Application of artificial neural network (ANN) for investigation of the impact of past and future land use–land cover change on streamflow in the Upper Gilgel Abay watershed, Abay Basin, Ethiopia

Abstract

Changes in land area features, such as vegetation type and soil conditions, have an impact on carbon sources and sinks and support food production; this is critical in addressing global sustainability challenges such as climate change, biodiversity loss, and food security. The study's major goal was to determine how LULC changes in the past and future might affect streamflow in the Upper Gilgel Abay watershed. The modeling was conducted using the MOLUSCE Quantum GIS plugins cellular automata simulation method and streamflow modeled using SWAT. Landsat 5 TM (1995), Landsat 7 ETM + (2007), and Landsat 8 imaging (2018) satellites were used to collect the images, which were then categorized using ERDAS 2014 software, and the kappa coefficient becomes 84.04%, 82.37%, and 85.54% for 1995, 2007, and 2018 LULC, respectively. SWAT model better performed the simulation which is R² of 0.77 for calibration and 0.68 for validation and E_NS becomes 0.71 and 0.62 for calibration and validation, respectively. The output change in streamflow due to past and future LULC maps shows an increase in LULC in cultivated areas and resulted in 39%, 46.81%, and 52.45% in each of the years 1995, 2007, and 2018, respectively. The three LULC modifications in the land cover maps from 1995, 2007, and 2018 had simulated mean monthly peak discharges of 62.20 m³/s, 66.51 m³/s, and 72.10 m³/s, respectively. The projected LULC 2027 also shows a similar increase in the study area, and dominantly cultivated land illustrates the highest change at around 53.77% but the highest change occurs on grassland during (2018–2027) land use at around 12.29%. And the highest streamflow was found around a monthly average of 1400 m³/s. The study primarily provides insight into how LULC fluctuation affects streamflow, and it is crucial for water planners and natural resource professionals whose focus is on the Upper Gilgel Abay basin.

Introduction

Background of the study

A change in how a certain land area is used or managed by humans is referred to as a land use change, while a change in some persistent characteristics of the land, such as the type of vegetation and the soil conditions, is referred to as a land cover change (Nedd et al. 2021). Since land use has a significant impact on carbon sources and sinks (Houghton et al. 2012; Arneth et al. 2014), destroys habitat (Powers and Jetz 2019), and supports food production, it is essential to address global sustainability issues such as climate change, biodiversity loss, and food security. Due to the mitigation potential of land use activities, particularly those related to agriculture and forestry, which have been accepted as essential in achieving climate targets under the Paris Agreement, land use has been a key topic in many international policy debates (Grassi et al. 2017). To support these discussions, it is crucial to measure and comprehend global land use change and its spatiotemporal dynamics. Among the highest in the world, sub-Saharan East Africa is putting further pressure on the conversion of land cover (Wolde et al. 2021). An estimation of the areas and trends in nine land cover classes in Ethiopia, Kenya, Uganda, Malawi, Rwanda, Tanzania, and Zambia from 1998 to 2017 was made using a planned sample of satellite-based observations of historical land cover change. In East Africa, cropland increased by 18,154,000 (1,580,000 hectares), or 34.8%, according to our analysis (Bullock et al. 2021). Several investigations on the changes in land use and land cover were carried out in Ethiopia (Negese 2021; Ogato et al. 2021; Kuma et al., 2022; Leta et al. 2021). One of the factors that keep coming up in the studies is the high population expansion. Ethiopia has a population of 73.9 million, up from 20 million in 1950 (Derbew et al. 2021). Between 1950 and 1960 and 1994 and 2007, the yearly growth rate grew as well, reaching 1.65% and 2.5%, respectively (Gashaw et al. 2014). Over the past few years, population growth has had a significant impact on natural resource availability. Rural communities in tropical nations like Ethiopia expand into the wet areas to create additional fields for agricultural cultivation (Zerssa et al. 2021). The strongest indicator of this is the humid area, which makes up more than half of the country's total size and sustains a larger portion of the rural population who rely mostly on agriculture for their livelihood (Barros et al. 2015).

Eighty-four percent of the people in Ethiopia rely on agriculture for subsistence. Land use has changed over time and space as a result of increasing agriculture, urbanization, deforestation, and human activity. Land cover has an impact on water flow patterns and water balance (Jamal and Ahmad 2020). The negative effects of land use and land cover change are wide ranging and threaten one of the foundations of the nation's economy and the welfare of its citizens. Also, both temporal and spatial changes in LULC have had a consistent impact on a variety of water resource development sectors (hydropower, irrigation, urban and rural water supply, etc.) (Welde and Gebremariam 2017). Land cover and land use patterns have changed as a result of both urban and agricultural activities.

The Upper Gilgel Abay watershed is one of the headwaters of the Tana sub-basin in the study area. The watershed is one of the most populated and agriculturally prosperous areas in the Upper Blue Nile Basin. Many problems, including deforestation, overgrazing, development of croplands, soil erosion, and the expansion of agricultural land, affect the watershed (Gashaw et al. 2020; Debie and Anteneh 2022; Gebrehiwot et al. 2014). Additionally, (Kidane et al. 2019; Alemu et al. 2022) also examined how LULC affected streamflow. There are numerous researchers studied LULC predication (Gidey et al. 2017; Tadese et al. 2021) as well as changes in LULC on hydrological streamflow in the Upper Gilgel Abay watershed, including (Mulu and Dwarakish 2016; Gashaw et al. 2020; Abuhay et al. 2022), intending to determine how changes in land use and land cover (LULC) impacted the watershed's hydrology, which is found in the Upper Blue Nile Basin. Unlike those studies, the article investigates the variation of future LULC on the streamflow with respect to the actual LULC. In addition, none of them make predictions on how such changes may affect the Upper Gilgel Abay watershed in future. Environmental degradation brought on by these changes has adversely affected the hydrologic regimes of several Ethiopian catchments, notably the watershed. For water resource planning, it is essential to comprehend the spatiotemporal LULC fluctuations and their effects on significant hydrological components. The catchment has a large population that depends on scarce natural resources like wood and water and is situated in a high-potential area where intensive agriculture is conducted (Bogale et al. 2023; Khadim et al. 2020). Hence, it is essential to analyze the change in future LULC on the streamflow in the study area.

