1 Introduction

The degradation of agricultural land as a result of soil erosion is a worldwide phenomenon leading to the loss of nutrient-rich surface soil and increased runoff from the more impermeable sub-soil that leads to the lowering of agricultural productivity (Erkossa et al. 2015; Taguas et al. 2015; Ganasri and Ramesh 2015; Keesstra et al. 2016; Nigussie et al. 2017). Soil erosion is more severe in the sub-Saharan African countries where the population livelihood is dependent on the soil (Sunday et al. 2012; Erkossa et al. 2015). In the Ethiopian highlands, deforestation for crop production, cultivation of marginal lands, and overgrazing are the major factors that dramatically increased the vulnerability of agricultural lands to rainfall-driven soil erosion (Nyssen et al. 2000; Vancampenhout et al. 2006; Belay et al. 2014; Adimassu et al. 2014; Erkossa et al. 2015; Addis et al. 2016). Intensive rainfall during rainy seasons (June to September) contributes to severe land degradation in mountainous regions, especially on steep sloping and unprotected areas (Addis et al. 2016).To tackle the soil erosion problem in the Ethiopian highlands, constructing soil and water conservation structures is considered to be a top priority in halting land degradation and thus to improve agricultural productivity.

Since 2010, a massive effort has been undertaken by the government of Ethiopia in constructing soil and water conservation structures on privately owned and community lands through community mobilization (Kebede 2014; Dagnew et al. 2015; Teshome et al. 2016; Dagnew et al. 2017; Grum et al. 2017; Guzman et al. 2017). Examples of soil and water conservation practices include stone bunds, soil bunds, and percolation ditches (Teshome et al. 2016). However, the effectiveness of these soil and water conservation measures on the dynamics of watershed-scale runoff and sediment loading has not been sufficiently studied and identified clearly for long- and short-term effects in the Ethiopian highlands.

In the northern highlands of Ethiopia, different studies have been carried out on the impacts of soil and water conservation structures on erosion process at the field scale (Kaltenleithneret al. 2014; Rieder et al. 2014; Strohmeier et al. 2015; Klik et al. 2016; Obereder et al. 2016). These studies reported that the soil and water conservation (SWC) structures are effective at the plot scale in the Gumara-Maksegnit watersheds. However, studies on the impacts of soil and water structures on erosion process at the watershed scale are limited. As data from field experiments cannot be extrapolated to a watershed scale (Verstraeten et al. 2006), the use of mathematical models for evaluating soil and water conservation measures is quite common.

Insufficient information on soil erosion and streamflow could lead to inefficient planning and inadequate design and operation of soil and water resource management projects (Poitras et al. 2011). Changes in the extent of seasonal precipitation, frequency, and intensity of extreme precipitation events directly affect the amount of seasonal streamflow (Poitras et al. 2011). The prediction and assessment of streamflow and sediment yield using a watershed model are important for agricultural watershed management in the Ethiopian highlands as watershed models are crucial tools to illustrate hydrological processes and to scale up the model results.

The soil and water assessment tool (SWAT; Arnold et al. 1998) is a continuous-time, semi-distributed, process-based river basin- or watershed-scale model. The model is one of the most comprehensive models able to evaluate hydrologic processes (Gassman et al. 2007). SWAT has been employed to simulate the discharge in the Ethiopian highlands (Setegn et al. 2008; Setegn et al. 2009; Easton et al. 2010; Setegn et al. 2010; Betrie et al. 2011; Setegn et al. 2011). Hence, the objectives of this study were (1) to calibrate and validate the SWAT model for two watersheds with and without SWC structures, (2) to study the impact of these structures on runoff and erosion processes, and (3) to provide feedback on the efficiency of the structures in reducing soil erosion in the watersheds and to advise future upscaling.

2 Materials and methods

2.1 Description of the study area

The two study watersheds, treated watershed (TW) and untreated watershed (UW), are located in the Gumara-Maksegnit watershed in northwest Ethiopia (Fig. 1). The watersheds drain into the Gumara River which finally drains into Lake Tana. They are located at 12° 25′ 24″ and 12° 25′ 54″ latitude and at 37° 34′ 56″ and 37° 35′ 38″ longitude and at an altitude ranging from 1998 to 2150 m above sea level (Fig. 1). The two study watersheds are neighboring each other at a distance of 1 km between the outlets. The TW has an area of 27.1 ha, and the UW has an area of 31.7 ha. The slope of the watershed ranges from 2 to 69% with an average slope of 14.8%. The soil types in the watersheds are Cambisol and Leptosol which are found in the upper and central parts of the watershed, while Vertisols are found in the lower catchment. The watershed is characterized as sub-humid. It has a long term (1997–2015) mean annual rainfall of 1157 mm with 80% raining from June to September and mean minimum and maximum temperatures of 13.3 and 28.5 °C, respectively (Addis et al. 2016).

