations and flux calculation
Field measurements of R S were carried out weekly from 14 January to 22 December 2020. On each of the gas sampling days, the chamber collars were enclosed using a matching lid (internal diameter = 20 cm; height = 24.5 cm) covered with reflective aluminum foil to minimize heating inside the chamber during the period of actual measurements (Fig. 2B). In addition, the duration of the measurement was 5 min at maximum to avoid saturation effects. A small fan was mounted on the chamber ceiling to circulate the air within the chamber and ensure a well-mixed gas sample. The chamber lid had a small vent on it to prevent pressure build-up in the chambers during measurements. The lid was fitted with a CO 2 probe (Vaisala CARBOCAP® GMP343, CO 2 sensor, Vaisala Oyj, Helsinki, Finland) and a temperature probe (HMP75, Vaisala Oyj, Helsinki, Finland). The CO 2 sensor measured the CO 2 concentration in parts per million (ppm). CO 2 has a characteristic absorbance band in the infrared (IR) region at a wavelength of 4.26 µm. Therefore, when IR radiation is passed through a gas containing CO 2 , part of the radiation is absorbed. The CO 2 and temperature probes were connected to a Vaisala MI70 indicator to measure changes in the temperature and CO 2 concentration at the same time for 5 min at an interval of 15 s per chamber closure time, giving us 20 readings per chamber deployment time. R S was calculated using linear least squares regression of CO 2 concentration within the chamber headspace versus time. The first 20 s of data points were discarded due to noise, as recommended by the manufacturer, ensuring that we used the linear portion of the curve. Individual chamber fluxes were calculated if the linear regression for each individual CO 2 yielded R 2 > 0.9 or normalized root mean square error < 0.2 (Christiansen et al., 2011). We also discarded data points for which the rate of change of CO 2 was negative, as photosynthesis can be assumed to be zero inside an opaque chamber. Therefore, R S was calculated using Eq. (1) as follows: where ∆c/∆t is the change in the chamber headspace CO 2 concentration over time (i.e., slope of the linear regression), V ch is the volume of the chamber headspace (m 3 ), M w is the molar weight (g mol − 1 ) of C for CO 2 (12), A ch is the area covered by the chamber (m 2 ) and M v is the pressureand temperature-corrected molar volume (see details in Brümmer et al., 2008aBrümmer et al., , 2008b. The values 60 and 10 6 are constants used to convert minutes into hours and grams into micrograms, respectively. Fluxes are expressed as mg C m − 2 h − 1 .

Microclimate data
Soil temperature (T S, • C) and volumetric soil water content (W S , %) at the depth of 0-5 cm were measured next to the collars using a handheld data logger connected to a GS3 sensor (ProCheck, METER Group, Inc., USA) on each gas sampling day. The air temperature and  Table 1  atmospheric pressure were recorded using a handheld digital thermometer and Garmin GPSMap 64. Daily rainfall (mm) over the year was measured with a tipping bucket (ARG100, Campbell Scientific, USA) and air temperature (T A , • C) with a digital thermometer. Both were installed at a weather station installed at site LG in THWS in an open area to avoid interference from trees. Measurements were taken every 5 min, averaged and stored every 30 min on a data logger (built in CR200X, ARG100, Campbell Scientific, USA). In addition, continuous T A , T S and surface temperature (T Surf ), and W S measurements were taken at 15-minute intervals in the proximity of the collar chambers using TMS-4 data loggers (TOMST®, Czech Republic). The TMS-4 has three temperature sensors for T A , T Surf and T S respectively positioned at + 15, 0 and − 8 cm relative to the soil surface (Wild et al., 2019). The W S sensor, located in the bottom part of the TMS logger probe, measures volumetric soil moisture at a depth of about − 14 cm (Wild et al., 2019). The sensors were kept in place and fenced using wire cages to prevent interference by animals (Fig. 2B). However, we had to replace the TOMST sensors at the HG site after they were burnt in a wildfire in August 2020.

