Tropical Montane Forest Conversion Is a Critical Driver for Sediment Supply in East African Catchments

Land use change is known to affect suspended sediment fluxes in headwater catchments. There is however limited empirical evidence of the magnitude of these effects for montane catchments in East Africa. We collected a unique 4‐year high‐frequency data set and assessed seasonal sediment variation, water pathways, and sediment response to hydrology in three catchments under contrasting land use in the Mau Forest Complex, Kenya's largest tropical montane forest. Annual suspended sediment yield was significantly higher in a smallholder agriculture‐dominated catchment (131.5 ± 90.6 t km−2 yr−1) than in a tea‐tree plantation catchment (42.0 ± 21.0 t km−2 yr−1) and a natural forest catchment (21.5 ± 11.1 t km−2 yr−1) (p < 0.05). Transfer function models showed that in the natural forest and the tea‐tree plantations subsurface flow pathways delivered water to the stream, while in the smallholder agriculture shallow subsurface and surface runoff were dominant. There was a delayed sediment response to rainfall for the smallholder agriculture and the tea‐tree plantations. A slow depletion in sediment supply suggests that the wider catchment area supplies sediment, especially in the catchment dominated by smallholder farming. In contrast, a fast sediment response and depletion in sediment supply in the natural forest suggests a dominance of temporarily stored and nearby sediment sources. This study shows that the vegetation cover of a forest ecosystem is very effective in conserving soil, whereas catchments with more bare soil and poor soil conservation practices generated six times more suspended sediment yield. Catchment connectivity through unpaved tracks is thought to be the main explanation for the difference in sediment yield.


Introduction
The conversion of native ecosystems to agriculture leads to the degradation of soil properties (Githui et al., 2009;Morgan, 2005;Owuor et al., 2018), which can increase soil erosion rates (Bruijnzeel, 2004). Soil erosion does not only deplete fertile topsoil from agricultural land but also leads to water quality deterioration caused by an increase in fine suspended sediments (Brown et al., 1996;Horowitz, 2008;Quinton et al., 2001). Hence, suspended sediment physically affects the fluvial network, (Owens et al., 2005) polluting drinking water for communities, livestock and wildlife, and impacting downstream water reservoirs and hydropower generation because of the accumulation of the sediments (Foster et al., 2012;Mogaka, 2006;Wangechi et al., 2015). Additionally, pollutants such as pesticides (Brown et al., 1996) and nutrients (phosphorus or nitrogen) (Fraser et al., 1999;Horowitz, 2008;Quinton et al., 2001) can be attached to sediments and can harm aquatic biota (Gellis & Mukundan, 2013;Kemp et al., 2011;Owens et al., 2005). The increase in nutrient concentrations can result in eutrophication of water bodies (Foy & Bailey-Watts, 1998;Hilton et al., 2006;Mogaka, 2006).
Streams in montane headwaters are major contributors to suspended sediment yield because of the steep terrain that leads to a strong hillslope to channel connectivity (Grangeon et al., 2012;Morris, 2014;Wohl, 2006). Significant increases in suspended sediment yield can be expected from tropical montane headwater catchments, which are heavily affected by deforestation followed by cultivation of erosion prone areas, often without soil conservation measures (Ramos-Scharrón & Thomaz, 2016;Wohl, 2006). Land use change may impact catchment hydrology and runoff mechanisms in the tropics (Muñoz-Villers & McDonnell, 2013;Ogden et al., 2013), which are key processes determining sediment yields.
In East Africa, land use change in tropical montane forests is mainly driven by scarcity of arable land (Pellikka et al., 2004) with the most fertile lands located in the proximity of natural ecosystems (Krhoda, 1988). The Mau Forest Complex exemplifies this case, with one quarter of the forest converted to agricultural land over the last four decades (Brandt et al., 2018), in addition to the clearances at the beginning of the twentieth century to establish commercial tea plantations (Binge, 1962). Due to its location in the highlands of Kenya, the Mau forest is a critical catchment area for the country; it is the headwater to twelve rivers, one of which is the Sondu River, a tributary of Lake Victoria (Mogaka, 2006;UNEP, 2005).
Eutrophication and sedimentation are major environmental problems affecting Lake Victoria, where sediments are estimated to accumulate at a rate of 2.3 mm yr −1 (Verschuren et al., 2002). Although authorities in Kenya acknowledge the need to reduce sediment pollution, the linkages between land use change and changes in sediment dynamics in the headwater catchments are not well quantified (Nyssen et al., 2004;Vanmaercke et al., 2010Vanmaercke et al., , 2014. There are limited data on sediment export for montane catchments in Sub-Saharan Africa in general and in East Africa in particular (Ntiba et al., 2001;Walling & Webb, 1996). Clearly, this is a significant gap in our knowledge of these environments that requires empirical measurements to address it. Not only will these measurements improve our understanding of these underresearched environments, but they will also assist in the development of targeted soil and water conservation strategies to disconnect sediment source areas in the upper catchments of the Mau Forest from the fluvial system and downstream environments, including Lake Victoria.
The overall aim of this study was to elucidate the spatial and temporal dynamics of suspended sediment and to quantify suspended sediment loads in tropical montane streams under contrasting land uses using a 4 year high-temporal resolution data set. The main objectives were (a) to quantify rainfall, streamflow, and suspended sediment transport dynamics, (b) to compare the seasonal responses in suspended sediment yield, (c) to assess the timing of the response of suspended sediment to rainfall and discharge, and (d) to improve our understanding of the dominant water flow pathways.