Fig. 1
figure 1

Maps of Ethiopia (left top), the larger Gumara-Maksegnit watershed (middle), and the two paired watersheds (right)

In 2011, stone and soil bunds were constructed as SWC structures in TW. Stone and soil bunds are SWC structures made of stones or soil constructed along a contour in order to best slow the water flowing down the slope, which prevents soil erosion (Fig. 2). On farmlands, 40-cm-high stone and soil bunds at distances ranging between 15 and 25 m depending on the steepness of the land were constructed along the contour. Gullies were treated with check dams. The UW was used as a reference without any SWC structures. Similar dominance of land use, soil texture/type, and topography of these two watersheds were used as basis for the comparison between TW and UW (Fig. 3). The main staple food crops in the watershed are sorghum, teff, barley, beans, wheat, and corn. The main crop rotations are chickpea followed by teff followed by sorghum (chickpea-teff-sorghum), chickpea followed by wheat followed by sorghum (chickpea-wheat-sorghum), and beans followed by barley followed by corn (beans-barley-corn).

Fig. 2
figure 2

Erosion plot experiments (left) and SWC structures (right) at the treated watershed

Fig. 3
figure 3

Slope classes, soil map, land use classes, and sub-basins of both watersheds

2.2 Runoff discharge and sediment yield

Runoff and sediment yield were determined at the outlet of both watersheds where rectangular v-notch weirs with flow sensors and automatic cameras were installed. Gauges were marked on the walls of the weirs, and the cameras took pictures every 2 min (Fig. 4). Discharges were then calculated using a rating curve developed on measurements by Zehetbauer et al. (2013). During each rainfall event, runoff samples were collected at three times during the runoff event. Samplings were made approximately at the start of runoff, the middle of the runoff event, and towards the end of the event. Samples were taken in triplicate with a volume of approximately 3 L each. For each event, an average sediment concentration in each sample was determined in the soil laboratory. Sediment yield was then obtained by multiplying runoff discharge by the sediment concentration. The data from 2011 to 2013 were used to calibrate and those from 2014 to 2015 were used to verify the distributed SWAT model.

Fig. 4
figure 4

Pictures taken from the automatic camera at day (left) and nighttime (right)

2.3 SWAT model and model input

The SWAT was used to estimate runoff and sediment yield in the TW and UW watersheds. Surface runoff was modified by the adjustment of the runoff ratio (curve number) (Bonta 1997a, b) while SWC structures’ impact on sediment yield was adjusted through the support practice factor (P-factor) and/or the slope length factor (LS).

Curve number values were modified by editing the management (.mgt) input table from the field experiment data results (Klik et al. 2016) while the average slope length (SLSSUBSN) value was modified by editing the hydrologic response unit (HRU) (.hru) input table. The watershed was divided into a number of sub-basins with different HRUs. The HRU represents the unique property of each parameter in the watershed. Parameters like surface runoff, sediment yield, soil moisture, nutrient dynamics, and crop growth were simulated for each HRU, aggregated, and processed to sub-basin level on a daily time step resolution. Input data are required for SWAT model that can be supplemented with GIS data and the model interface (Di Luzio et al. 2002).

For this study, SWAT offers finer spatial and temporal scales, which allow observation of output at a particular sub-basin on a daily basis. Digital elevation model (DEM), land use, and soil and climate data were used as inputs for the SWAT model.

2.4 Model input

A DEM of the watershed was developed with 5-m resolution at 2556 coordinate points using a total station to derive topographical parameters and automatically delineate watershed boundaries and channel networks. The study watersheds were divided into the following five slope steepness classes: 0–10, 10–20, 20–30, 30–40, and greater than 40% (Fig. 3). Nine land use classes were developed for the study watersheds (Fig. 3) using Google Earth imagery taken on 14 October 2011 which were cross-checked using ground truth data. The land use percentages of each watershed are summarized in Table 1.

Table 1 Land use and land cover in the untreated and treated watersheds

Soil physical and chemical properties of both watersheds were determined from soil samples collected from a 100 m by 100 m grids at 0–25-, 25–60-, and 60–100-cm soil depths. The UW has soil textural classes comprised of 3.7% clay, 52.9% clay loam, 36.3% loam, 4.7% silty clay loam, and 2.9% silty loam. Similarly, the TW has textural classes of 12.4% clay, 50.7% clay loam, 23.4% loam, 0.16% silty clay loam, and 12.9% silty loam (Fig. 3).