Evaluating the intensity of animal use at each grazing site
In order to monitor and estimate the intensity of animal use at each site, we installed motion-activated wildlife cameras (APEMAN Trail Cam H45, China) to record the number of animals visiting the sites during the study period. The cameras had a detection angle of 90 • and a field of view (FOV) of 73 • . They were fixed on a tree (2.5-3 m above ground) to avoid disturbance by animals that typically scratch their body against the trees. We set the cameras to record three consecutive images with a 3-s interval each time the sensor was triggered by animal movement within the camera's detection zone. The recovery time was set at 1 min in order to record images of the same animal or group of animals passing in front of the camera only once ( Van Berkel, 2014).
We defined the intensity of animal use as the total number of animals captured in an image at any given time during the whole study period at each site. In addition to counting, we recorded the type of animal species and whether it was grazer or browser, the time and date, and the total number of animals in each image. When a group of animals such as a herd of livestock triggered a camera, the intensity of use (number of animals) was recorded as the maximum number of individuals in that image. This was then binned into the total and mean per month to evaluate the number of animals at each site and across the year to assess the effect on the monthly vegetation index (a proxy of vegetation status) and monthly R S . Data were downloaded and the batteries and memory card were checked every month.

Measurements of vegetation characteristics at GI sites
Since the differences in vegetation cover between the sites were mostly caused by grazing and browsing, changes in the remotely sensed vegetation index (Huete et al., 2002) were assumed to reflect the grazing intensity (Ma et al., 2019). Therefore, to investigate differences in the vegetation status and its effect on R S at each site, we used remotely sensed composite enhanced vegetation index (EVI) data from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data calculated by the following equation (Eq. (2)): where ρblue, ρred and ρNIR are atmospherically corrected surface reflectance in blue, red and near infrared (NIR) wavelength bands, respectively, L is the canopy background adjustment factor (L = 1), C 1 and C 2 are coefficients of the aerosol resistance terms (C1 = 6 and C2 = 7.5), and G is a gain factor (G = 2.5) (Huete et al., 2002).
We acquired MOD13Q1 (MODIS Level 3 products) EVI data generated at 16-day intervals with a 250-m spatial resolution from https:// ladsweb.modaps.eosdis.nasa.gov on 23 January 2021. To maintain temporal consistency with the gas sampling data, we selected EVI data captured closest to the gas sampling day or at maximum 5 days before or after this day, assuming that no substantial change in vegetation would have occurred in such a time. We extracted EVI pixel values closest to each site based on their latitude and longitude.
In addition to the EVI data, five (1 m × 1 m) quadrats were randomly selected at each site to assess the herbaceous aboveground biomass (AGB). This was done by clipping all the grasses and small shrubs within the quadrats to ground level (< 1 cm) using hand shears. Sampling was carried out monthly from April to December 2020. The fresh weight of the samples was recorded, and after drying the samples in an oven at 70 • C for 48 h, the dry weight to the nearest 0.01 g was recorded. The biomass of the five quadrats per site was averaged to represent the biomass of each site.

Soil sampling and analysis
Soil sampling was conducted in April and June 2020 at each site by randomly collecting four soil samples at a depth of 0-20 cm using a soil auger (Eijkelkamp Agrisearch Equipment, Giesbeek, The Netherlands). All samples were mixed in a bucket, and all plant debris and roots were removed. A composite soil sample (500 g) was taken and stored in airtight polyethylene bag. This exercise was repeated three times, giving us three composite soil samples per site. Another sample (at a depth of 0-5 cm) was taken using a bulk density ring (Eijkelkamp Agrisearch Equipment, Giesbeek, The Netherlands) for bulk density and water content assessment. The samples were labeled accordingly and transported in cooler boxes to the Mazingira Centre at ILRI (Nairobi, Kenya) for the analysis. In the laboratory, the samples were stored at 4 • C until processing within 14 days. The subsample of field moist soil was airdried at room temperature and sieved using a 2-mm sieve to determine soil organic carbon (SOC), the total nitrogen content (TN), soil pH and texture. Soil texture was determined using the hydrometer technique (Reeuwijk, 2002;Scrimgeour, 2008) and soil pH was measured in a soil:distilled water suspension (1:2.5) using a pH meter (3540 pH and conductivity meter, Bibby Scientific Ltd, UK). To determine SOC and TN, a duplicate of 20 g of the fresh sample was oven-dried at 50 • C for 48 h and then pounded into a fine powder (< 0.25 mm) using a ball mill (Retsch ball mill, Haan, Germany) followed by analysis using an elemental analyser (Vario MAX Cube Analyzer Version 05.03.2013). BD were determined by drying soil samples collected with bulk density rings of known volume at 105 ℃ for 48 h.