Catchment Characteristics and Site Description
The three catchments studied are located in the headwaters of the Sondu River Basin (3,470 km 2 ) in the Western Highlands of Kenya ( Figure 1). Each catchment is dominated by a distinct land use: (1) natural forest (NF; 35.9 km 2 ), (2) smallholder agriculture (SHA; 27.2 km 2 ), and (3) tea-tree plantations (TTPs; 33.3 km 2 ). The Sondu River drains into Lake Victoria, which is the second largest fresh water lake in the world, an important water and economic resource for five countries and one source of the Nile River.
The three catchments (Table 1) are characterized by steep hillslopes with a maximum slope gradient of 72% in the NF catchment. The streams are mostly firstand second-order perennial streams that merge together to form the River Sondu (a sixth-order stream). The rainy seasons are bimodal with a long rainy season between March and June and a short rainy season between October and December with a continued intermediate rainy season between the two wet seasons. Mean annual precipitation is 1,988 ± 328 mm (period 1905-2014) with rainfall peaks in April and May (>260 mm month −1 ). January and February (<95 mm month −1 ) are the driest months.
The NF catchment is located in the South-West Mau block of the Mau Forest Complex. The Mau Forest is an afromontane mixed forest dominated by indigenous broad-leafed evergreen trees and shrubs with 10.1029/2020WR027495

Water Resources Research
STENFERT KROESE ET AL. a complex vegetation pattern. Riparian forests with a mixture of indigenous vegetation are present throughout the catchment. The NF catchment is characterized by high infiltration rates with the occurrence of shallow to deeper subsurface water pathways, whereas the TTP and the SHA catchments have lower infiltration rates with a dominance of surface runoff (Jacobs, Timbe, et al., 2018;Owuor et al., 2018) (Table 1).
In the SHA catchment, subsistence farmers grow maize interspersed with beans, potatoes, millet, and cabbage on small farms (~1 ha). Small-scale tea plantations, eucalyptus (Eucalyptus spp.), cypress (Cypressus spp.), and pine (Pinus spp.) woodlots are interspersed with crop fields and grazing land (Table 1). A combination of hand weeding, hoeing, and herbicides is used for weed control. Bamboo (Bambusa spp.) is generally found around natural springs. The whole catchment is connected by a dense network of unpaved tracks either bare or sparsely covered by grass; stream crossings rarely have bridges. The heavily traveled unpaved tracks have commonly become eroded gullies (Figures 2a and 2b), which run down the slope to rivers, connecting surface runoff from surrounding fields with the stream network (Figures 2c-2e). Cattle entrance points to the stream are generally highly disturbed and have degraded riverbanks (Figures 2d-2f). The natural riparian vegetation is in many areas replaced by eucalyptus woodlots or small bushes. In some places, riparian wetlands are found.
The TTP catchment has tea fields alternated with Eucalyptus spp. and Cypress spp. woodlots that are used for fuelwood at the tea factories. Some of the tea companies use mulch and rows of oat grass between rows of tea to control soil erosion during the establishment of new tea bushes. Herbicides are commonly used to control weeds. Cover crops with mature tea trees during the establishment of a new tea crop, terracing and sited cutoff drains are also used within the catchment to control soil erosion. The catchment is covered by a network of well-maintained paved and unpaved roads, linked to drainage systems, such as open culverts along the roads that connect them to the streams. The riparian vegetation includes a mix of indigenous tree species that cover densely the ground and form a buffer of approximately 30 m ( Table 1).
The study area is composed of folded volcanics from the early Miocene times. Porphyritic phonolites, a member of the sequence of basic and intermediate lavas (igneous rocks) are predominant in the study area (Binge, 1962), where pyroclastic rocks cover the upper part of the catchment (ISRIC, 2007). The study area comprises well-drained, very deep (>1.8 m) dark-red and dark-brown loamy soils (Sombroek et al., 1982), with moderate to high amounts of organic matter under the forest cover (Dunne, 1979).

Automated Hydrological and Sediment Monitoring
This study uses a 4 year data set (January 2015 to December 2018) on rainfall, discharge, and turbidity with a 10 min resolution ( Figure 1). A radar sensor (VEGAPULS WL61, VEGA Grieshaber KG, Schiltach, Germany) collected continuous water level measurements. Water level ("stage") was used to determine stream discharge based on a site-specific second-order polynomial stage-discharge relationship . The calibration was checked over a wide range of stream flows using salt-dilution gauging (Shaw et al., 2011), an Accoustic Doppler Velocimeter (ADV; FlowTracker, SonTek, San Diego CA, USA) or an Acoustic Doppler Current Profiler (ADCP; RiverSurveyor S5, SonTek, San Diego, USA) depending on river size and discharge . Specific discharge (mm day −1 ) was determined by integrating instantaneous discharge taken at 10 min intervals over a day and relating it to the catchment area. Precipitation was measured using eight automatic tipping bucket rain gauges calibrated to measure cumulative rainfall every 10 min with a 0.2 mm resolution (five tipping bucket rain gauges: Theodor Friedrichs, Schenefeld, Germany, and three weather stations: ECRN-100 high-resolution rain gauge). Using Thiessen polygons, we estimated the weighted contribution of rainfall of every tipping bucket in each catchment. A more detailed description of the study sites and instrumentation can be found in . Turbidity was measured in situ as a surrogate for suspended sediment concentrations using a ultraviolet/visible (UV/Vis) spectroscopy sensor (spectro::lyser, s::can Messtechnik GmbH, Vienna, Austria). Turbidity is measured in formazin turbidity unit (FTU) by transmitting a beam of light to an optical receptor. With an increase in water turbidity the transmission of light decreases. To calculate sediment concentrations, a site-specific turbidity-suspended sediment calibration was established (section 2.3.1). Before each turbidity measurement, the window of the sensors was automatically cleaned by compressed air to remove any interfering particles. The sensors were additionally cleaned manually on a weekly basis using a specific cleaning agent recommended by the manufacturer to reduce biofouling on the measurement window and by manually removing debris and sediment.