To evaluate the effects of LULC changes on hydrological processes and simulate future changes in land use of the Upper Gilgel Abay watershed and Upper Blue Nile Basin between 1986 and 2021, this study used an artificial neural network (ANN) and the Soil and Water Assessment Tool (SWAT) model to examine the effects of previous and future LULC changes on streamflow.

Materials and methods

Description of the study area

The Upper Gilgel Abay watershed is located in Ethiopia (Fig. 1 A) particularly at Abay Basin (Fig. 1 B), which drains an area of around 1665.65 km², and is the largest tributary in the Lake Tana sub-basin. It comes from the highland spring in the northeastern Ethiopian town of Gish Abay. It supplies Lake Tana with 44.2% of the water. The watershed is located in the southern Lake Tana sub-basin and the northern Upper Blue Nile basin, with latitudes and longitudes that range from 10° 56′ 53′′ to 11° 21′ 58′′N and 36° 49′ 29′′ to 37° 23′ 34′′E, respectively, and the elevation ranges between 1887 and 3534 m above mean sea level.

Fig. 1

Area of study watershed of Upper Gilgel Abay

The Upper Gilgel Abay watershed has a cold, semi-humid climate, with mean monthly temperatures ranging from 17.5 to 20.4 °C. From October and May is primarily the dry season, and between June and September is primarily the wet season. In general, the climate is tropical at lower levels and temperate at higher elevations. The Inter-tropical Convergence Zone (ITCZ) moves north to south, bringing moist air from the Atlantic and Indian oceans into the Upper Gilgel Abay basin (Kebede et al. 2006). It has a wide variety of climates, from humid and moderate in the southwest to semiarid desert types in the lowlands. The Upper Gilgel Abay watershed's long-term mean annual rainfall (1995–2018) is 1478 mm.

Data collection and analysis

The hydrologic model for this study was constructed using a variety of datasets, including observed streamflow, meteorological data (rainfall, maximum and lowest temperatures, wind speed, sunshine hour, and relative humidity), land use, soil maps, DEM, etc. Self-prepared Landsat photographs were created using Google Earth images (Table 1).

Table 1 Data collecting techniques for the SWAT model and source

Hydrological data (flow data)

The hydrological data of Upper Gilgel Abay River flow data gathered from the Ministry of Water Resources Ethiopia (MWRE) from six gauging stations include a longer period, spanning 1995 to 2012 for a total of 18 years. The flow data were taken into consideration for the sensitivity analysis, calibration, and validation of the model. To comprehend the correlations between model parameters and watershed characteristics, these data have been used to calibrate and validate the SWAT model. The data from 1995 to 2012 have high quality, according to the findings of hydrological inspections (Fig. 2).

Fig. 2

Monthly average flow of Upper Gilgel Abay station (1995–2012)

Drainage pattern

Slope

The slope of a catchment has a significant impact on flow velocity; at smaller slopes, the equilibrium between rainfall and runoff can be temporarily stored throughout the area and can progressively drain out over time. Hence, a rise in surface slope was accompanied by a rise in surface runoff. The same slope ranges employed in the SWAT model were utilized for the slope categorization in ArcGIS (Fig. 3).

Fig. 3

Upper Gilgel Abay watershed slope map

Soil data

One of the main inputs to the SWAT model is data on the soil. The information came from the Ethiopian Ministry of Water Resources and Energy (MWRE). The Food Agriculture Organization of the United Nations provided the data used in this study's soil classification (Table 2). Geoprocessing in Arc View GIS was used to trim the soil map for the research area using the watershed shape file (Fig. 4). Haplic Alisols and Haplic Luvisols are the dominated soil types in the watershed. Abuhay et al. (2023) also found out similar soil type on the study area (Fig. 5).

Table 2 FAO soil classes of Upper Gilgel Abay watershed

Fig. 4

Upper Gilgel Abay watershed soil classification map 2010 E.C

Fig. 5

The location of the Upper Gilgel Abay watershed's meteorological stations

Meteorological data

The daily statistics on precipitation, maximum and minimum temperatures, relative humidity, wind speed, and solar radiation are among the weather input variables used for SWAT simulation. The Amhara Meteorological Agency provided these (AME). Adet, Dangla, Wotet, Abay Sekela, and Enjibara stations, as well as other weather stations in and near the Upper Gilgel Abay watershed, provided the meteorological information from 1995 to 2018 (Fig. 6). While the remaining three stations (Wotet, Sekela, and Enjibara) (Fig. 5) only have records for temperature and rainfall, the first two stations (Adet and Dangla) are the first classes to have records on all climatic variables (Table 3). The two weather stations provided estimates for the monthly generator parameters (Adet, Dangla).

Fig. 6

Average monthly rainfall of the five stations in Upper Gilgel Abay

Table 3 Meteorological locations, station names, elevations, and variables

Land use–land cover data

The study used downloaded Landsat data to examine the dynamics of land cover and land use in the Gilgel Abay basin over the past 24 years. Three land uses with a 12- and 11-year gap (1995, 2007, and 2018 GC) have been constructed using Landsat photographs for this investigation. The land use maps from 1995, 2007, and 2018 were created using multispectral imagery from Landsat 4–5, Landsat 7 ETM + , and Landsat 8 ETM + C1 level 1, respectively. The spatial resolution of the pictures used to create the land use maps for the research area is 30 m by 30 m, and all of the images were gathered from photographs taken from January to March to ignore the effect of cloud cover. There is no seasonal fluctuation in the images, and categorization entails measuring temporal impacts using multitemporal data sets, but the resolution of the image and the study's objectives determine the land class units. The research region has been identified using a supervised classification based on field observation and GPS coordination data (Table 4).