Data analysis
Descriptive statistics (mean ± SE and coefficient of variation (CV)) were used to examine differences in the measured variables between the sites and seasons. Monthly and seasonal variations in R S were calculated from the daily arithmetic mean of hourly R S , which were the averages of all chamber measurements taken on each gas sampling day. R S data were tested for normality (Kolmogorov-Smirnov coefficients) and equality of variance (Levene test) before and after data transformation. R S data were not normally distributed and data transformation (log and square root) did not improve data normality, therefore the original data were used for statistical analyses. We used boxplots (displaying the minimum, first quartile (Q1), median, third quartile (Q3), and maximum of data) to examine the temporal variability of R S , W S and T S between sites across the study year. The effects of (i) grazing site and (ii) month on R S were tested using a generalized linear mixed model (GLMM) fit by maximum likelihood (Gbur et al., 2012). Grazing sites were treated as fixed effects and the months of the year as random effects in the model. A distribution test revealed gamma distribution to provide a good fit for the R S data and therefore a GLMM Gamma model was used.
The mathematical representation of the model is presented by the equation (Eq. (3)), therefore the model using Rstudio package lme4 (Bates et al., 2014).
where y = response variable, x = explanatory variables, μ = random effects and ε = error term.
Another GLMM model was developed using W S , T S and EVI as fixed effects and grazing sites as random effects to examine the effect of environmental variables on R S . Spearman correlation coefficients were applied to assess the relationship between the cumulative monthly and annual mean R S (calculated by trapezoidal integration of daily fluxes with time) against animal intensity of use and environmental variables that were measured once or twice during the study period, including soil pH, BD, soil texture, SOC and TN, herbaceous AGB. The aim was to identify the most significant variables driving the spatial and temporal variation in R S . All statistical tests were performed at the 5% level of significance. Summary values are expressed as the mean value ±standard error of the mean (SE). Statistical procedures and plotting were performed using the software packages in R 3.5.2.

Effects of grazing intensity on soil CO 2 flux
Soil CO 2 efflux (R S ) rates measured during the study period at all sites ranged from 10 to 356 mg C m − 2 h − 1 . Overall, mean R S was lowest at OG (59 ± 7 mg C m − 2 h − 1 ) and NG (67 ± 7 mg C m − 2 h − 1 ) and highest at MG (96 ± 9 mg C m − 2 h − 1 ) ( Table 2). With sites as the fixed effect and the month of year as the random effect, a GLMM model revealed that the difference in grazing sites had a significant effect on R S (p-values < 0.001; Table 3) accounting for ca. 60% of the spatial variation in R S . A post hoc analysis that involved a pairwise comparison of the R S across the sites revealed that OG and NG were significantly lower than all the other grazed sites (HG, MG, M-LG, LG), and the pairs M-LG and HG, M-LG and MG were significantly different. We considered individual months across the year, and observed a significant increase of approximately 10-25% in R S with the decrease in the grazing evident in January, February, May and November (OG < HG < MG < M-LG < LG; Fig. 3c). In April, no significant differences between the sites was observed, however, in the months of June, July and December, R S increased with decrease grazing (OG < HG < MG) and thereafter gradually decreased in M-LG and LG. This only applied to the grazed sites, as we observed a decrease in R S at NG throughout the rainy season months. On average, R S was lowest at OG in the dry season months and at NG site in the rainy season months. In dry months, during the earliest months of the dry season in June and July, the highest R S was from MG but latter months in September and October the highest was from NG.