Water Resources Research
STENFERT KROESE ET AL.

Calibration, Quality Assurance, and Analysis 2.3.1. Sediment-Turbidity Rating Curve
We used a site-specific, ex situ incremental suspension calibration to convert long-term turbidity records into an estimate of instantaneous suspended sediment concentrations (mg L −1 ). A river water-sediment suspension with 16 to 18 concentration increments was established to simulate changing stream water suspended sediment concentrations occurring from low flow conditions (minimum 0 mg L −1 ) to storm events (maximum 4,607 mg L −1 ). The sediment suspension consisted of fine suspended sediment collected from sediment traps (time-integrated Phillips samplers) and fine soil material mixed with turbid river water collected during storm events. To ensure that only the clay size fraction remained in the suspension, the sediment suspension was decanted twice after the settling time for coarse particles (particle size >2 μm) had elapsed (Stokes law: 98 s). The spectro::lyser probes from each monitoring station measured each concentration increment starting with river water representing low flow conditions (0 to 8 FTU). Small quantities of the synthetic sediment suspension were added at each concentration increment until the maximum measurable turbidity of 1,500 FTU was reached. The exact concentration was then determined gravimetrically from a 250 ml subsample at each increment. Total suspended sediment (TSS) load was determined by multiplying suspended sediment concentration by discharge. Suspended sediment yield was calculated by integrating the sediment load over time and relating it to the catchment area. The sediment mass is reported in tonnes (t = Mg) to conform with other published values.

Data Quality Assurance
Quality assurance of the turbidity, discharge, and precipitation data set was performed in two different ways. First, during equipment maintenance and manually downloading of the data any observed anomalies were recorded in a log book. Potential causes of anomalous values included (i) sensor above water level, (ii) turbidity sensor completely buried by deposited sediment during storm periods, (iii) biofilm or other phenomena on the measurement window due to malfunctioning of automatic cleaning with compressed air, (iv) measurement gaps due to incidents of power supply failure, or (v) counting of number of tips by the rain gauges restricted by blocked funnel or spiderwebs. The readings for these periods were flagged with Not-a-Number (NaN).
After anomalous values were replaced by NaN, the median absolute deviation (MAD) was used to detect local outliers. The MAD has the following form: where x i is the whole data set, M i1 is the median of the data set, and M i2 is the median of the absolute deviation from the data set from its median. The constant b estimates the standard deviation and was set to 1.4826 for normal distribution (Leys et al., 2013).
A moving window of k measurements around observation x i at time t i was used to detect local outliers with where a = 6 is the threshold for outlier selection, M j,i is the median, and the MAD j, i is the MAD for x j , while the moving window k was set to 16. Missing sediment data were interpolated using a linear function.

Data Analysis
All data were tested for normality with the Shapiro-Wilk test. We tested significant differences on suspended sediment, rainfall, and discharge values among the different land uses using Kruskal-Wallis test for analyses of variances. To detect the significance of the effect of land use on the hydrological and sedimentological parameters, and within and among seasons on suspended sediment load, we used the pairwise Wilcoxon rank sum test.
Five seasons, dry season, start of long rains, long rains (long rainy season), intermediate rains (season between the long and short rainy season), and short rains (short rainy season), were identified to calculate their contribution to annual suspended sediment yield . The periods were chosen based on exceeding a threshold of monthly specific discharge for each catchment. The seasons for each year vary in length and timing due to variations in the onset of the rains and monthly streamflow ( Figure 3).