Table 4 The resources utilized in this investigation and the software used

Summary of materials used in this study

Methodology of the study

Image processing

The land use maps for the years 1995, 2007, and 2018 were prepared using Landsat MSS-5, Landsat 7, and Landsat 8 levels 1, respectively. Using the composite band tool, all of the satellite image bands that were accessible were blended. Using the Arc GIS 10.1 software program, the combined Landsat data in the GCS WGS 1984 raster format was converted into the Universal Transverse Mercator (UTM) projection raster form by taking into account the zone of the study area WGS 1984, UTM Zone 37N. Red, Green, and Blue (RGB) systems were used to composite all of the images to improve viewing. Several band combinations still revealed the true hue of the land sat picture, even though many various band combinations can be formed for classification (Helber et al. 2019). The genuine color combinations for Landsat MSS-5 (3,2,1; Lyzenga 1981), Landsat 7 (4, 3, 2; Moges and Bhat 2017), and Landsat 8 (4, 3, 2 band combination; Darge et al. 2019) (Table 5).

Table 5 The acquisition dates, sensor, path/row, resolution, and the producers of the images

Image classification

Using the ERDAS IMAGIN 2014 software, there are three alternative classification algorithms for LULC classification (Toma et al. 2023). However, the maximum likelihood approach was used for this classification since it classified the land cover classes better than the minimal distance algorithm (Priyadarshini et al. 2018). As a result, the maximum likelihood classifier was employed moving forward in the investigation. The method most frequently employed for quantitative assessments of remote sensing picture data is supervised classification (Rwanga and Ndambuki 2017; Priyadarshini et al. 2018). This is accomplished by choosing representative sample locations, also known as training sites or areas, of a specified cover type. The computer algorithm then classifies the entire image using the spectral signatures from these training areas. The methods that went into creating the land cover map, which was based on pixel-based supervised classification, include first choosing the training sites, which are typically representative of the land cover classes.

Based on real color band combinations, the training locations were compiled. Moreover, image retouching and composition were used to improve the differentiation of the land cover classes (Wang et al. 2019). These methods resulted in the collection of 45 training sites from each image (1995, 2007, and 2018). Next, classify the images using the maximum likelihood classifier. Finally, evaluate the classification accuracy by comparing the classified images to Google Earth images. Random samples of points from each class were chosen for 1995, 2007, and 2018 maps, and the confusion matrix was analyzed. Ground control points with coordinates gathered from Google Earth Pro are used as a training sample set for supervised classification. Since these band combinations enable viewing of the photographs in their actual color, the sample set for the classification is constructed using the combination of bands 3, 2, and 1 (for images from 1995 to 2007) and bands 4, 3, and 2 (for the image from 2018). The study area is divided into five groups, including cultivated land, forest, grassland, shrubs land, water, and marshy land, following Andualem and Gebremariam (2015) and Gumindoga et al. (2014) classification of the watershed and with extra real values from Google Earth (Table 6).

Table 6 Land use–land cover classification of Upper Gilgel Abay watershed

Accuracy assessment of classification

One of the crucial last processes in the classification of land use and land cover change is the accuracy evaluation, whose goal is to analyze how well the pixels were sampled into the appropriate land cover class (Zhao et al. 2019). Google Earth pictures from 1995, 2007, and 2018 were used in this study's accuracy assessment, and ground-truth measurements were taken in the study region to determine how accurately the land uses were classified. Random points chosen for proof were compared to the categorized maps using Google Earth photographs. Similar to this, 140 random points taken for 2018 and 135 random points taken for 1995 and 2007, were utilized to validate the identified photographs. The image map is used to create the confusion matrix. Nonetheless, overall accuracy was determined by dividing the total number of ground-truth sample units by the sum of the correctly identified sample units. The user's accuracy and the producer's accuracy evaluated each category's accuracy with regard to commission and omission errors (Olofsson et al. 2014). The probability that a pixel has been accurately classified by the user is defined as the accuracy, while the accuracy of the producer is defined as the likelihood that a pixel represents the ground on the map (cite). There are several ways for evaluating the accuracy of land use classification available (Islam et al. 2021), but the kappa coefficient (K) is by far the most effective approach (Eq. 1). The Khat statistics (an estimate of kappa) produced by the kappa analysis serve as a gauge of agreement or correctness (Congalton 1991).

$$K = \frac{{\mathop \sum \nolimits_{i = 1}^{r} Xii - \mathop \sum \nolimits_{i = 1}^{r} \left( {X_{{1 - X_{X + 1} }} } \right)}}{{N^{2} - \mathop \sum \nolimits_{i = 1}^{r} \left( {X_{{ii + X_{x + 1} }} } \right)}}$$

(1)

where r = number of rows and columns in the error matrix, N = the total number of observations (pixels) Xii = observation in row i and column I, X_ii +  = marginal total of rows i, and x + 1 = marginal total of column I. Perfect agreement is indicated by a kappa coefficient of 1, whereas a number near to 0 indicates that the agreement is just slightly better than would be predicted by chance. The category of kappa statistics is frequently used and replicated (Landis and Koch 1977) (Table 7).

Table 7 Rating criteria of kappa statistics

Preparation of confusion matrix

The creation of a classification error matrix, also known as a confusion matrix or a contingency matrix, is one of the most popular ways to describe classification accuracy (Comber et al. 2012). The classification results are contrasted with extra ground-truth data in the confusion matrix. Both the nature and the quantity of categorization errors are identified (Table 8).