Temporal variations in soil CO 2 flux
Overall, R S displayed a bimodal pattern for all sites throughout 2020 (Fig. 3d), with a CV ranging between 60% and 80%. Mean R S values significantly differed across the seasons (p < 0.01), with the highest R S ranging from 50 to 356 mg C m − 2 h − 1 occurring in the two wet seasons (rainy seasons) and in the short dry season (SDS) (mean R S for each site per season is provided in Table 2). During the expected SDS, sporadic rainfall were observed from last week of January, which led to higher W S (12-18%) and thus higher R S . On the other hand, the lowest R S values (ranging from 5 to 60 mg C m − 2 h − 1 ) were observed in the long dry season (LDS), which lasted from June to the third week of October. The observed variation in R S for all the sites followed the changes in W S (Fig. 3d) rather than those in T S , and thus R S substantially increased with the onset of rainfall events (Fig. 3b). W S at 5 cm also followed a bimodal seasonal pattern, with higher W S (mean from 15% to 30%) during the long wet season (LWS: long rainy season from end of March to May) and short wet season (SWS: short rainy period from November to December) and the lowest W S (< 10%) during the February in the short dry season (SDS: period between January to end of February) and LDS, with a high temporal CV across the study period ranging between 60% and 75% at different sites. On the other hand, temporal CV for T S across the study period was low ranging between 15% and 25% at different sites The difference in R S between the LWS and SWS was not significant, but R S in the SDS was 50% higher than in the LDS. Overall, the mean R S in the combined LWS and SWS period was not significantly different to SDS, but > 45% higher than in the LDS.

Effects of volumetric soil moisture and soil temperature on soil respiration
Overall mean W S (13 ± 3%) did not differ between the sites during the study period (Fig. 3d). However, during the dry season (LDS), W S was slightly higher at NG than the grazed sites, but throughout, OG had the lowest W S . Soil T S , on the other hand, was slightly higher at OG and HG (ca. 33 ± 7 • C) and the lowest at NG (29 ± 6 • C) (Fig. 3e) through the study period. Site NG had tall shrubs and bushes, which might have caused shading, thereby cooling the soils compared to the other sites. Surface temperature (T Surf ) was also lowest at the NG site (Fig. A1). W S was found to have a linear relationship with R S , (p-value = <2e-16). The random effect of the grazing sites were found to be significant due to its large variance of 4.97. Parallel slopes were therefore produced for the different grazing sites to summarize the relationship between W S and R S    Tables 4 and 5 for model parameter).
( Fig. 4a; Table 4). W S in M-LG and LG sites led to higher amounts of R S while W S levels in NG had the least R S . Overall, the field measurements of W S account for about ca. 70% (Table 4) of the variation in the R S . Soil T S , on the other hand, did not show any significant (p > 0.09) relationship with R S within each site and across months during the study year. Our results indicate that W S was essential in controlling the seasonal dynamics of R S overall and at each site. Combining the effects of W S and T S on R S did not improve the model with W S as a single factor.

Intensity of animal use at each grazing site
Altogether, the camera traps at all sites captured a total of 12,711 animals from last week of January to December 2020. We experienced camera failures for a week in August at the HG due to a wildfire (8-9 August), and in October at M-LG, LG and NG due to a failure of the memory card and/or batteries, which might have affected the data for that period of time. However, the frequency of animal use (images captured) was the highest at OG (ca. 33% of the animals) and HG (ca. 23%) and the lowest at MG (ca. 6%) (Table A1). About 97% of the of the animal captured were grazers and browsers, while 3% were predators, such as lions or hyenas, or baboons and stray dogs. Furthermore, 78% of the grazers and browsers were livestock, including cattle (89% of the livestock), sheep and goats. The total number of animal recorded were highest at OG, HG and MG compared to M-LG and LG (no livestock were photographed at the LG). At the MG, 85% of the livestock were captured between 26 September and 28 October, and at NG, about 142 cattle intruded the site at night between 3 and 8 September due to a damaged fence (Table A1). Overall, the highest number of wildlife grazers and browsers was recorded at LG, the number decreasing with increasing grazing by cattle (Fig. 5b). Buffalos were the most abundant (ca. 39%) wildlife captured, of which about 62% predominantly occurred at the LG. Over the year, the results from OG showed a trend towards a higher intensity of animal use between May and August followed by a drop in September and October and then a slight increase in November and December (Fig. 5a). The reason for the drop is evidently that in September to October the land is mostly barren.

