Evaluating Impact of Land-Use/Land-Cover Change on Surface Runoff using Arc SWAT Model in Sore and Geba Watershed, Ethiopia

The aim of this study is to assess the impact of LULC change on surface runoff in the watershed of Sore and Geba, the upper Baro-river basin covering an area of approximately 6551 km in Ethiopia. Landsat images were used to analyze the LULC change trends for the periods of three decades (1987-2017). LULC map of the area produced using the maximum likelihood algorithms based supervised image classification. Trends in LULC change showed that farm land has augmented by 16.55% within the periods of between 1987 and 2017with annual expansion by 36.15 km at the expense of other land use types such as open forest, dense forest and wood land. The Soil and Water Assessment Tool (SWAT) model has been used to evaluate the impacts of LULC change on surface runoff. Nine sensitive flow control parameters were identified and used for calibration of the model. In both calibration and validation periods, good performance was obtained. Results show between 1987 and 2017, a 16.5% cultivated land expansion was observed which may explain an increase of about 6.65 m/s (32 mm) in annual surface runoff. In general, during the study period significant influence of LULC change were reflected in changes to the hydrologic system of the region with an important management implication for this region as well as other similar regions in Ethiopia.

source-based measure for the problem and develop a new plan in the catchment, evaluating the effects of land use practice on hydrology is imperative step in making a better decision about the land use planning by the government and the decision makers.
The aim of this research is to assess land use change impacts on surface runoff using Soil and Water Assessment Tool (SWAT) model. Specifically,  understanding the effects of different land use practices on surface runoff response and identifying land cover change trends in the area  determining the components of water balance in the watershed, and  understanding implications and recommend a land use management best practice 2. METHODOLOGY 2.1. Description of study area Baro river is one of the major rivers in Baro-Akobo basin which lies in southwestern part of Ethiopia located between 33 0 23'39'' to 36 0 18'21''E and 9 0 25''2'' to 7 0 27'8''N, which defines part of Ethiopia border with South Sudan. From its source in the southwestern Ethiopian highlands it flows west for 306 kilometers to join the Pibor River that flows to White Nile after forming Sobat River system. Regionally it lies in the Ethiopian administrative regions of Gambela, Oromia and Southern Nation Peoples regional state. The drainage area of the basin including its tributaries are about 41,400 km 2 and is bordered by the Sudan in the Northwest, Abbay Basin in the east and Akobo basin in the Southwest. The elevation of the catchment ranges from 3244 m in the Southwestern highlands of Ethiopia and 390 m at the point where the border of South Sudan. For this specific study, Sore and Geba watershed about 6551.07 km 2 which is sub watershed of Baro river basin (one of the main tributaries) was selected ( Figure 1). The elevation of Sore and Geba watershed ranges from 937 to 3001meter above sea level.
Depending on the altitude and temperature of the area, the climate of Sore and Geba watershed categorized under the warm to cool, semi humid zones. The average annual rainfall of Sore and Geba watershed varies from 1533 mm up to 2046 mm. The rainfall distribution pattern of the study area is experiencing a unimodal rainfall pattern with continuous highest rain from March to October. The land use types of Sore and Geba watershed is predominantly broad leaf montane forest, wood land, cultivated land and water bodies (Wubie, 2015).

Data
For this study, different climatic and topographic data were acquired such as DEM, LULC and hydrological data. The land cover satellite image and DEM were obtained from United State Geological Survey (USGS) website. The climatological data can be gained from NMA of Ethiopia. Hydrological and Soil data was obtained from Ethiopian MoWIE and FAO. Satellite imageries of different bands for each year were used to identify the LULC change spreading in Sore and Geba watershed in the last 30 years. Landsat-5-TM, Landsat-7-ETM+ and Landsat Figure 2. Framework of the study Image processing The selection of the acquired data date was made as much as possible within the same annual season to avoid a seasonal variation in vegetation pattern and distribution throughout the year. The images were georeferenced to UTM projection UTM zone 37N. All satellite images are composite to the composition of RGB colors. There are several image classification techniques exist in literature for remote sensing image classification. Broadly categorized as supervised or unsupervised image classification techniques. The most widely used and found to be accurate method is supervised image classification techniques (Hasmadi, et. al. 2009). For this study, by selecting training sites and generating signature files, supervised image classification techniques was applied to classify the image using a maximum likelihood classification algorithm. Image classification was done by using ArcGIS image processing tools.