Table 8 Generalize algorithm of accuracy assessment

2.5 Transitional potential modeling

Artificial neural network (ANN): There are components of computational intelligence. Artificial neural networks' fundamental component is the adjustment of the weight connections between connected neurons and their interactions (Maind and Wankar 2014). This adjustment is dependent on both the input data and the anticipated network output.

Many software programs are available for simulating future LULC scenarios, including LCM (Land Change Modeler), CA-MARKOV, DINAMICA-EGO, and CLUE-S, which use empirical methodologies based on prior LULC. But recently, the Modules MOLUSCE (QGIS Plugin) became available for assessing present LULC changes and predicting future LULC and found out it is easy and gives better LULC predicted simulation output (Lukas et al.2023; Kamaraj and Rangarajan 2022; Gao et al. 2023). Utilizing four different models, including a simulation model that has been trained and a transition potential/possibility matrix built using artificial neural networks (ANNs) and multi-criteria evaluation (MCE) (Sajan et al. 2022). Using a Monte Carlo cellular automata model, the simulated land use map is created (Sajan et al. 2022).

Prediction of future land use–land cover

To predict LULC, two Landsat classified maps were used as a reference (2007 and 2018) and preparation of the slope and aspect of the study region were employed as input variables for the transition model. These data were then fed into the MOLUSCE Plugin to create a map of the change in land cover and determine the trend of change for the research area between 2007 and 2018. The area change map produced by the plugin displays the change in the land from 2007 and 2018 in all five classes, including cultivated land, forest land, grassland, shrubland, and water, according to the dominant reference coordinated land use areas as well as Google Map classification point. The plugin measures the percent of area change in a given year and provides a transition matrix that shows the proportions of pixels changing from one land use cover to another land cover. An approach that MOLUSCE used in this investigation was the artificial neural network (ANN). To project the change in LULC, a cellular automata simulation approach was employed in the plugin. Cellular automata is an effective method that uses transitional potential modeling (ANN) to forecast future changes (Basse et al. 2014; Qiang and Lam 2015; Sajan et al. 2022).

Flow charts of the MOLUSCE plugin in QGIS

2.6. SWAT modeling

A SWAT model's primary usage is to simulate the watershed under different land use, land management, and climate change scenarios (Fan and Shibata 2015). The SWAT model is a physically based, constantly operating watershed model that replicates the hydrological processes of the catchment. To process the datasets and create the necessary input for the initial modeling setup, the SWAT model is integrated with Arc SWAT in the ArcGIS Geographical Information System interface. The hydrologic cycle's equation for water balance is (Eq. 2):

$${SW}_{t}={SW}_{O}+\sum_{i=1}^{t}({R}_{\mathrm{day}}-{Q}_{\mathrm{surf}}-{E}_{a}-{W}_{\mathrm{Seep}}-{Q}_{\mathrm{gw}})$$

(2)

where ${SW}_{o}$ is the initial soil water content on a day i (mm), ${SW}_{t}$ is the final soil water content, ${Q}_{\mathrm{surf}}$ is the amount of surface runoff on a day i(mm), E_a is the amount of evapotranspiration on a day i(mm), ${Q}_{\mathrm{gw}}$ is the amount of return flow on a day i, (mm), ${R}_{\mathrm{day}}$ is the amount of precipitation on a day i(mm), ${W}_{\mathrm{seep}}$ is the amount of water entering to vadose zone from the soil profile on a day i(mm), and t is the time in days.

Surface runoff was estimated using the Soil Conservation Service Curve Number (SCS-Curve number) method (Eq. 3):

$${Q}_{\mathrm{surf}}=\frac{({{R}_{\mathrm{day}}-{I}_{a})}^{2}}{({P}_{\mathrm{day}}-{I}_{a}-S)}$$

(3)

where I_a is the initial abstraction which includes surface storage, interception, and infiltration before runoff and S is the retention parameter (mm).

Retention parameter is defined by (Eq. 4):

$$s=25.4\left(\frac{1000}{CN}\right)-10$$

(4)

where CN is the curve number for the day, which varies from 0 to 100 depending on soil permeability, land use, and the antecedent soil water condition. The initial parameter is approximated as 0.2 S; the equation becomes (Eq. 5):

$${R}_{\mathrm{sur}}=\frac{({{P}_{\mathrm{day}}-0.25)}^{2}}{({P}_{\mathrm{day}}+0.85)}$$

(5)

Figure 7 illustrates the overall stages of SWAT modeling using the process: the first one is data preparation, watershed delineation, HRUs definition, Run SWAT simulation, sensitivity analysis, calibration, and validation, Rerun SWAT simulation, and then simulate streamflow.

Fig. 7

Detail flow chart of simulation/prediction of Upper Gilgel Abay LULC using QGIS (MOLUSCE)

Sensitivity analysis

The most sensitive parameters that have a substantial impact on the model calibration and validation can be found through sensitivity analysis (Gou et al. 2020). Sensitivity analysis explains how a given input variable's range of effects on model output using data from both the actual and simulated flows, sensitivity analysis was carried out. The SWAT Calibration and Uncertainty Procedures (SWAT-CUPsemi-automated)'s Sequential Uncertainty Fitting (SUFI-2) calibration method was employed for calibration and validation (Kumar et al. 2017).

There would be a difference between measured data and simulated results when a SWAT simulation was conducted. The parameters that are influencing the outcomes and the degree of variation must therefore be identified to reduce this disagreement. Hence, to verify this, sensitivity analysis one of the SWAT model's tools was requested to reveal the rank and the mean relative sensitivity of parameter identification. The p value, which is the number you need to be looking at, is determined by comparing the parameter sensitivity based on the t-stat of a parameter with the values in the Student's t-distribution table (Abbaspour 2015) and The Student’s t-distribution describes how the mean of a sample with a certain number of observations is expected to behave (Hooft et al. 2010).