Modeling Flow Pathways
A linear continuous-time (CT) transfer function (TF) model with rainfall-runoff nonlinearity was used to identify the dynamics that explain the response of water flow pathways to rainfall within a catchment by conceptualizing a Single-Input, Single-Output (SISO) system (Young & Garnier, 2006). These types of models are equivalent to systems of linear differential equations and can be applied in numerous mass and energy transport as well as chemical or biochemical process applications, including flow and sediment delivery models (Chappell et al., 2006). The TF modeling process follows the Data-Based Mechanistic (DBM) modeling philosophy, searching through a range of model structures, ordering them according to statistical criteria, then retaining models that have a physical explanation (Young & Beven, 1994). The DBM approach produces parsimonious models describing rainfall-runoff relationships that include very few tunable parameters (Lees, 2000). Hourly time series of rainfall (mm hr −1 ) was used as input and the corresponding hourly time series of discharge (m 3 s −1 ) as output (the units conversion is absorbed into the model coefficients, in case of Equation 6-into coefficient b 0 ). These parameters have hydrological interpretation, describing hydrological pathways, so while these are not strictly hydrological models derived from the process models, they still apply to hydrological systems (Beven, 2012).
Hydrological processes are known to be nonlinear (Beven, 2012), with the effectiveness of rainfall (amount of rainfall converted into discharge) dependent upon the state of saturation of the catchment. Therefore, the rainfall-runoff nonlinearity is modeled using the Hammerstein model structure, with the input (rainfall) transformed using a nonlinear function into what is termed "effective rainfall," which then drives the linear dynamics of the transport process model. This study uses a power law relationship between measured rainfall and effective rainfall as surrogate for soil moisture to translate rainfall to effective rainfall (see Text S1 in the supporting information for more details). Effective rainfall is the dynamically changing proportion of rainfall representing the volume of streamflow generated after soil moisture storage is deducted from the total rainfall (Beven, 2012). The RIVCBJ (Refined Instrumental Variable Continuous Time Box-Jenkins Identification, for continuous models, Young & Garnier, 2006) algorithm was used to estimate model parameters. RIVCBJ is a component of the CAPTAIN toolbox, which runs within MATLAB® (Taylor et al., 2007).
The linear CT TF model has the following form: In Equation 3 Y(s) denotes the Laplace transform of the output signal as hourly streamflow (m 3 s −1 ), U(s) is the Laplace transform of the input signal hourly rainfall (mm hr −1 ), E(s) are the model residuals, and e −sτ is the Laplace transform of time delay τ representing the pure time delay (as opposed to the dynamic lag resulting from the system's dynamics) in time units between the input and output signals.
In this study, up to third-order models were tested for all the sites and model fit was evaluated according to the coefficient of determination (R 2 t ) (also known as the Nash-Sutcliffe efficiency) (see Text S2) and the Young Identification Criterion (YIC) (see Text S3 and S4). First-order TF models were selected for the three catchments, where each system has a different depletion time, determined by its time constant (TC). A firstorder CT TF model is written as follows: where 1/a 1 is the TC and the parameter b 0 /a 1 represent the steady state gain (SSG) of the hypothetical pathway of rainfall through the catchment with a and b as the dynamic response characteristics. The TC reflects the response between the input (rainfall) and the output (runoff or streamflow) (Young & Garnier, 2006). Linear CT TF models were identified for each year between 2015 and 2018 for all three catchments.

Sediment Response to Hydrological Variables
We used the cross-correlation function (CCF) to identify the statistical correlation between two sets of time series at different time lags (Lee et al., 2006;Mayaud et al., 2014). Rainfall and discharge time series were cross correlated with the suspended sediment concentration time series. The peak response time 10.1029/2020WR027495

Water Resources Research
STENFERT KROESE ET AL.
between either precipitation or discharge to sediment concentrations was calculated as the delay time in time lags together with its correlation strength between these variables. CCFs were calculated as follows: where CCC(τ) is the cross-correlation coefficient at time lag τ, τ = 0, ± 1 ± 2 … ± m between the two time series (sampled every 10 min), where x i is observed rainfall or positive derivative of discharge at sample number i and y i − τ is the suspended sediment concentration at sample number i − τ, x is the mean rainfall or positive derivative of discharge, y is the mean suspended sediment concentration, σ x is the standard deviation of rainfall or estimated positive derivative of discharge, and σ y is the standard deviation of suspended sediment and n is the number of data points. At the 95% confidence interval, lag-time correlations are significant when CCC(τ) exceeds the standard error of 2/√N, where N is the length of the data set (Diggle, 1990). The positive derivative of discharge, that is, the estimated rate of change on the rising limb of the hydrograph, was selected because the main sediment pulses are mostly generated during the rising limb (Alexandrov et al., 2003;De Girolamo et al., 2015). A similar derivative effect has been observed in dynamic sediment load models by Walsh et al. (2011). The CCF analysis was carried out for each year between 2015 and 2018 for all three catchments.

Hydrological Response of the Three Catchments
Mean annual rainfall for the study period was 1,842, 1,730 and 1,554 mm yr −1 with maximum hourly rainfall over the whole observation period of 37.4, 33.1, and 27.5 mm hr −1 for the NF, TTPs, and SHA catchments, respectively. The wettest year for the SHA catchment was 2018 with 1,823 mm yr −1 of rainfall, while for the NF and TTPs precipitation was highest in 2015 with 1,986 and 1,928 mm yr −1 , respectively. The annual mean specific discharge was 632 ± 157, 610 ± 153, and 621 ± 224 mm yr −1 for the NF, TTPs, and SHA catchments, respectively. The catchment runoff coefficient was similar for the NF and the TTPs with a mean of 0.34 and 0.35, respectively, and 0.39 for the SHA (Table 2).
Discharge in all catchments was flashy and varied seasonally. Rising limbs were generally steep and had variable falling limbs depending on event size. The highest discharge peaks were measured during the long rainy seasons between April and July in 2015, 2016, and 2018. In contrast, 2017 was the driest year with a late onset of the rains and the highest discharge peaks between August and November for the SHA and the NF catchments. However, in the TTP catchment the rains started in May lasting until November 2017, resulting in discharge peaking in May and September 2017. High discharges were also recorded in January 2016 because the 2015 rains continued through November and December and were followed by an unusually wet January (Figure 4).

Relationship Between Turbidity and Suspended Sediment Concentration
We obtained one rating curve for all three catchments to predict suspended sediment concentration from the measured turbidity values. A linear model provided the best fit between the in situ turbidity and suspended sediment concentrations, and there was no significant difference between slopes for each site-specific calibration (p value > 0.1). The intercept of the linear model was forced through the origin to prevent negative sediment concentrations at low turbidity, yielding an equation of the form TSS = 2.4*turbidity (R 2 = 0.98, p value < 0.001, n = 50; Figure 5). This equation was used to convert the turbidity data to suspended sediment concentrations.