The effects of grazing intensity on vegetation
Overall, mean EVI values across the study period were highest at the NG (0.39 ± 0.02), with the least disturbance, and the lowest at OG (0.24 ± 0.02) (Fig. 6b). HG had a significantly higher EVI (p = 0.02) than other sites in January and February, while during the rest of the year, the highest EVI was generally at NG and MG. NG had a greater cover of shrubs and smaller trees because grazing and browsing were prohibited. Although cattle intruded on the NG in September, this had no observable effect on the EVI. At all sites, we observed an increase in the EVI at the onset of the rainy season in March, but at OG, there was a gradual decline in EVI as the number of animals increased from April and May (Fig. 5a). The herbaceous ranged between 200 and 1200 g dry weight (DW) m − 2 across the sites. AGB was slightly significantly highest at the MG and lowest at the OG and HG ( Fig. 6a; p < 0.05). Still, EVI and AGB at HG were affected by a wildfire in August. Therefore, HG recorded the lowest EVI (Fig. A2) in August and September. Overall, the mean herbaceous AGB displayed a positive relationship with the mean EVI ( Fig. 6d; p < 0.05; R 2 = 0.31). Mean animal count (Fig. 6c) revealed a negative relationship against mean AGB ( Fig. 6e; p < 0.05; R 2 = 0.26) and mean EVI ( Fig. 6f; p < 0.05; R 2 = 0.58), indicating that an increase in the number of animals led to a decrease in the mean EVI and AGB. Across the year, the highest EVI values were observed from January to the end of May, and they decreased in June at the onset of the LDS, reflecting the variation in eco-climatic conditions across the year. GLMM results of R S against EVI with grazing sites as random effect revealed a positive relationship (p < 0.05) with the EVI explaining about 60% of the variation in R S across the year ( Fig. 4b; Table 5). Furthermore, including the EVI and W S in the model led to the explanation of 76% of the variability in R S (Table A2).

The effects of soil properties on soil CO 2 efflux
Soil organic carbon (SOC) concentrations were significantly higher at NG, LG and M-LG compared to OG and HG, with the corresponding total nitrogen (TN) content exhibiting a similar trend (Table 6). Bulk density (BD) increased with increased grazing. OG showed a higher (approximately 10%) BD (p < 0.05) than the NG, and declined with a lower grazed sites. However, the other soil properties did not differ significantly between the sites and thus were minor in explaining the difference in R S between the sites. AIC 2377, R 2 = 0.71, Significance code: p < 0.05 = *, p ≤ 0.01 = **, p ≤ 0.001 = ***. LGmoderately to lightly grazed, LG -lightly grazed and NG -no grazing) over the study periods.

Discussion
In southern Kenya, we studied soil CO 2 flux (R S ) at five grassland sites under different levels of grazing and one enclosed control area. Overall, R S values from all grazing sites were ranging from 10 to 356 mg C m − 2 h − 1 . These results are within the range of previous studies in the same region (Wachiye et al., 2020(Wachiye et al., , 2021 and from other land use types in tropical savanna ecosystems similar to our study (Ardö et al., 2008;Livesley et al., 2011). Low R S values from tropical savanna ecosystems have often been related to a low SOC (Grace et al., 2006;Scholes et al., 1997), as observed in this study (Table 6).