The p value for each term evaluates whether the coefficient is equal to zero and has no effect, which is the null hypothesis. If your p value is low (0.05), you can rule out the null hypothesis. As a result of the relationship between changes in the predictor's value and changes in the response variable, a predictor with a low p value is likely to be a useful addition to your model. A higher p value, on the other hand, indicates that there is no correlation between changes in the predictor and those in the response. That parameter is therefore not particularly sensitive. The generally recognized threshold for rejecting the null hypothesis is a p value of less than 0.05. (i.e., the coefficient of that parameter is different from 0). One may assert with a 95% probability of accuracy that the variable is having some influence because there is only a 5% chance that the findings that are seen would have occurred in a random distribution when the p value is 0.05 (Abbaspour et al. 2007).

Model performance criteria

To gauge how well-simulated outcomes matched actual data in the watershed, the statistical performance indices Nash and Sutcliffe efficiency (NSE) and coefficient of determination give good agreement on the study area (Abuhay et al. 2022; Andualem and Gebremariam 2015). The R² statistic offers a measure of how well the variation of observed values is duplicated by the model predictions and ranges from 0 to 1, where 0 indicates no correlation and 1 represents perfect correlation (Arnold et al. 2012). The R² statistic is derived as follows (Eq. 6):

$${R}^{2}=\frac{\sum_{i=1}^{n}\left({x}_{i}-\overline{x }\right)\left({y}_{i}-\overline{y }\right)}{(n-1){s}_{x}{s}_{y}}$$

(6)

NSE values can range between −∞ and 1 and a perfect fit between the simulated and observed flow is indicated by an NSE value of 1. NSE values ≤ 0 indicate that the observed data mean is a more accurate predictor than the simulated output. Both NSE and R² are biased toward high flows. The total is calculated throughout the entire time frame of the calibration data. The variance is better explained by the model when the value is nearer to unity. A poor modeling efficiency indicates that the model's forecast is less accurate than when the mean of the observed flows is used instead (Nash and Sutcliffe 1970) (Eq. 7).

$$\mathrm{NSE}=1-\left\{\frac{\sum_{\mathrm{i}=1}^{\mathrm{n}}({{\mathrm{Z}}_{\mathrm{Obs}}-{\mathrm{Z}}_{\mathrm{simu}})}^{2}}{\sum_{\mathrm{i}=1}^{\mathrm{n}}({{\mathrm{Z}}_{\mathrm{Obs}}-{\mathrm{Z}}_{\mathrm{mean}})}^{2}}\right\}$$

(7)

Results and discussion

This study's findings were divided into three categories. The accuracy of land use classification results for the years 1995, 2007, and 2018 is broken down into two parts. The daily flow calibration and validation results of the first-class (1995–2018) SWAT model are described in the second section. The outcome of the modeling and prediction of the LULC change and its effect on the streamflow of the Upper Gilgel Abay watershed is shown in the third section (Fig. 8).

Fig. 8

Overall workflow chart for the SWAT model simulation

Analysis of land use–land cover change

According to the results of the LULC analysis of the watershed of Upper Gilgel Abay, there were considerable changes between 1995 and 2018. Kidane et al. (2019)'s discovery of the Upper Blue Nile River's substantial differences between these periods lends credence to this assertion. Figure 9A–C shows the land cover maps that were gained for 1995, 2007, and 2018, respectively.

Fig. 9

Land cover map of the Upper Gilgel Abay watershed in 1995 (A), 2007 (B), and 2018 (C)

According to the 1995 cover map, the watershed was covered by 39.16% farmland, 37.75% forest, 3.08% grassland, 19% bushes, and 0.57% water and marshy land. Table 9 and Fig. 8A also illustrate the area coverage of each land cover type. In contrast, much of the catchment's agricultural land increased in size with a similar distribution of cultivated land found during such period due to the rapid increase of deforestation, population growth, and habitation growth (Gashaw et al. 2020; Andualem and Gebremariam 2015).

Table 9 Land use–land cover classification results in 1995, 2007, and 2018 maps

Following this, Fig. 8B also depicts the land cover in 2007 revealing that around 46.81% of the watershed was covered by cultivated land, 23.78% by forest, 17.81% by shrubland, 11.58% by grassland, and very little by water and marshy land. According to the distribution of land cover classes depicted in the figure, cultivated land cover was present in the majority of the watershed, particularly in the northern portion. Also, the coverage shown in Fig. 8C for the year 2018 shows an increased change of about 52.45% over cultivated land, 25.75% in forest, 13.38% in shrubland, 8.39% in grassland, and a small change in water and marshy land.

Overall, the cultivated area increased by 7.66% while the forestland fell by 13.97% between 1995 and 2007. From 2007 to 2018, cultivated land increased by 13.30%, while forestland decreased by 12.00%. From 1995 to 2018, cultivated land increased by 5.64%, while forest land increased by 1.97%. The majority of the catchment area was converted into agricultural lands (cultivated land) during this time due to the high growth in population density, which also caused a decrease in the amount of marshland and forest cover (Wubie et al. 2016; Tolessa et al. 2020). This led to a shortage of agriculture in low-land areas.

Accuracy analysis

Table 10 displays the user accuracy of the Landsat 5, Landsat 7 ETM + , and Landsat 8 ETM + C1 image-supervised categorization results. Table 10 A depicts classified Landsat 5 and the results of cultivated land, woodland, grassland, shrubland, and water all having user accuracy of 91.17%, 87.87%, 88%, 84%, and 83.33%, respectively. The highest accuracy countered on cultivated land; this meant that it had been correctly classified. And the minimum was found at water land. The classification was accurate since this image's average kappa coefficient was 84.04%. On the other hand, the Landsat 7 ETM + image in the year 2007 shows that 91% of the highest accuracy is found in cultivated land, 84.3% becomes forest land, and 88%, 80%, and 82% become grassland, shrub land, and water land, respectively. And the final kappa coefficient becomes 82.37% (Table 10 B).