Water Resources Research
STENFERT KROESE ET AL.

Suspended Sediment Dynamics
Sediment yield for the NF was lower than for the other two catchments (Figure 4). The sedigraph of the NF had smaller peaks with a short event time, whereas the SHA and TTPs showed a steep increase followed by a flat recession with a long event time. The sedigraph of the SHA catchment showed a flashy sediment response to rainfall and increased discharge. The maximum sediment yield in the NF catchment was 0.2 t km −2 hr −1 , followed by the TTPs with 1.5 t km −2 hr −1 and the SHA with a maximum sediment peak of 1.6 t km −2 hr −1 (Figure 4). The mean annual suspended sediment yield was significantly higher for the SHA catchment (131.5 ± 90.6 t km −2 yr −1 ) than for the TTPs (42.0 ± 21.0 t km −2 yr −1 ) and the NF (21.5 ± 11.1 t km −2 yr −1 ) (p < 0.05) ( Table 2). The lowest mean suspended sediment concentration (33.8 ± 73.8 mg L −1 ) was also lowest for the NF catchment, followed by the TTPs (47.4 ± 90.7 mg L −1 ). Concentrations were 3 to 4 times higher at the outlet of the SHA catchment (128.6 ± 233.4 mg L −1 ) than at the other catchments (p = 0.04). The daily mean suspended sediment load from the NF was the lowest, followed by the TTPs and the SHA (2.1, 3.9 and 10.0 t day −1 , respectively). The TSS load for the entire study period (2015-2018) for the SHA (14,308 t) was 4 times higher than that of the NF (3,083 t), whereas sediment load for the TTPs (5,595 t) was only twice that of the NF. These loads represent a mean of 771, 1,399, and 3,577 t yr −1 for the NF, TTPs, and SHA, respectively. Suspended sediment yield increased from 2015 to 2018 in the NF and SHA catchments. In the TTPs, a similar sediment and rainfall pattern was observed with a decline from 2015 to 2017 and then an increase again in 2018 ( Table 2). The NF had the longest period of missing data lasting for 95 days in November 2015 to February 2016, followed by a shorter gap in the SHA of 50 days from March to April 2015 and the TTPs had the shortest period of missing sediment data of 13 days between March and April 2017. Besides these periods, minor gaps were usually of less than 24 hr with a total of missing sediment data of 7% for the NF, 2% for the TTPs and 4% for the SHA catchments between 2015 and 2018.

Seasonal Variations in Suspended Sediment
During the study period, suspended sediment yield showed pronounced seasonal variability, with most sediment being transported during the long rains in all catchments, and the highest monthly yields being recorded for the SHA catchment ( Figure 6). Overall, more than half of the sediment yield (45-52%) was attributed to the long rains, which cover less than one third of the year. The sediment contribution during the long rains, intermediate rains, and short rains in the SHA was significantly greater than in the NF and TTPs (p < 0.05). For the NF the streams carried significantly more material during the long rains (mean yield of 4.3 ± 1.8 t km −2 month −1 ) than during any other season. In the TTPs and SHA, the sediment yield for the long rains (mean 10.9 ± 6.9 and 30.7 ± 10.6 t km −2 month −1 , respectively) differed significantly from the dry season (mean 0.4 ± 0.2 and 0.6 ± 0.3 t km −2 month −1 , respectively), the intermediate rains (mean 1.1 ± 0.6 and 5.4 ± 3.3 t km −2 month −1 , respectively) and the short rains (mean 3.2 ± 2.2 and 23.9 ± 26.1 t km −2 month −1 , respectively), while there was no difference between the sediment yield for the start of the long rains (mean 5.7 ± 6.5 and 7.8 t km −2 month −1 , respectively) and the long rains (p < 0.05).

Flow Pathways and Streamflow Dynamics
We compared hydrological flow pathways for each catchment as these are key in delivering sediments to the streams. Rainfall-runoff response was modeled over a continuous period of one year for each monitoring year 2015-2018 (see Figure S1). In all three catchments, first-order linear models (Equation 6) were selected because these had the highest coefficients of determination (R 2 t Þ ranging between 86% and 93% and explained the data with the most negative YIC ranging between −12.5 and −10.4. The TTPs had a lower model performance in 2017 compared to the other years with a R 2 t value of 54% and a YIC of −8.3, where the first-order model was identified as the optimal model (Table 3). A simple first-order model was used to derive the TC to compare the dynamic relationship between rainfall and runoff response among the three catchments. Our interpretation of the CT TF model suggests a slower rainfall-runoff catchment response in the NF and TTP catchments in contrast to the fast flow response to rainfall in the SHA catchment. The TCs calculated ranged from 8.4 to 11.5 days in the NF catchment and from 9.6 to 13.0 days in the TTPs. The SHA had TCs between 6.2 to 8.6 days (Table 3).