Daily precipitation and temperature data were collected from a weather station located at the outlets of UW. Daily solar radiation, relative humidity, and wind speed data were recorded from an automatic weather station installed approximately at a distance of 5 km from the watersheds. Missing daily weather data were simulated using SWAT weather generator (Schuol and Abbaspour 2007). Daily weather data (January 1, 1997 to December 31, 2015) recorded at both weather stations were used to create the weather statistics using the weather generator.

2.5 Project setup

For each watershed, a separate SWAT project was set up. The modeled period was from 2004 to 2015. Runoff and sediment yield data collected from the watersheds during 2011 to 2013 were used for model calibration while data from 2014 to 2015 were used for validation. Mean daily runoff and sediment data from both watersheds were used to calibrate the SWAT model. Some of the appropriate parameters were adjusted until the predicted daily runoff (Table 2) and sediment yield (Table 3) approximately matched the measured ones at the outlets of the watershed. Based on the given threshold areas and manual input data, automatic sub-basin delineation was done for the UW and TW (Fig. 3d). The SWAT model divided the sub-basin into detailed HRUs. The model delineates each HRUs with a user-defined threshold based on the percentage of the slope classes, soil type, and land use (Arnold et al. 2011).

Table 2 List of parameters adjusted for runoff during the calibration process
Table 3 List of parameters adjusted for sediment during the calibration process

The HRUs for this study were delineated using the soil type and the land use thresholds set at 5% area coverage. Any soil type and land use type each covering more than 5% of the sub-basin area were considered as an HRU. Based on the thresholds selected, there were a total of 760 HRUs in the UW and 658 HRUs in the TW. These HRUs were used for analyses on a particular land use, soil type, and slope class.

2.6 Model performance evaluation

Graphical and statistical model evaluation techniques were used to see how well the simulated results match the observed data. SWAT and SWAT-CUP calibration tools provide multiple model evaluation statistical criteria to be selected as an objective function for model calibration and validation based on the recommendations suggested by Santhi et al. (2001) and Moriasi et al. (2007). The SUFI-2 (Sequential Uncertainty Fitting 2) program which is linked to SWAT-CUP2012 version was used for a combined model sensitivity analysis, calibration, and validation procedures (Abbaspour et al. 2004, 2007). The SUFI-2 algorithm accounts for different sources of parameter uncertainty, conceptual model uncertainty, and input data uncertainty (Gupta et al. 2006). In this study, coefficient of determination (R 2) (Krause et al. 2005), Nash-Sutcliffe efficiency (NSE) (Nash and Sutcliffe 1970), and percent bias (PBIAS) (Gupta et al. 1999) evaluation statistics were used to see the goodness of fit of the model (Santhi et al. 2001; Moriasi et al. 2007). The equations used are

$$ {R}^2={\left[\frac{\sum_{i=1}^n\left({O}_i-\overset{\_}{O}\right)\left({E}_i-\overset{\_}{E}\right)}{\sqrt{\sum_{i=1}^n{\left({O}_i-\overset{\_}{O}\right)}^2}\sqrt{\sum_{i=1}^n{\left({E}_i-\overset{\_}{E}\right)}^2}}\right]}^2 $$

where n is the number of observations or samples, O i is the observed values, E i is the estimated values, Ō is the mean of observed values, Ē is the mean of estimated values, and I is the counter for individual observed and predicted values. The R 2 ranges between 0 and 1, where 1 means that the predicted value is equal to the observed value and zero means that there is no correlation between the predicted and observed values.

$$ \mathrm{NSE}=1-\frac{\sum_{i=1}^n{\left({E}_i-{O}_i\right)}^2}{\sum_{i=1}^n{\left({O}_i-\overset{\_}{O}\right)}^2} $$

The range of E lies between − ∞ and 1.0 with E = 1 describing a perfect fit. Values between 0 and 1.0 are generally viewed as acceptable levels of performance, whereas values < 0 indicate that the mean observed value is a better predictor than the model.

$$ \mathrm{PBIAS}=\left[\frac{\sum_{i=1}^n\left({O}_i-{E}_i\right)\times 100}{\sum_{i=1}^n\left({O}_i\right)}\right] $$

The optimal value of PBIAS is 0, with low-magnitude values indicating accurate model simulation (Moriasi et al. 2007).