LULC mapping
The LULC classes of Sore and Geba watershed has been differentiated and identified depending on the available sources such as, the prior local knowledge of the study area, remote sensing, google satellite image and previous research output of the study area. The Landsat images can be classified into five different types of LULC classes. 1) Cultivated land: Areas used for annual crops by rainfed agriculture, Cereal Land Cover system which is moderately stocked and scattered rural settlements that are closely associated to cultivated fields. Scattered settlements were difficult to separate from remotely sensed images, due to the fact that it combined under the categories of cultivated land during classification Journal of Environment and Earth Science www.iiste.org ISSN 2224-3216 (Paper) ISSN 2225-0948 (Online) Vol.9, No.10, 2019 2) Dense forest: Forest montane, broadleaf, dense (50-80% crown cover) which includes evergreen forest land 3) Open forest: Forest montane, broadleaf, open (20-50% crown cover) which includes vegetation's with trees 4) Wood land: Areas with bushes, open (20-50% tree cover) mixed with some grasses, and 5) Water bodies: Marsh lands, Rivers and its main tributaries. Accuracy assessment Quality analysis was done by the confusion matrix. The various parameters describing the quality of image classifications are derived from the confusion matrix. The confusion matrix is a table with rows which represented the mapped (classified) categories derived from remote sensing data and the columns which representing the reference (observed) classes (Olofsson et al., 2014). The most frequently used index is overall accuracy, accuracy of the producer, and accuracy of the user. The resulting error matrix was used to calculate these.
Overall accuracy shows the percentage of pixels properly classified. This measure can be calculated by the total number of pixels in the confusion matrix as the ratio of the total number of accurately classified pixels (diagonals).
The accuracy of the producer is calculated by dividing the number of correct pixels by the total number of pixels as derived from the reference data into one class. It is an error of omission (exclusion). The accuracy of the user is derived from the ratio of the number of pixels in each category correctly classified to the total number of pixels in that category. It describes error of commission (inclusion).
The Kappa coefficient (K) is another accuracy assessment statistic used for this study. It reflects the difference that is expected by chance between the actual agreement and the agreement. It can be calculated by the following equation.
= * * * (2) The value of kappa ranges between 0 and 1, where 0 represents agreement due to chance only and a value of 1 represents complete agreement between the two data sets (classified map and reference data).

SWAT modeling
Data need for the modelling effort include DEM, LULC, soil properties, weather data and observed stream flow data for calibration and validation of the model. The calculation of flow direction, flow accumulation, stream networks, watershed delineation and calculation of sub-basin parameters using SWAT watershed delineator tools requires DEM data. Land use is one of SWAT model's main input data to define the study watershed's hydrological response units (HRU). Soil data, including physical and chemical properties, is an input data for the SWAT model. The weather data used for the study were represented from Five stations in and around the Sore and Geba watershed, such as Alge, Bedele, Gatira, Gore and Metu as shown in (Figure 3). Only Gore station are first class that has records on all climatic variables, whereas the other ones have only precipitation and temperature records. From 01 January 1996 to 31 December 2017, the climate data used for this study cover 21 years. While in some of the climatic variables in all stations, some missing values have been identified. These missing values have been selected and assigned with the SWAT no data value code (-99) that could be filled out from monthly weather generator parameter values embodied in the SWAT model. These monthly weather generator values were estimated by SWAT preprocessing software from Gore Meteorological Station. Data on stream flow is required for model calibration and validation. Because of the availability of daily stream flow data from continuous records from periods  was selected and by calculating the average monthly flows it was prepared according to the required SWAT-CUP format. Finally, as required by the model, all the input data were prepared.
and Geba watershed (Sore river gauge). The available stream flow data of sore river gauge covered 21years . The data were split in to two for calibration period (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) and validation period of (2010-2017). The expressive flow parameters identifiers were chosen from SWAT database, previous study output and its initial ranges of the parameters used in the calibration were assigned from absolute SWAT values that linked in SWAT-CUP programs. The sensitive parameters included in the calibration were selected based on an analysis of Global sensitivity that simultaneously varies all parameters. A t-test was then used to identify each parameter's relative significance. SUFI-2 used 2 iterations, 400 simulations for each iteration on the Sore river gage, and the best simulated parameter values were used for editing the SWAT model. The model was calibrated using the simulated flow obtained from the land use map of the year 1987. Then the best simulated parameter values were used for each land use map (1987,2001 and 2017) model run. The p-factor quantifies all the uncertainties, which is the percentage of measured data bracketed by the 95PPU (95 percent uncertainty prediction). The quality of the calibration was measured by r-factor, which indicates the thickness of the 95% prediction uncertainty (95PPU). The validation of the model for stream flow was carried out for the period of (2001-2017) by using the calibrated SWAT-CUP parameters ranges (without any further changes).