Effects of grazing intensity on vegetation, soil and soil respiration
Grazing can alter vegetation cover and biomass, specifically grass cover, and soil properties, depending on the intensity of grazing and management practices, which in turn can affect on R S (Cao et al., 2004;QingJi et al., 2009;Wang et al., 2016). Few data are available for a detailed comparison of R S under different levels of grazing in East Africa (Oduor et al., 2018). In this study, difference in grazing sites were significant in explaining 60% of the spatial variation in R S . Generally, R S were significantly lower in heavier grazed OG than the less grazed sites (HG, MG, M-LG and LG), which agrees with previous similar studies in other savannas in Kenya (Oduor et al., 2018;Ondier et al., 2020), and also studies by Cao et al. (2004) and Abagandura et al. (2019). In their study, the authors attributed this to the lower SOC at intensively grazed sites, and we also found a significantly lower (p = 0.02) SOC at the OG in comparison to the less grazed sites (Table 6).
In the OG, vegetation cover was lowest on average (based on EVI, which is a proxy for vegetation cover), and increased to MG dropped in M-LG and LG, suggesting that heavy grazing greatly reduced aboveground vegetation. Mean AGB was also lowest at OG and higher at less grazed sites (Fig. 6a). This was supported by a negative relationship  AIC 2377, R 2 = 0.60, Significance code: p < 0.05 = *, p ≤ 0.01 = **, p ≤ 0.001 = ***.
In the private wildlife sanctuary THWS hosting M-LG and LG, the intensity of grazing is low, since only wild mammals are allowed to graze and browse, causing less damage to the vegetation cover and biomass compared to that LUMO community grazing area (OG, HG and MG), which is frequented by both livestock and wildlife ( Fig. 5b; Table A1). Comparatively to THWS sites, OG and HG wildlife cameras captured large herds during study period ( Fig. 5b; Table A1). OG and HG also had large patches of bare soil, which indicated heavy grazing. However, the number of animals captured at the LG site was lower than at MG (Table A1) but LG displayed a lower EVI and R S than MG (Fig. 6b). We noted that most of the animals at LG were buffalos (Syncerus caffer), which might have caused more destruction on the grass cover through trampling than other animal species at the other sites. A previous study also showed that megaherbivores, such as elephants and buffalos were the cause of damage on woody biomass in THWS (Amara et al., 2020).
Heavy grazing can shift plant species composition by replacing highly palatable grass species with undesirable plants (Klumpp et al., 2009). This might be the case at OG, where Sodom apple (Solanum incanum L) plants are densely spread throughout the site. Although native to East Africa, Sodom apple is considered among the 100 worst invasive plants in the region (Pringle et al., 1785). At HG, we observed Indigofera tinctoria, which was extremely dense, especially during the wet season. As a species that is undesired by cattle, it displayed a high EVI at this site between January and May (Fig. A2). This shift in plant species can affect the soil organic matter and root biomass (Klumpp et al., 2009), and may also respond differently to the onset of the rainy season compared to grasses that are known to respond quickly than shrubs (Merbold et al., 2009), thus affecting overall R S .
Regarding the effect of dung and urine deposition on R S , we cannot draw any conclusions, since animal excreta were not visible in any of our chambers during the measurement of R S . In addition, soil pH and TN did not differ significantly between the sites. Nonetheless, BD was higher at OG and dropped with decreased grazing, thus providing an additional explanation for the lower R S at OG (Bahn et al., 2008;Cao et al., 2004). Increased grazing increases BD through animal trampling, thus reducing water infiltration and gas diffusion in and out of the soil (Jafari et al., 2014;Li et al., 2008;du Toit et al., 2009;Oduor et al., 2018). This in turn reduces the diffusion of R S in soil, resulting in a lower mean R S at the OG site compared to the other grazed sites (Cao et al., 2004).
With grazing and browsing prevented due to animal exclusion by fencing, NG showed signs of recovery of the degraded grazing land, since this site had a higher cover of shrubs and small trees compared to the other sites (mean EVI was the highest in NG). The forest restoration enclosure was established in 2010 in a similar landscape to the M-LG site. Although cattle intruded into NG in September, this was not enough to cause a notable change in EVI, since the disturbance only lasted for a few days.
Site NG recorded the lowest R S , which was opposite to our expectations and contradicts several earlier studies that have reported higher R S values at ungrazed sites (Oduor et al., 2018;Jia et al., 2006). According to these studies, higher R S at ungrazed sites was explained by higher vegetation cover, biomass, W S and SOC, and reduced BD (Yan et al., 2015;Wang et al., 2016), which was also true in this study, however, due to the resultant shading, there was less herbaceous understorey compared to the other sites. Shading might have interfered with the growth and regrowth of herbaceous vegetation below the canopies, and therefore, there is a likelihood that root production in the other open grazed areas was higher (Janssens et al., 2001), although we cannot confirm this since we did not measure root biomass. Additionally, due to the higher trees and shrubs canopy coverage, the NG site also recorded the lowest soil and surface temperatures overall (Fig. 3d &  Fig. A1). Hence, while W S was adequate to support biological activity, T S became an important determinant of R S (Carbone et al., 2011).
After the wildfire on 8-9 August at the HG, vegetation and animal use remained low during September and October but increased once the rainy season began in November (Fig. 5a). Burning releases most carbon, destroys aboveground vegetation, thus affecting carbon allocation and root respiration. Previous studies reported lower R S values in burned areas a few weeks after fire events in Congo (Castaldi et al., 2010), Ethiopia Michelsen et al., 2004) and Brazil (Poth et al., 1995). In the present study, we did not detect any effects of the wildfire on R S , even after the onset of the rainy season. This agrees with studies carried out in Nigeria (Adedeji, 1983) and South Africa (Zepp et al., 1996). Since the fire occurred in the dry season, most grasses were dry and W S was already very low (below 10%), which may have overridden any effects of the fire on R S .