3 Results

3.1 Runoff calibration and validation

Results showed that the observed mean daily discharge was 0.03 m3 s−1 for the calibration period and 0.02 m3 s−1 for the validation period whereas the estimated mean daily discharge was 0.03 m3 s−1 for the calibration period and validation period in the UW (Table 4). The simulation results showed that the R 2 and NSE values for the daily runoff in the UW were 0.77 and 0.75 for the calibration period and 0.72 and 0.56 for the validation period, respectively (Table 4). PBIAS was − 8.9 for the calibration and 14.8 for validation for the UW.

Table 4 Mean daily discharge, sediment yield and summary statistics of treated and untreated watersheds

Similarly, the estimated and the observed daily discharge for the TW was 0.02 m3 s−1 for the calibration and validation periods (Table 4). The daily runoff simulation results showed better model efficiency with a R 2 value for the daily runoff of 0.78 for the calibration period and 0.70 for the validation period (Table 4). The NSE values were 0.63 for calibration and 0.58 for validation periods (Table 4). The mean daily results give PBIAS of 29.2 for calibration and 24.3 for the validation periods (Table 4) indicating that the model performed well according to Moriasi et al. (2007). Results showed that there is good agreement between the observed and predicted daily runoff for both TW and UW during calibration and validation periods (Figs. 5 and 6) indicating that SWAT performs well. This indicates that the model predicts the daily discharge very well (Figs. 5 and 6).

Fig. 5
figure 5

Observed and simulated daily runoff for calibration (a) and validation (b) periods at the outlet of the untreated watershed (UW)

Fig. 6
figure 6

Observed and simulated daily runoff for calibration (a) and validation (b) periods and at the outlet of the treated watershed (TW)

The observed runoff on the same day was often under predicted for the calibration period and for the validation period. Based on the model results, the mean daily runoff from both watersheds shows better agreement with the measured runoff calibration and validation periods (Fig. 7a, b). The evaluation coefficients of the simulated daily runoff of different objective functions for both the TW and UW indicated satisfactory model fit according to the assessment criteria (Moriasi et al. 2007). Khelifa et al. (2016) reported daily runoff with NSE values of 0.64 for calibration and 0.68 for validation. Similar studies are in better agreement with these results (Addis et al. 2016; Zimale et al. 2016). For a study in the Gumara watershed by Zimale et al. (2016), the NSE values for daily flows obtained were 0.70 for calibration and 0.77 for validation period.

Fig. 7
figure 7

Observed and simulated daily discharge (Q) for calibration and validation periods at UW (a) and TW (b)

3.2 Sediment calibration and validation

Daily sediment yield was calibrated and validated using the measured data from the two watersheds. Sediment yield prediction results gave a R 2 of 0.69 for calibration and 0.65 for validation period in the UW (Fig. 8a) and a R 2 of 0.65 for calibration and 0.55 for validation period for the TW (Fig. 8b).

Fig. 8
figure 8

Observed and simulated daily sediment for calibration and validation periods at UW (a) and TW (b)

Daily sediment yield calibration and validation results showed NSE of 0.47 and 0.31, respectively, for the TW and 0.54 and 0.33 for calibration and validation, respectively, for the UW (Table 4).

Results showed that SWAT model underestimated the generated sediment yield. The model predicted sediment yields of about 33.5 and 44.8 t ha−1 year−1 for TW and UW, respectively. The observed sediment yields were 39.9 and 64.6 t ha−1 year−1 in TW and UW, respectively. The model underpredicted the annual sediment yield of the TW and the UW. This indicates that there is a potential impact of the SWC for sediment yield reduction in TW.

4 Discussion

The results show that the soil and water conservation structures constructed by the farmers reduce the surface runoff and soil losses in the highlands of Ethiopia. The results show that the untreated watershed had higher sediment and runoff losses than the treated watershed, given similar climatic and land use patterns. The intervention of SWC measures by the mobilization of the community has a significant soil loss reduction to protect their land from the rainfall-driven soil erosion. The effectiveness of SWC on runoff and sediment yield reduction has been reported in other studies in northern Ethiopia (Gebremichael et al. 2005; Haregewyn et al. 2005; Descheemaeker et al. 2006; Mitiku et al. 2006; Nyssen et al. 2007, 2009; Dagnew et al. 2015, 2017).

In this study, SWAT was used to assess the impacts of SWC on runoff and erosion processes, and the model has been found to be a useful tool for understanding the hydrologic processes and the sediment dynamic in the study area in both watersheds. The evaluation coefficients of the simulated daily runoff of the different objective functions for both the TW and UW indicated satisfactory model fit according to the assessment criteria (Moriasi et al. 2007). The NSE values found in the UW and the TW agree with Khelifa et al. (2016) who studied the impact of SWC on runoff and sediment yield in Tunisia. These authors report NSE values for runoff simulations using SWAT of 0.64 and 0.68 for calibration and validation, respectively. In another study in the Gumara watershed by Zimale et al. (2016), the NSE values for daily flows obtained were 0.70 for calibration and 0.77 for validation period, which is comparable with the UW and TW NSE values in the Gumara-Maksegnit watersheds.