Table 10 The confusion matrix of the LULC classifications of 1995 (A), 2007 (B), and 2018 (C)

The final Landsat 8 TM image year 2018 supervised classification results, the user accuracy, and producer accuracy were shown. The user accuracy of cultivated land, forest, grassland, shrubland, and water was 90%, 86.6%, 95.8%, 88%, and 82%, respectively. The maximum user accuracy was cultivated land, which indicated the correctly classified. Moreover, the minimum was water land and the overall accuracy was 82%. The average kappa coefficient of this image was 85.54%, which represented the classification was accurate (Table 10 C).

3.3. Modeling and prediction of future land use–land cover

Correlation evaluation

Modeling of LULC prediction includes different input driving variables including the slope and aspect variables' correlation ratios, which are shown in Fig. 10A and B, respectively. According to Szabó et al. (2015), the aspect is the angle formed by the steepest slope's direction and the vertical. Additionally, both of them were computed using a digital elevation model, which is a grid of elevations.

Fig. 10

Slope of Upper Gilgel Abay watershed (A) and Upper Gilgel Abay watershed aspect (B)

Area change land use change from 2007 to 2018

Table 11 illustrates the reference classified land use map used for the prediction of LULC, and it displays the past land use cover variations in the Upper Gilgel Abay watershed between the classification maps from 2007 (Fig. 9B) and 2018 (Fig. 9C). The change shows a decrease in shrubland and an increase in cultivated land. It is evident that during the research period, cultivated land and forest replaced the grassland and shrubland. It indicates the agricultural land in the study area increases over time. Between 2007 and 2018, the area of land that was farmed increased by 5.64% (Table 12). Abuhay et al. (2023) and Tasgara and Kumar (2023) indicate an increment of LULC on the Upper Gilgel Abay causes a fluctuation of streamflow.

Table 11 Land use change from 2007 to 2018 (area change %) from the MOLUSCE plugin

Table 12 Transition matrix of land use cover in 2007 and 2018

On the other hand, the forest land area has significantly increased with a total percentage of 1.97% and the shrubland decreased by 4.43% the grassland and water also decreased as shown in Table 11. Generally, all land uses have different trends of change. Agricultural area increased, whereas others decreased, removing the other land uses.

Transition matrix

The transition matrix represents the replacement of one land use to change to another land use (Berihun et al. 2019). The highest probability change in shrubland is that transports to cultivated land, while the lowest probabilities were presented in water. Other land use activities, such as shrubland and grasslands, have taken the place of the cultivated area in Table 11. The regeneration of farmed land usage in the study area is what caused this condition. It demonstrates that changes occur very quickly within a short period, and other land cover classes alter following earlier land cover change transition matrices.

Cellular automata simulation

Using the variation in LULC between 2007 and 2018 as a basis, the forecast anticipated future trends and changes in the research field. The effective cellular automata approach is utilized by Basse et al. (2014) to forecast future changes. Additionally, the geographical variables that I employed, slope and aspect, have an impact on this strategy. Based on the input data from the previous two years' land use and land cover maps of the research region, Fig. 11's map depicts the change in each land use for the 10 years beginning in 2027. The grid code of the image and the coverage area of the land cover class of the research region were used by the model to provide the results.

Fig. 11

Predicted Upper Gilgel Abay watershed land use–land cover for 2027 using MOLUSCE

According to Table 13, the probability of cultivated land was covered by 53.77% and that of forestland was covered by 24.78%, but that of grassland, shrubland, and water was covered correspondingly by 20.68%, 2.74%, and 0.01% of the entire research area.

Table 13 Predicted land use–land cover change in 2027 by ANN modeling in MOLUSCE

3.4. Sensitivity analysis

Using the daily observed flow data, the sensitivity of the Upper Gilgel Abay watershed was assessed to determine the most sensitive parameter and to further calibrate the simulated streamflow. Hence, 18 influential parameters were selected for sensitivity analysis (Adem et al. 2016; Tassew et al. 2019; Abuhay et al. 2023), and from those, 12 parameters’ flow gives a reasonable sensitive performance on the study area, which is shown in Table 14. The outcome shows that the SWAT model of the watershed is more sensitive to the examination of the 12 factors that were chosen. SLSOIL: Slope length for lateral subsurface, CN2: Initial SCS runoff curve number flow and SOL_AWC; Available water capacity of the soil layer gives the highest sensitive parameters which are the watershed is highly suspected to those parameters similar findings on the watershed conducted (Abuha et al. 2023). It seems the watershed is more suspected of runoff and direct impact on the soil water capacity and causing serious flood issues (Pour et al. 2020). According to the P-value and t-statistic values, the first seven parameters were extremely sensitive, while the other ones are moderate. In the SWAT-CUP model, these delicate parameters are regarded as model calibration. The calibration period for the models is 1995–2006, while the validation period is 2007–2012.

Table 14 More sensitive parameters and their rank with value for streamflow

SWAT model calibration and validation

Following model calibration and validation, sensitivity analysis is performed. For the warm-up period from 1995 to 1996 (1-year data), the calibration period from 1997 to 2006, and for validation from 2007 to 2012, the data are divided into 80% for calibration and 20% for validation. For the model calibration, initial LULC classified map of 1995 was used for conducting calibration and validation of initial streamflow analysis and the corresponding LULC which is 2007, 2018, and predicting 2027 land use map used simultaneously for validation of streamflow analysis variation on the study area. Streamflow calibration sensitivity analysis was performed manually and automatically until the model output closely matched the observed data. The SWAT operates on monthly output intervals but receives daily input data.