Table 4 Summary of Cross-Correlation Functions (CCFs) Between Rainfall/Positive Derivative of Discharge to Suspended Sediment With Time Lag (in Hours) and Peak Cross-Correlation Coefficients Reported for the Natural Forest, Tea-Tree Plantation, and Smallholder Agriculture Catchment in the South-West Mau, Kenya, Over Four Years (Study Period 2015-2018)
Site Year

Time Lags Between Rainfall and Discharge to Sediment Concentrations
The analysis using CCFs (Equation 7) showed statistically significant correlations between rainfall and discharge to suspended sediment exceeding the 95% confidence interval of 0.001 for a sample size of 61,201 in all three catchments for the four years. The CCF indicates the impulse response time between the peak of rainfall and discharge to the suspended sediment peak (see Figure S2). The NF had almost instantaneous (<10 min) to rapid responses (1.5 hr) in suspended sediment to both rainfall and discharge in all four years. A period shorter than a year was used for the NF in 2015 and 2016 because of missing sediment data ( Table 4). The rainfall to sediment cross correlogram in the TTPs differed from the other two catchments, with a first fast peak within 1 hr followed by a delayed second peak after 5 to 8.5 hr. The discharge to sediment response in the TTPs was similar to that of the SHA catchment with a time lag of around 2 to 3 hr. The SHA had variable time lags between either rainfall or discharge to sediment ranging between 1.5 and 3.8 hr. The impulse response time between the peaks of discharge and sediment concentration was in general longer compared to the rainfall peak (Table 4).

Suspended Sediment Dynamics
This study shows that the annual suspended sediment yield is around 6 times greater for the drier SHA catchment and twice greater in the TTP catchment compared to the wetter NF catchment. The sediment yield difference is likely the result of more vegetation cover and low surface hydrological connectivity in the NF, where consequently erosion processes (including mass wasting) are not as active. Similar findings were reported for the neighboring Mara River Basin, where a semiarid catchment had higher suspended sediment yields (44 t km −2 yr −1 ) with half of the annual rainfall in contrast to a wetter but less populated and less disturbed catchment (33 t km −2 yr −1 ) (Dutton et al., 2018) (Table 5).
Missing sediment data and linear interpolation to fill these gaps could have increased the uncertainty in the sediment yields calculated. However, data gaps during dry periods, such as those in the sediment data for the NF during 2015 and 2016, would not have a large influence on yield calculations because of the small amount of material transported. Gaps during periods of heavy rainfall, which occasionally occurred in the SHA catchment due to siltation, could have contributed to underestimation of the sediment yield, as peaks in the sediment concentration could be missing.
For Africa, suspended sediment yield was estimated to be 634 t km −2 yr −1 at continental scale based on 682 catchments and rivers with larger drainage areas (mean >1,000 km 2 ) (Vanmaercke et al., 2014). Compared to this, and other sediment studies in Kenya with varying catchment sizes of 24-42,000 km 2 and 8.2 to 6,330 t km −2 yr −1 (Dunne, 1979;Vanmaercke et al., 2014), our annual suspended sediment yields (21-131 t km −2 yr −1 ) are within the lower reported ranges. Suspended sediment yields from the TTPs and NF catchment of our study are comparable to those observed in the neighboring upper Mara river basin (33 t km −2 yr −1 ; Dutton et al., 2018), which is dominated by small-scale farming and urban development.
The SHA catchment in our study had slightly higher suspended sediment yields than the Athi catchment in Kenya (109 t km −2 yr −1 ; Kithiia, 1997) and had lower suspended sediment yields than disturbed agricultural catchments in montane headwaters such as the May Zegzeg catchment in Ethiopia (850 t km −2 yr −1 ; Nyssen et al., 2009) or the Arvorezinha, Conceição, and Guapore catchments in South Brazil (140-298 t km −2 yr −1 ; Didoné et al., 2014;Minella et al., 2018). Guzman et al. (2013) found that in Ethiopia the highest suspended sediment yields (2,470 t km −2 yr −1 ) in small catchments were in those with the largest proportion of agricultural land. Their reported annual yields were significantly higher than the suspended sediment yields observed in this study with similar annual rainfall. Annual suspended sediment yields of 117 up to 2,740 t km −2 yr −1 from undisturbed to highly disturbed forest or upland grassland catchments were measured in Southeast Asia (Borneo, Thailand, and the Philippines) subjected to mass wasting during typhoon or posttyphoon events (Douglas et al., 2014;Ziegler et al., 2014).

Factors Controlling Sediment Yield 4.2.1. Vegetation Cover
The low annual suspended sediment yield measured in the Mau forest shows that forest vegetation is the most effective surface cover to limit soil erosion despite the steepest slopes of the forested 10.1029/2020WR027495 Water Resources Research catchment (Table 1). The dense vegetation, diverse strata, and complex rooting systems prevent soil detachment and trap potentially erodible material. Similarly, a dense perennial tea vegetation covers the soil surface in the TTPs, which can buffer erosive rainfall (Edwards & Blackie, 1979). Nevertheless, the annual suspended sediment yield for the TTP catchment was twice that of the NF despite the soil conservation practices applied by tea companies such as mulching and planting of buffer strips (oat grass) between rows of young tea bushes or cover trees on newly planted tea plots. This indicates that high sediment loads originate from unprotected bare surfaces during renovation of tea plantations or logging activities of woodlots. Logging activities trigger overland flow and erosion processes and lead sediments to the streams, when there are no buffer strips (Chappell et al., 2004;Douglas et al., 1993). The dense vegetation contrasts with the land management in the SHA catchment, where steep slopes tend to be bare between crop harvest and the start of the next cropping season. During that period, bare surfaces are prone to soil erosion, although cropland surface erosion was not observed, which may be explained by the high infiltration rates previously measured on these croplands (401 ± 211 mm hr −1 ) (Owuor et al., 2018). The routing of main flow paths was observed on compacted gullied tracks, which act as ephemeral channels during a storm event. Based on these observations, we hypothesize that rural unpaved tracks in the SHA generate a larger contribution to the total sediment load than agricultural land, but further work is required to confirm this. A potential reason for the annual increase in suspended sediment yield in the SHA and the NF catchments during the study period could be the reduced tree cover and increasing areas under annual crops and forest disturbance indicated by the study of Brandt et al. (2018).

