Cities around the world are growing, attracting millions of people due to the opportunities they provide to improve livelihoods. In 2018, around 55% of the world’s population lived in cities, this is expected to rise to more than two-thirds by 2050 (UN 2021). This is abruptly changing the land use of catchments hosting cities (Ahmed et al. 2020) and altering rainfall-runoff relationships (Guzha et al. 2018; Birhanu et al. 2019; Bulti and Abebe. 2020) which in turn causes an environmental impact that is hindering sustainable development (Wagner et al. 2013; Degife et al. 2019). However, land use land cover (LULC) changes and their hydrological impacts are still among the 23 unsolved problems in hydrology as identified by hydrologists (Saddique et al. 2020).

Published systematic reviews exist with an effort to draw conclusions on the relationship between changes in specific LULC classes and streamflow regimes (e.g., Zhao et al. 2016; Kayitesi et al. 2022). However, these efforts were not fully successful partly because most of the existing studies focused on small catchments and annual streamflow. Kayitesi et al. (2022) concluded that despite the fastest LULC changes in tropical regions, only a few studies evaluated the corresponding hydrological impacts. Results of these few studies on LULC change impact on streamflow are not consistent at a large scale (Reintjes et al. 2011). Hence, there is still a lack of understanding about the impacts of LULC change at scales relevant to water resources planning and management, i.e., for medium and large catchments. As a result, based on a systematic review of existing literature Zhao et al. (2016) advocated for more studies on the impact of LULC change on the entire flow regime changes in large basins.

The major issues in assessing the impact of LULC change are: (i) accuracy of LULC maps, and (ii) separation of hydrological impacts of LULC change from impacts due to climate change. Despite the availability of global land cover maps (e.g., Chen et al. 2015; Peter et al. 2020; Potapov et al. 2022) their use for local studies is constrained by scale, aggregation, lack of validation and incomplete documentation describing the details of the data collected (Verburg et al. 2011). As a result, hydrological studies often involve the preparation of a bespoke LULC map for their study catchment. However, the accuracy of locally prepared LULC maps is constrained by the type of classifier, number of ground control points (GCPs), its spatial distribution, images’ characteristics, catchment’s characteristics and other factors (Shetty et al. 2021).

The limitations of classifiers have a significant impact on the accuracy of LULC classification (Shetty 2019b; Qu et al. 2021) and hydrological behaviors that motivate the ongoing research on improving the performances of classifiers. The non-parametric classifiers (e.g., machine learning) have some advantages over the parametric traditional classifiers (e.g., maximum likelihood) for LULC mapping. These advantages include improved performance using noisy reference or training data, and the ability to deal with complex land cover. Adoption of machine learning classifiers for LULC classification is increasing due to their higher accuracy and performances (Kelsey et al. 2018; Tassi and Vizzar 2020). However, most studies still use traditional parametric classifiers (e.g., Andualem et al. 2018; Engida et al. 2021; Leta et al. 2021) and the comparative performance assessment to inform the selection of classifiers among the many types of machine learning and parametric classifiers is lacking in the literature. In this study, instead of arbitrarily selecting single LULC classifiers (e.g.,Astuti et al. 2019; Koneti et al. 2018; Phan et al. 2021), we based our selection on a comparison of multiple classifiers. We also assessed the LULC change effect variation on streamflow using different LULC classifier product. This is in contrast to past LULC mapping in the study area that either partially covered the catchment by focusing only on Addis Ababa city (Arsiso et al. 2018) or used the traditional maximum likelihood classifier that is arbitrary selected for use (Worako 2016).

In LULC mapping, processing of big data, access to cloud-free satellite imageries, searching, downloading, and mosaicking the images are other factors that consume researchers’ time (Gorelick et al. 2017). However, the recently available Google Earth Engine (GEE) cloud-based computing platform provides a useful resource to overcome these issues. The platform uses Java or Python scripts for image classification without the need to download the data to the local computing resources (Nyland et al. 2018). Furthermore, it contains several packages of machine learning classifiers, useful for LULC classification (Zurqani et al. 2018; Chung et al. 2021; Shafizadeh Moghadam et al. 2021). In this study, GEE was used for two purposes: (i) to access Landsat images, and (ii) to classify LULC of the Akaki catchment, Ethiopia, using machine learning classifiers.

This study aims to quantify the effects of historical (1990–2020) LULC change on streamflow in the Akaki catchment. The Akaki catchment is a very complex catchment due to ongoing rapid development in Addis Ababa city and its surrounding towns such as Sendafa, Legedadi, Dire, and Burayu, with the major portions of the catchment having a large impervious surface. The catchment also includes three artificial water supply reservoirs: Gefersa, Legedadi, and Dire. The presence of those artificial reservoirs adds complexity to understanding the catchment through rainfall-runoff modeling. However, in the previous study (e.g., Shibeshi et al. 2019; Zeberie. 2019) the reservoir effect was not accounted in their modeling.

In this study, a semi-distributed HEC-HMS model was used for LULC change impact assessment. HEC-HMS model was tested and calibrated for a variety of purposes, including flood forecasting (Bhuiyan et al. 2017), rainfall-runoff simulation (Gumindoga et al. 2016; Bitew et al. 2019), assessing the effects of land-use change on hydrological responses (Guzha et al. 2018; Shanshan et al. 2020; Dipak and Shirish. 2021) and climate change impact assessment (Meenu et al. 2012). This study adds to the gap in the literature on the impact of LULC change on the streamflow of a complex catchment with multiple reservoirs and rapid urbanization. Therefore, the approach and findings of this study will inform similar research in urban–rural catchments that are experiencing rapid change.