Temporal variation in soil respiration
Temporal variation in R S at all sites showed a clear bimodal pattern that mirrored the variation in W S and rainfall over the year. The highest R S values were observed during the two rainy seasons (LWS and SWS), and also in the expected short dry season (SDS) due to sporadic rainfall that occurred in January, which resulted in relatively high W S at all sites. The lowest R S was observed in the long dry seasons (LDS) when W S was the lowest at all sites. Overall, W S explained about 71% ( Fig. 4; p < 0.05) of the seasonal variation in R S , which is within the range of previous reports Wachiye et al., 2020). Other African savannas have also reported higher R S results during rainy seasons due to an increase in W S (Castaldi et al., 2010;Livesley et al., 2011;Wachiye et al., 2020;Pinto et al., 2002). An increase in W S increases microbial activity, which speeds up the decomposition of soil organic matter and stimulates plant and root growth, thus increasing root respiration (Carbone et al., 2011;Orchard and Cook, 1983). On the other hand, T S showed no significant influence on seasonal variation in R S due to its narrow temporal variation, and the same result has been observed in other studies of savannahs (Castaldi et al., 2006;Grover et al., 2012;Livesley et al., 2011;Wachiye et al., 2020).
Furthermore, higher EVI values were observed during the rainy seasons at all sites, indicating a greater amount of vegetation and thus more root respiration since grasses sprout more rapidly with the first rains (Merbold et al., 2009). Results from the GLMM model revealed a positive relationship (ca. 60%; p < 0.05; Fig. 4b) between R S and the EVI at all sites. The inclusion of W S in the model further improved the relationship, explaining 76% of the variation in R S . According to these findings, seasonal variation in R S was therefore strongly dependent on the vegetation status and W S at high grazing intensity sites, as well as over a broader area.
During the wet season, the differences in R S between the sites HG, MG, M-LG and LG due to grazing were more apparent. Similarly, EVI significantly differed, with the lowest value at OG and values increasing as a function of decreasing grazing intensity. We assume that vegetation recovered more quickly in less grazed areas because the vegetation was not as heavily damaged by animals grazing and trampling as in heavily grazed areas. Moreover, OG and HG sites displayed more large patches of bare soil. Furthermore, in heavily grazed areas, the higher soil bulk density due to animal trampling, increased evaporation and soil erosion of nutrients lead to drier and nutrient-poor soils, which may limit plant establishment and growth (Belnap et al., 2009;Wang et al., 2016).
At OG, EVI increased at the onset of the rainy season in March but started to gradually decline as the number of animals started to increase from April onwards (Fig. 5a). While conducting gas sampling exercises during this period, we often encountered large herds of livestock. Large patches of bare soil were more visible throughout the overgrazed areas, especially in the dry season. According to the management of the LUMO community grazing area, local herders not only bring their livestock there but also neighboring communities use the area for their livestock during the dry season. LUMO is a common location for herders to move their cattle to find better forage and water when they have exhausted forage on their land and have most of their land under crop production.
In the dry season, a drop in W S causes senescence of the grass layer, which in turn reduces root respiration (Pinto et al., 2002). The drying of grasses and other plant tissues was evident in the progressive reduction in the EVI, as well as a decline in W S and R S . Reduced W S can suppress root growth, as a decline in W S affects the diffusion of organic C substrate, resulting in microbial stress and reduced plant growth (Li et al., 2018;Orchard and Cook, 1983).
Soil W S was slightly higher at the NG in the dry season an indication that shading provided by shrubs and bushes allowed soils to cool, minimizing evaporation, as NG recorded the lowest T S (Fig. 3e) and T Surf (Fig. A1). During the latter months of the dry season in September and October, R S was slightly highest in NG. At all grazing sites and even in NG, grass cover had dried out by this time, reducing R S , but since the roots of shrubs and trees in NG are deeper than grass roots, they could tap into deeper soil moisture (Carbone et al., 2011). Microbial activity is highest and most variable in the upper soil layers, which are first wetted up and first to dry down, but deeper roots can access to water reserves in deeper soil, which takes longer to exhaust and therefore and so remain active during the dry season (Carbone et al., 2011).