The results of this study were better graphically represented by using the monthly output intervals. With an E_NS simulation efficiency of 0.71 and R² of 0.77, the results of the calibration for the average monthly streamflow over the Upper Gilgel Abay watershed demonstrated good agreement between observed and simulated streamflow (Fig. 12). The validation for monthly streamflow using the sensitivity parameters listed in Table 15 resulted in E_NS of 0.62 and R² of 0.68. Similar research was done by JEMBERIE et al. (2016) at the Dedissa sub-basin, where the amount of cultivated land increased by 12% and the amount of forest area fell by 3.61%. The SWAT model's effectiveness was also examined. Andualem and Gebremariam (2015) also obtained findings with the values of E_NS and R² on the calibration showing 0.72 and 0.77, respectively. Both the calibration and validation results involving observed and simulated streamflow had E_NS and R² values of 0.7 and 0.79, respectively. However, the validation appears to have produced results with a little different magnitude of 0.77 in E_NS and 0.88 in R². According to this article, SWAT can produce outcomes that are adequately reasonable, and depending on how they perform, the results might be very comparable to what they were in the past.

Fig. 12

The calibration and validation plot of observed, simulated, rainfall and 95PPU of monthly streamflow

Table 15 Mean monthly streamflow for calibration and validation

Table 15 shows that there is good agreement between observed and simulated streamflow in the validation and calibration of the streamflow result. As a result, the streamflow findings (Table 15) show that the Upper Gilgel Abay watershed's streamflow can be accurately predicted using the SWAT model. Figure 9 shows a comparison of the Upper Gilgel Abay station's monthly observed and simulated discharges across the calibration and validation periods. With higher values for the coefficient of determination and Nash–Sutcliffe efficiency (ENS) for the Gilgel Abay watershed, similarly, Setegn et al. (2008) demonstrated good agreement between the observed and simulated flows using SUFI-2 algorithms.

The measure used to evaluate the accuracy of the calibration was the proportion of data included within the 95% prediction uncertainty (95PPU) (Setegn et al. 2010). The overall impact of the hydrological model, parameter, and input data uncertainties is the 95PPUs. Figure 12 shows the comparison of the monthly observed, best-simulated streamflow rainfall with the 95% prediction uncertainty (95PPU) of the SWAT model, which depicts rainfall more precisely and with lower uncertainty.

The scatter plots of observed and simulated values for calibration and validation are shown in Fig. 10A, B. Between the simulated and observed values, there is a strong linear correlation (R²) of 0.76 for calibration and 0.59, indicating that the model correctly predicted the observed streamflow values.

The impact of predicted land use–land cover change on streamflow

On a monthly and seasonal level, the effect of anticipated land cover change on river flow has been examined. By comparing the observed river flow from 1997 to 2012 with the 2027 land cover map, the effect of land cover change on river flow has been examined. Table 16 shows the examination of the projected mean monthly streamflow for 2027, which is 72.10 m³/s, revealing an increase from previous streamflow studies. The findings show that between 1995 and 2027, streamflow rose by 9.90 m³/s and by 38.57 m³/s during the wet season (July to September). The flow drops by 4.64 m³/s during the dry season (January to March) over the years 1995 to 2027.

Between 1995 and 2018, the Upper Gilgel Abay watershed's mean monthly streamflow increased by 6.46 m³/s. Additionally, the simulated streamflow result between 1995 and 2027 land cover reveals an increase in streamflow of 9.90 m³/s.

Table 16 Mean monthly simulated river flow (m³/s) of the study area (1997–2027)

According to the impact of land cover change on streamflow, streamflow has generally decreased in dry seasons and may even increase in wet seasons over the previous 24 years of land use change and for the future 10 years as well. If future water resources planning and management are not anticipated considering the change in the land cover, there may be a significant challenge in balancing the water demand of the river caused by an increase in human population, stabilizing aquatic ecosystems, agroforestry, etc.

3.7. Evaluation of river flow due to land use change

Hydrological responses to land cover change

This study's evaluation of the effects of current and projected changes in land use and cover on the streamflow in the Upper Gilgel Abay watershed was one of its key goals. The influence of land use change has been included in the study at various time scales, including monthly and seasonal.

The change in average monthly streamflow values for the years 1995, 2007, and 2018 LULC maps for the calibration and validation of the streamflow from (1997–2012) was 62.20, 66.51, and 68.66 m³/s, respectively, and demonstrates a little variability in change. The relationship between rainfall and runoff is significantly impacted by changes in land use and cover.

Land use changes have been analyzed using land uses of three different periods that show the variation and extent of land use in the past few decades and the present. The land use change affects river flow as indicated in Table 17. In the calibration and validation period, the mean monthly streamflow had increased by 4.31, 2.15, and 6.46 m³/s for the land use change periods between (1995–2007), (1995–2007), and (1995–2018), respectively. Mostly, the streamflow has increased throughout the study period over 24 years. The change in streamflow was due to land use–land cover change (mainly due to an increase in cultivated land). The streamflow component of the river flow shows an increase in all the land use change periods. According to Mechal et al. (2022), such cases are mainly due to a population increase that demands additional land for agriculture, settlement, and infrastructure seems the same with the Ziway Lake watershed. Land use and land cover change very often due to the growing population and economy, in human history land, a fundamental factor of production has been coupled with economic growth (Melese 2016). The interaction of human and natural systems could influence the watershed water cycle, which can then bring variation in river flow. For instance, the reduction of forests and shrubs lands, caused by population growth, might lead to a surface runoff increase (Dibaba et al. 2020).

The mean annual streamflow increased, as indicated in Table 18, as a result of various maps showing changes in land use and cover. The outcome demonstrated that, on the land cover maps from 1995, 2007, and 2018, respectively, the streamflow is 744.34, 766.87, and 803 m³/s over the calibration period. Change in land cover maps' respective detection rates was 22.53, 36.13, and 58.66 m³/s. This illustrates how changes in land use and land cover affect streamflow and cause watersheds.