Connectivity Between Sediment Sources and the Streams
The TTPs and the SHA have higher catchment surface connectivity than the NF, which may be causing higher sediment transfer by connecting multiple source areas with the streams. Lateral linkages (tracks, gullies, or drains) connect sediment source areas at catchment scale with the stream network (Lane & Richards, 1997), which can be an important driving force for the total sediment load into the rivers (Sidle & Ziegler, 2010).
In the SHA catchment, unpaved tracks are the main pathways for people and livestock to access streams, thus being frequently used and heavily trafficked also by motorbikes ( Figure 2). This activity generates highly compacted surfaces, where soil infiltration is impeded. Ziegler et al. (2001) observed that unpaved rural roads, similar in appearance to the unpaved tracks of our study, generate significantly more overland flow compared to adjacent hillslopes. As a consequence of low infiltration rates and downslope-orientated tracks, surface runoff energy increases generating more volume and velocity of flow that can transport large quantities of soil, eventually eroding tracks into gullies (Sidle & Ziegler, 2010;Svoray & Markovitch, 2009). Other researchers found a strong influence of subsurface water tables in valleys on gully formation and the development of large-scale sediment mobilization (Tebebu et al., 2010;Zegeye et al., 2018). High sediment loads at the outlet of the SHA catchment are thought to originate from the eroded unpaved tracks and its connecting adjacent source areas. Catchment drainage density is higher in the SHA (0.64 km km −2 ) compared to the NF and TTPs (0.48 and 0.42 km km −2 , respectively) suggesting a link to increased erosion rates. The TTP catchment is hydrologically connected through a network of tracks and well-engineered paved and unpaved drains in between the tea fields. The design of the well-engineered drains took into account the appropriate routing of surface runoff to the riparian zones before entering the streams, suggesting a higher hydrological connectivity than in the SHA catchment. However, the drains are well maintained with a densely forested riparian zone, which reduces sediment export, thereby reducing sediment transport connectivity. The strong hydrological connectivity of the TTP catchment could lead to high sediment transport and loads with poor maintenance of the drainage network.

Riparian Zones
Dense riparian vegetation can trap sediments before they reach the stream (Pavanelli & Cavazza, 2010). An intact forested riparian zone in the NF and a riparian buffer of 30 m, as predescribed by the Kenya's Water Act (Kenya, 2012), of mixed indigenous vegetation in the TTPs seem to be reducing sediment delivery to the streams by trapping eroded soil. In contrast, high sediment loads are expected in the SHA catchment, where the riparian vegetation is highly degraded or replaced by crops or woodlots planted on the river banks. Small floodplains in a steep, narrow valley floor provide limited space for sediment storage. In the same river basin, other studies reported that highly degraded riparian zones adjacent to areas cultivated by SHA lead to 10.1029/2020WR027495 increased suspended sediment concentrations (Masese et al., 2012;Njue et al., 2016). In the SHA catchment, livestock access the streams through the riparian area for watering (Figures 2d-2f), which damages the riverbank and the riparian vegetation and further increases sediment supply.

Water Pathways Are Key for Sediment Production
Hydrological pathways such as surface runoff or subsurface flow are key in determining sediment response in catchments. Our analysis showed that the NF and TTP catchments, with lower suspended sediment yields, had the longest streamflow response time to rainfall using the CT TF model (Equation 6). This supports our hypothesis that shorter pathways indicate that surface runoff mobilizes soil particles causing six times more suspended sediment yields.
The number of pathways and their response time depends on catchment characteristics (Chappell et al., 2006;Ockenden & Chappell, 2011). Forest ecosystems are generally characterized by complex catchment behavior (Chappell et al., 1999), where soils with high infiltration rates promote infiltration to deeper subsurfaces. These pathways can be divided into shallow water pathway and deep groundwater pathway (Chappell & Franks, 1996). The high infiltration rates (760 ± 500 mm hr −1 , Owuor et al., 2018) and the long TCs derived through modeling for the NF catchment (Table 3) point to subsurface flow pathways. Jacobs, Timbe, et al. (2018) reported the occurrence of shallow to deeper subsurface flow by using an endmember mixing analysis in the same NF catchment. Groundwater seemed to be an important stream water source (Jacobs, Timbe, et al., 2018), which agrees with the long response time we calculated (Table 3). Our findings corroborate those of other studies in tropical forest catchments, which demonstrated that subsurface flow is the main water pathway of forest ecosystems (Boy et al., 2008;Muñoz-Villers & McDonnell, 2013;Noguchi et al., 1997). Consequently, low suspended sediment yields are associated with limited surface erosion and sediment delivery to the streams in the NF catchment.
The main pathways in the TTPs with slightly shorter TCs (Table 3) and almost half the infiltration rate (430 ± 290 mm hr −1 , Owuor et al., 2018) compared to the NF suggest that shallow subsurface flow and surface runoff may dominate. Overland flow was observed to be routed through the well-engineered drainage network along roads and surface water drains between tea plantations to the well-buffered fluvial network. Overland flow was also thought to be significant by , where nitrate concentrations in stream water seemed to be diluted by surface runoff. The well-maintained drains explain the lower suspended sediment yields, despite the prevalence of surface runoff observed in the TTPs.
The analysis of the SHA catchment showed a relatively fast pathway. However, Owuor et al. (2018) measured infiltration rates of 401 ± 211 mm hr −1 on croplands in the catchment, suggesting that subsurface flow is the most likely pathway. Nevertheless, field observations of runoff along with poorly maintained highly compacted tracks, and the shape of a classification of hysteresis loops by  provides a contrasting insight suggesting that surface runoff in the SHA is an important vector for sediment. Our hypothesis is that the tracks act as ephemeral streams and receive water from the surrounding areas as shallow lateral flow, thus explaining the shorter water response times. The NF and TTP catchment with high tree cover showed similar TCs, whereas the much lower tree cover in the SHA catchment lead to faster pathways. The generally slow pathway component in each model can be explained by the presence of deep and well-drained soils in all catchments (Sombroek et al., 1982).