Model performance evaluation:
There are different model performance evaluation techniques such as, Nash-Sutcliffe efficiency (NSE), coefficient of determination(R 2 ), percent bias (PBIAS) and by the ratio of the root mean square error of the standard deviation of measured data (RSR) etc. To judge the simulation of streamflow as satisfactory NSE >0.50, RSR<0.70 and PBIAS+ 25% (Moriasi et al., 2007). For this study, the calibration and validation performance were carried out using the p-factor, r-factor, R 2 and NS model performance techniques.

Evaluation of LULC change on water balance components
To evaluate the changes of the simulated water balance components such as, surface runoff influence to river flow, ground water and lateral flow influence to stream flow, actual and potential evapotranspiration due to LULC change, the study was carried out for three different years (1987, 2001 and 2017). In order to evaluate the variability of the water balance components due to LULC changes from 1987LULC changes from to 2001LULC changes from , 2001LULC changes from to 2017LULC changes from and 1987 Figure 4 blow. In the last 30 years , there has been an upsurge in cultivated land and a decrease in dense forest, open forest, woodland and water bodies in the watershed of Sore and Geba. The total area coverage of the cultivated land in 1987 was around 33% and in 2001 about 34% of the total area of the watershed, but in 2017 it increased rapidly to 50%. This is because of the gain of land from the shrinkage of other types of land use due to population growth and deforestation. For example, the total area coverage of dense forest, open forest, wood land, and water bodies in 1987 LULC maps was about 3%, 49%, 4% and 11% respectively. However, in 2017 land use and land cover map it decreased to around 0%, 44%, 1% and 5% of Sore and Geba watershed area. The individual area coverage and change statistics for the three periods 1987 to 2001, 2001 to 2017 and 1987 to 2017 are summarized in (Table 2). The results revealed from this study is consistent with the results of previous study in different parts of Ethiopia (Wubie, 2015;Bewket and Abebe, 2013).

Evaluation of accuracy
To evaluate the correctness of the classified image, the accuracy assessment is used. For this study confusion matrix was used. Using google earth image as a reference and the original mosaic satellite image, there were randomly selected reference points for each land cover types. With the corresponding classified map, the randomly selected points were compared.  Anderson (1976) for overall accuracy is 85%. Therefore, for all land cover maps, the Journal of Environment and Earth Science www.iiste.org ISSN 2224-3216 (Paper) ISSN 2225-0948 (Online) Vol.9, No.10, 2019 classification conducted for this study produces an overall accuracy that meets the minimum value suggested by Anderson (1976).
The producer's accuracy results for land use map of 1987, 2001 and 2017 ranges from 60% to 100%, 72.7% to 100% and 63.6% to 100% respectively. Whereas, the user's accuracy ranges from 75% to 96.7%, 85.7% to 100% and 88.9% to 100% respectively. Due to the spectral similarity of different land uses, the lowest values were misclassified.
The accuracy assessment results from the Kappa coefficient statistics is 0.85, 0.94 and 0.95 for the period of 1987, 2001 and 2017 respectively. For the period 1987, a kappa of 0.85 means that agreement is 85 percent better than by chance alone. Kappa of 0.94 for the period of 2001 means 94% better agreement than by chance alone. And there is 95% better agreement for the 2017 period than by chance alone.

Calibration and Validation
The model was calibrated using a sequential uncertainty fitting (SUFI-2) algorithm using the SWAT-CUP computer program. The model was evaluated using Determination Coefficient (R 2 ) and Nash-Sutcliffe efficiency (NS) values for fit measurement goodness. The value of R2 ranges from 0 meaning poor model performance to 1 showing good performance, and the value of NS ranges from negative infinity to 1. The larger the value of R 2 and NS the better the agreement between measured and simulated flows. The other model prediction uncertainty measures selected for this study was p-factor and r-factor. The p-factor value ranges from 0 to 100 %, while the rfactor value ranges from 0 to an infinity. A threshold value suggested by (Abbaspour, 2015) >70% of p-factor and having r-factor of around 1 is recommended for a better result of calibration of a model by SWAT-CUP. The pfactor shows the percentage of measured data bracketed by the predictive uncertainty of 95 percent (95PPU) and provides the capability of the models to capture uncertainties. The r-factor shows the 95PPU thickness and is a measure of calibration quality.
The model was calibrated for Sore and Geba watershed with Sore river gauging station. Figure 5 shows the five-year (2005-2010) calibration and (2011-2016) validation periods of measured and simulated monthly stream flows for the Sore and Geba watershed at the Sore river. The SWAT model accurately tracked the lowest observed stream flows but some of the highest observed flows in the first year of simulation was over predicted for both calibration and validation period. However, the simulated flows in calibration period is closely followed the observed flows than the validation period. The model performance measures statistics during calibration were stronger than those computed for validation period. According to a threshold values suggested by different scholars the computed statistics for Sore and Geba watershed showed satisfactory result for both calibration and validation periods. For example, p-factor 0.78, r-factor 1.02, determination coefficient (R2) 0.8 and Nash-Sutcliffe efficiency (NS) 0.79 during the calibration period, where p-factor 0.78, r-factor 1.26, determination coefficient (R2) 0.75 and Nash-Sutcliffe efficiency (NS) 0.54 are used for the monthly data simulation. The statistics generally show that there is a strong correlation between the values simulated and measured. As shown in Table 3, the simulated average monthly stream flow at Sore river gage for both calibration and validation periods were pretty much close to observed flows with the agreement slightly is better in calibration period. Overall, the model performance was good.  Mengistu and Sorteberg (2012) in the Baro-Akobo river basin to calibrate and validate the monthly streamflow simulation at the Gambela gauge station gives Nash-Sutcliffe efficiency (NS) 0.9 and Determination Coefficient (R2) 0.92 for the calibration period and 0.81 and 0.89 for the validation period.