Conclusion
This study examined soil carbon dioxide flux (R S ) under various grazing intensities in the savannah lowlands of southern Kenya. Overall, results indicated that R S decreased with greater grazing intensity, because higher grazing intensity reduced vegetation cover, increased bulk density, and reduced herbaceous aboveground biomass, resulting in a lower R S . Additionally, SOC was lower on highly grazed sites compared to less grazed sites, supporting these findings. The variations in volumetric soil water content (W S ) and vegetation cover (EVI) were prominent factors in explaining temporal variations in R S during the study period. Results presented here could be integrated into the management of savannah grasslands in East Africa to identify the potentially adverse consequences of expanding livestock numbers and create climate-smart pastoral practices.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability
The data associated with the manuscript can be obtained from https://figshare.com/articles/dataset/Effects_ of_livestock_and_wildlife_grazing_intensity_on_soil_carbon_ dioxide_efflux_in_the_savanna_grassland_of_Kenya/14791380.

Acknowledgments
We acknowledge the management of the Taita Hills Wildlife Sanctuary and the LUMO Community Wildlife Sanctuary for allowing us access to the study sites. We thank the Taita Research Station of the University of Helsinki for technical and fieldwork support and the Mazingira Centre of the International Livestock Research Institute for technical support with the laboratory work. Specifically, we would like to thank Peter Mwasi, Darius Mwambala Kimuzi and Mwadime Mjomba for helping in data collection, as well as Margaret Muthoni and Collins Ouma for assistance with the laboratory work.

Financial support
The Schlumberger Foundation under the Faculty for the Future program funded this study. The work was conducted under the project "Environmental sensing of ecosystem services for developing a climatesmart landscape framework to improve food security in East Africa", funded by the Academy of Finland (318645), and "Earth observation and environmental sensing for climate-smart sustainable agropastoral ecosystem transformation in East Africa", funded by the European Commission, DG International Partnerships under the DeSIRA program (FOOD/2020/418-132) for PP. A research permit from NACOSTI (P/18/ 97336/26355) is acknowledged. LM acknowledges the CGIAR Fund Council, Australia (ACIAR), Irish Aid, the European Union, IFAD, the Netherlands, New Zealand, the UK, USAID and Thailand for funding the CGIAR Research Program on Livestock and Programme of Climate Smart Livestock (PCSL, 2017.0119.2), funded by the German Federal Ministry for Economic Cooperation and Development (BMZ issued through GIZ).

Appendices.
See Figs. A1 and A2 and Tables A1 and A2. LGmoderately to lightly grazed, LG -lightly grazed and NG -no grazing) from January to December 2020. Fig. A2. Variation in EVI in each grazing sites (OG -overgrazed, HG -heavily grazed, MG -moderately grazed, M-LGmoderately to lightly grazed, LG -lightly grazed and NG -no grazing) from January to December 2020.

Table A1
Intensity of animal use in relation to grazing intensity (OG -overgrazed, HG -heavily grazed, MG -moderately grazed, M-LGmoderately to lightly grazed, LG -lightly grazed and NG -no grazing) from January to December 2020 as captured by wildlife cameras.