Sediment Response Times and Event Duration
The shorter time lag between the peaks of rainfall to sediment than to discharge can be attributed to exposed and easy erodible material adjacent to the outlet. The almost instantaneous sediment response to rainfall in the NF and TTPs suggests sediment supply from readily available, nearby sediment sources

10.1029/2020WR027495
Water Resources Research (De Girolamo et al., 2015;Francke et al., 2014;Tena et al., 2014) (Figure 7). Near-channel or in-channel sediment sources can originate from the stream bank or the stream bed (Chappell et al., 1999;Kronvang et al., 1997;Lenzi & Marchi, 2000). Temporarily stored sediment is thought to be mobilized very quickly during the first stages of a storm event (Eder et al., 2010). Fast response systems of short duration and low mass magnitude, as those observed in the NF catchment, are characterized by rapid sediment flushing and fast depletion of sediment supply (Chappell et al., 1999), due to limited availability of eroded material from the protected surface (De Girolamo et al., 2015;Fan et al., 2012;Fang et al., 2008). The paved drain network in the TTP catchment can act as conduits transporting sediment instantaneously from nearby logged plantations or tracks to the stream. The delayed sediment response after a rainfall event in the SHA and the second sediment response in the TTP catchment suggest a long travel distance between the sediment source and the catchment outlet. The delayed response can be related to the magnitude of mass, where more distant sediment source areas from the wider catchment accumulate more mass over a longer period. These responses may also indicate the breaching of barriers such as hedges, fences, or grazing land, especially in the SHA catchment. The long recession limb in the SHA and the TTP catchments is typically explained as a slow depletion of sediment supply (De Girolamo et al., 2015;Francke et al., 2014;Tena et al., 2014). The more pronounced sedigraph in the SHA catchment can be associated with the wide range of accumulated sediment source areas (Figure 7).

Seasonal Variability of Suspended Sediment
Throughout the study period, we observed distinct seasonal variability in suspended sediment yield for the NF, TTPs, and SHA with higher yields in periods of high discharge and rainfall. In climates with strong seasonality of rainfall, seasons can explain sediment dynamics (De Girolamo et al., 2015;Horowitz, 2008), as we observed in this study. Wet seasons or high-flow events generate the largest proportion (80-95%) of the annual sediment load, as observed by De Girolamo et al. (2015), Sun et al. (2016) and Vercruysse et al. (2017) in other areas, while the sediment load is the smallest in the dry season with less than 5% in our study. The seasonal differences were more pronounced in the SHA catchment compared to the other two catchments explained by the high catchment surface connectivity and low vegetation cover.

Conclusions
This study presents the first long-term high-resolution sediment data set in Kenya. The four years of continuous data for the NF, TTP, and SHA catchments provide critical insights in contrasting sediment dynamics of the tropical montane Mau Forest Complex. The analysis revealed that land use is a critically important driver for sediment supply, where SHA generates 6 times more annual suspended sediment yield than a catchment dominated by NF. Besides vegetation cover, a strong catchment surface connectivity through unpaved tracks and gullies from hillslopes to the fluvial network is thought to be the main reason for the differences in sediment yields. However, further work is required to test this hypothesis. Catchments with a high tree cover, such as the NF and the TTPs, seem to have similar water pathways with a dominance of subsurface flow. In contrast, in the highly disturbed landscape such as that of the SHA catchment, surface runoff dominates and soil erosion increases suspended sediment yield. This superficial water pathway results in the more pronounced seasonal impact of rainfall in the SHA compared to the other two catchments, also due to varying vegetation cover. Delayed sediment response to rainfall and a slow depletion in sediment supply in the SHA and TTPs suggests that the wider catchment area is supplying sediment from a range of sediment sources, especially in the catchment dominated by smallholder farming. In contrast, the fast depletion in sediment supply in the NF suggests the importance of nearby sediment sources and temporarily stored sediment.
Land scarcity and population growth bring enormous pressure on NF ecosystems. Forest conversion will increase sediment production, which will affect not only many people in the Sondu River basin who rely on the rivers for drinking water but also Lake Victoria which is already affected by increased sediment supply. The implementation of catchment management, such as soil conservation measures and better engineering of rural trackways, is essential to reduce sediment supply to water bodies. However, a detailed sediment source fingerprinting analysis is necessary to identify the main contributing sediment sources. This will support the application of better management strategies at the source to prevent sediment entering the stream network. Sediment yields reported in other sediment studies in montane SHA catchments were 10.1029/2020WR027495

Water Resources Research
higher than in our study, which provide a warning of potentially higher sediment loss in the future unless mitigation strategies are implemented.