Environmental implication of the observed surface runoff
The change of LULC has made a substantial consequence on the hydrology including surface runoff, stream flow, evapotranspiration, sediment loading and water yield of the study watershed. Vegetation cover helps to reduce the soil erosion by interrupting and dissipating the erosive power of rainfall, runoff and wind. It has also a role in reducing the volume of runoff through increasing the infiltration by following the root system and increases soil organic content which increase the aggregate stability of the soil. Within the study period there has been a decline of natural forests and expansion of agricultural lands. As can be quantified in this study the expansion of agricultural lands generates highest surface runoff. This highest surface runoff will accelerate the erosion process Journal of Environment and Earth Science www.iiste.org ISSN 2224-3216 (Paper) ISSN 2225-0948 (Online) Vol.9, No.10, 2019 such as detachment and transportation. Sediment loading was also highest due to the expansion of agricultural lands. All this phenomenon has an implication on the environment such as biodiversity loss, flooding and sedimentation on the downstream water storage structures.

CONCLUSION
LULC trend analysis within the periods of three decades from 1987 to 2017 in Sore and Geba watershed shows a significant change over the years. The area coverage of cultivated land, open forest, dense forest, wood land, and water bodies of Sore and Geba watershed in 1987 were 33%, 49%, 3%, 4% and 11% of the total watershed, respectively. Whereas, in 2017 open forest, dense forest, wood land and water bodies diminished to 44%, 0%, 1% and 5% of the total watershed respectively. Whereas the coverage of cultivated land increased to 50% of the total watershed. Cultivated land is gained from other types of land use. In the last three decades there has been an increase in scattered rural settlements closely associated with cultivated land. Due to deforestation, the highest land use and land cover change occurred between 2001 and 2017.
In the study watershed, nine sensitive flow control parameters were identified using SWAT-CUP computer program by SUFI-2 algorithm that were targets of the calibration process. During the calibration period, values of Nash-Sutcliffe (NS), coefficient of determination (R 2 ), p-factor and r-factor were 0.79, 0.8, 0.78 and 1.02 respectively. Whereas for the validation period the values were 0.54, 0.75, 0.78 and 1.26 respectively. For both calibration and validation period, this performance is viewed satisfactory. The effects of land use and land cover change on surface runoff and stream flow were evaluated after calibration and validation of the SWAT model. The observed LULC change shows a significant change on the hydrological process in the watershed particularly surface runoff, stream flow, evapotranspiration, and groundwater flow. The simulated average annual surface runoff from LULC map of 1987 was 189.28 mm while from land use map of 2017 the average annual surface runoff increased to 221.19 mm. This result showed that during the last three decades (1987-2017) LULC change indicated an increase of surface runoff by 16.86%. The average annual stream flow increased from 188.06 m 3 /s to 190.84 m 3 /s between the periods of 1987 and 2017 LULC change. The quantified LULC change and surface runoff has an implication on the environment, such as loss of biodiversity, loss of top soil by erosion, flooding and sedimentation problems in the downstream communities. This study suggests promoting non-timber forest products, planning and regulating the expansion of settlements and soil fertility management activities should be implemented. This would increase existing farming productivity and help in controlling the expansion of cultivated land. Finally, this study highlights the application of SWAT model integrated with GIS tools in the study basin provides a better understanding of the process of impacts of land use change in local hydrology.