Table 17 The simulated mean monthly streamflow (from 1997 and 2012) due to three LULC maps (1995, 2007, and 2018)

Table 18 Annual mean streamflow change from (1997–2012) for each land use–land cover type (1995, 2007, and 2018)

Seasonal streamflow Upper Gilgel Abay (1995–2018)

This procedure produced discharge outputs for both land use and land cover patterns, which were compared. The discharge change during the wettest (July–September) and driest (January–March) streamflow months was calculated and used as an indicator to determine the impact of land use and land cover change on the streamflow (Table 16).

Taking into account the streamflow measurements during the wet season (July through September) and the dry season (January through March) to discover changes in streamflow (Table 16), for the years 1995 to 2018, the mean monthly streamflow during the rainy months grew by 34.76 m³/s. On the other hand, streamflow during the dry season decreased by 3.43 throughout the period from 1995 to 2018. The amount of farmed land rose during the research periods, reflecting this tendency. Studies such as Chakilu and Moges (2017) and Gashaw et al. (2018) demonstrate in various regions of the area how the impact of changing land use and land cover on streamflow evaluates, indicating that the change in land use and land cover has been influencing the catchment's low flow or dry season flow. As a result, according to the findings, low flow and high flow in the dry season and rainy season have both had an impact on LULC change.

3.8 Comparison of present and future simulated LULC on streamflow over Upper Gilgel Abay watershed

Figure 11 shows the area coverage of the land cover classes indicates the cultivated land increased every decade. The area coverage of the cultivated land was 46.81%, 52.45%, and 53.77% in 2007, 2018, and 2027 land cover maps, respectively, which is the same scenario estimated over Tana Basin, Upper Blue Nile River Basin (Getachew et al. 2021). And also, the other land cover classes show change every decade. The forest land increased from 23.78% to 25.75% by the 2007 and 2018 land cover maps and decrease to 24.78% by the 2027 land cover map. These LULC variations directly influenced the streamflow of the study area watershed including an increase of streamflow, especially on the LULC 2027 estimated map, and it reaches the maximum monthly average streamflow of about 1400m³/s year in 2009 (Fig. 13). Such changes happen due to the cultivated land and forest land; similar research by Tewabe and Fentahun (2020) on the Lake Tana sub-basin shows an increase in streamflow.

Fig. 13

Scatter plot of monthly river streamflow (A) for the calibration period (1997–2006) and (B) for the validation period (2007–2012)

Each decade the streamflow increases with an increment of LULC change. According to Birhanu et al. (2019), land use and land cover have a significant impact on the rainfall–runoff process. Based on this instance, the streamflow in the research area increased from 1995 to 2027. In 1995, 2007, 2018, and 2027, the mean monthly calibrated and validated simulated streamflow results were 62.20, 66.51, 68.66, and 72.10 m³/s, respectively (Fig. 13). Additionally, the research area's overall actual mean monthly streamflow was 53.09 m³/s. Overall 12.29% of maximum change occurred in the grassland areas (Table 20) during the year 2018–2027 LULC with the same change occurring in the same basin by Abuhay et al. (2023) and Gumindoga et al. (2014). This shows how streamflow is changed due to changes in land use–land cover of the water shade (Figs. 14 and 15).

Fig. 14

LULC change area coverage for 2007 (A), LULC change area coverage for 2018 (B), and predicted LULC change area coverage for 2027 (C) in the upper gilgel Abay watershed

Fig. 15

Simulated streamflow from (1997–2012) of Upper Gilgel Abay watershed (1995, 2007, 2018, and 2027) LULC map result

Conclusions

The three changes in land use and cover were created using Landsat pictures taken in 1995, 2007, and 2018, using Landsat MSS-5, Landsat 7, and Landsat 8 level 1 photographs, respectively. The classification was done using ERDAS Imagine 2014. The Upper Gilgel Abay watershed's land use has altered dramatically between 1995 and 2018, according to assessments of land use cover. From 39.16%, 46.81%, and 52.45% in 1995, 2007, and 2018, respectively, cultivation underwent significant change. The growth of rural population and agricultural land affects the loss of forested land. The amount of forestland has declined by 13.97% from 1995 to 2007 and by 12% from 1995 to 2018; however, it has increased by 1.97% between 2007 and 2018. The main issue for farmers in the study area is a shortage of arable land, which may be the result of population pressure that has led to a high demand for new land. The simulated peak streamflow in the wet season in 1995 was 128.57 m³/s on the land use and land cover change map; in 2007 it was 131.69; and in 2018 it was 163.33 m³/s. This explains how the study area's 24 years of land use cover change have affected streamflow in the Upper Gilgel Abay watershed's streamflow generation. Additionally, the dry LULC map was 8.21 m³/s in 1995, 6.54 m³/s in 2007, and 4.78 m³/s in 2018.

The results of the classification show a considerable shift in the types of land cover, particularly an increase in cultivated land and a decrease in shrubland. As a result, these approaches successfully met the study's goal and provided an answer to the research problem. The future expected land use and land cover for 2027 between 10-year intervals from the previous 20 years has also had a significant impact on the streamflow of the Upper Gilgel Abay watershed, which brings us to the conclusion that the future LULC change on the streamflow in the Upper Gilgel Abay watershed will occur. Because the streamflow in the study region during the dry season was 3.57 m³/sec and the simulated wet season discharge was 166.83 m³/sec, this indicates that there will be an increase in variation during the wet season of 38.57 m³/sec and a decrease in variation during the dry season of 4.64 m³/s for the period of (1995–2027). The study gives insight contribution on how the variation of LULC impacts the streamflow and it is very essential for water planners and natural resource professionals which demonstrates their plan on the Upper Gilgel Abay basin (Tables 18, 19, and 20).

Table 19 Seasonal mean streamflow of the validation period of (1997–2012)

Table 20 Overall transition of actual and predicted LULC change over Upper Gilgel Abay watershed in (%) from (2007–2027)