However, the model tends to underestimate sediment yield during calibration and validation periods for both watersheds. The NSE for sediment yield in both watersheds showed lower values in the calibration and validation periods. The underestimation of the sediment yield by the model is because there are parts of the watershed severely eroded which created gully erosion in both watersheds that led to higher soil losses beyond the estimated sediment load. This is substantiated by the photo taken in Fig. 9 which shows the development of deep gully in the upper parts of the watershed that contributes higher soil erosion losses that generate higher sediment load in the outlets.

Fig. 9
figure 9

Gully development in the upper part of the watershed (left) runoff with high sediment concentration at the outlets (right)

The model result indicated that SWC structures may considerably reduce soil loss by 24.8–38.2% in the Gumara-Maksegnit watershed. A plot-level experiment conducted on the effects of stone bunds showed that stone bunds can reduce soil erosion by 33–41% (Rieder et al. 2014; Klik et al. 2016) in the TW which is close to the current finding of the soil loss reduction level due to SWC. Similarly, Strohmeier et al. (2015) reported that at plot scale, stone bunds reduced soil loss by 40% in the Gumara-Maksegnit watershed. In another study conducted in northern Tunisia on the effects of soil and water conservation structures on sediment load, Khelifa et al. (2016) reported 22% reduction in sediment yield at the watershed scale. Similar studies by Abouabdillah et al. (2014), Yesuf et al. (2015), Addis et al. (2016), and Licciardello et al. (2017) are in agreement with our findings. Betrie et al. (2011) also reported 41% reduction in sediment yield in the Blue Nile basin due to stone bunds. The soil loss reduction (25–38%) in this study due to SWC structures at the watershed scale agreed with the findings of Abouabdillah et al. (2014) who estimated an overall soil loss reduction by 25%.

The sediment yield estimated by SWAT model for the UW (44.8 t ha−1 year−1) and TW (33.5 t ha−1 year−1) was in agreement with other studies. Setegn et al. (2010) reported that sediment loads of 30–60 t ha−1 year−1 were exported from the Lake Tana watersheds while Easton et al. (2010) predicted a maximum soil loss of 84 t ha−1 year−1 in the Gumara watershed. Similarly, Zimale et al. (2016) reported an average sediment yield of 49 t ha−1 y−1 from the Gumara watershed. There are also a number of simulation studies on sediment load prediction at the gauging stations near Lake Tana (Easton et al. 2010; Setegn et al. 2010; Kaba et al. 2014; Zimale et al. 2016) which confirmed the results of the current study conducted in the Gumara-Maksegnit watershed.

5 Conclusions

Soil erosion is a serious problem in the Ethiopian highlands arising from agriculture intensification, deforestation, and land degradation. To find erosion hot-spot areas and to develop a soil conservation strategy, assessment of soil erosion is a useful tool. Empirical soil erosion modeling can provide a quantitative and consistent approach to estimate runoff and soil erosion under a wide range of conditions.

In this study, SWAT has been used to predict the impacts of SWC interventions on runoff and soil loss for two adjacent watersheds in the highlands of Ethiopia. SWAT-CUP has been used to perform calibration and validation of the observed and simulated runoff and soil losses. For UW and TW, separate SWAT and SWAT-CUP analyses were undertaken for estimating daily runoff and sediment yield.

The results of the simulation study showed good model performance for daily runoff prediction at each watershed with acceptable R 2, NSE, and PBIAS values. However, the model performance was poor in terms of predicting sediment loss with lower NSE values. Overall, the watershed modeling results indicated that soil and water conservation structures can reduce runoff and soil loss in the Gumara-Maksegnit watershed. Based on the calibrated SWAT model, the soil loss from the TW is found to be lower than that from the UW. The SWC structures reduced soil losses by 25–38% in the TW as compared to the UW. However, the soil erosion is still severe and above the world tolerable ranges (2–11 t ha−1 year−1). Therefore, land management strategies and SWC structures should be improved to achieve more sustainable soil erosion protection for sustainable agriculture for food security in the area. Generally, this study supports the application of SWAT model to other watersheds in the Ethiopian highlands to predict the impact of SWC structures on runoff and soil erosion processes.