Material and methodology

Study area

The Akaki catchment is found in central Ethiopia, and it is one of the headwaters of the Awash River. The catchment hosts the capital city of Ethiopia, Addis Ababa, and surrounding towns such as Legedadi, Sendafa, Dire and Burayu. The Akaki catchment is located between 8°36ʹ–9°12ʹ N and 38°40ʹ–39°4ʹ E with an area coverage of around 1500 km2. The elevation of the study area ranges from 2000 m.a.s.l around the bridge of Addis Ababa to Debre Zeit Road to 3400 m.a.s.l on Entoto Mountain (Fig. 1).

Fig. 1
figure 1

Location map a Ethiopian river basins, b Awash basin, and c Akaki catchment. where—RF is the rainfall gauging station, hydrological station is flow gauging station, and the DEM legend unit is a meter

The drainage system of the Akaki catchment is divided into the Little Akaki and Big Akaki river systems. The Big Akaki river sub-catchment covers about 62% of the Akaki catchment area, whereas the Little Akaki sub-catchment covers the remaining part of the catchment area. Big Akaki River has two major tributaries, the Kebena River that originates from the Entoto Mountains in the north of Addis Ababa, and Bulbula River that originates in the northeast of Addis Ababa. This river system contains two water supply reservoirs (Legedadi and Dire). The other river system Little Akaki River originates from the Entoto Mountains in the north of Addis Ababa and Wechecha Mountain in the northwest of Addis Ababa. Gefersa reservoir, the third water supply reservoir, is located in the upstream part of this river. Both Big and Little Akaki Rivers drain to the Aba-Samuel reservoir.

The main rainy season of the Akaki catchment is from June to September. It receives minor rains from mid-February to April. Between 1990 and 2018 the mean monthly maximum and minimum areal rainfall was observed in August (290 mm) and December (7 mm), respectively, with around 1265 mm mean annual rainfall amount over the catchment. The mean monthly maximum and minimum temperature of the study area varies from 7 to 11 °C and 21–25 °C, respectively, over the period of 1990 and 2018. During the same period, the lowest temperatures were recorded in July and August, while the highest temperatures were recorded in February and March.

Data sets

Spatial and temporal data

The main datasets used in this study included Digital Elevation Model (DEM), satellite images, ground control points (GCPs), soil map, and hydro-meteorological data. The DEM was collected from ALOS PALSAR product (, which has 12.5 m × 12.5 m spatial resolution. It was used for catchment delineation and extraction of other catchment characteristics such as time of concentration (Tc), basin storage coefficient (R), sub-basin area (A), longest flow path (L), and slope of the watercourse (S).

Time series images were obtained from Landsat thematic mapper (TM), enhanced thematic mapper (ETM+) and operational land imager (OLI). These images were used for image classification in GEE to derive the LULC maps of the study area. The dataset included full scenes for the years 1990, 2000, 2010, and 2020. The selection of the images was based on the criteria that the images had no cloud cover over the study area, and the resolution of the images were similar (30 m × 30 m). All images were acquired in the dry months of November and December to ensure cloud-free images. GCPs for the three historical periods (1990, 2000, and 2010) were collected from topographic maps and SPOT images together with historical LULC information. Topographic maps (1993, 1995, and 1997) with a scale of 1:250,000 and SPOT 6 and SPOT 7 (for the year 2000 and 2010) images with 5 m and 1.5 m spatial resolutions respectively, were collected from the Ethiopian Geospatial Information Institute (EGII). The daily rainfall data of 12 stations and daily temperature data (maximum and minimum) for 5 stations were collected from the Ethiopian Meteorological Institute (EMI) of Ethiopia. The rainfall and temperature data cover the period 1990 to 2004 to match the observation period of streamflow data. Daily streamflow data recorded at three stations (from 1990 to 2004), and soil data were collected from the Ministry of Water Energy (MoWE). See Fig. 1 for the locations of stations.

Reservoir data

Elevation–area–volume (E–A–V) curve is very essential to represent the effects of reservoirs in rainfall-runoff modeling. For the Legedadi reservoir, the E–A–V curves were established by a company called BECOM in 1992, and for Dire reservoir, it was developed by TAHAL company, in 1997. However, for the Gefersa reservoir, two E–A–V curves were collected from the 2018 bathymetry report. The first curve was prepared by the Addis Ababa Water and Sewerage Authority (AAWSA) in 1966 and then the revised curve was prepared in 2018 by the Ethiopian construction design and supervision works corporation (ECoDSWC). As a result, in this study, a new E–A–V curve was calculated for the calibration period from the original and revised E–A–V curves by interpolation technique. The daily water production data for the three reservoirs were collected from AAWSA. The dam feature data (bottom outlet elevation, size of bottom outlet, dead storage level, normal pull level, spillway length, dam crest level elevation and dam crest length) were collected from the 2018 bathymetry design reports.

Land use land cover data processing

Field survey

Before field visits, unsupervised image classification was conducted to determine the major LULC classes of the study area for ground truth data collection. This was followed by a detailed field survey which was undertaken in October 2020. The major objective of the field visit was: (i) to revise the major LULC classes of the study area which were obtained using the unsupervised image classification, (ii) to collect GCPs for the identified LULC classes, and (iii) to collect additional information about the historical LULC classes and changes in the study area. During the field visit, 6 major LULC classes and 27 sub-classes were identified. 450 GCPs were collected with the help of a geographic positioning system (GPS) with ± 3 m horizontal error. Following common practice in literature (Rientjes et al. 2011; Yimer et al. 2020), a split-sample validation method was then applied in which 80% of the GCPs were used for training the classification algorithms, and 20% of the GCPs were used for validation.

To collect historical information on the past LULC (i.e., 1990, 2000, and 2010), 25 people were interviewed. This information was combined, (i) with the GCPs for 2020, (ii) GCPs that were identified from the SPOT image and then used for the preparation of historical LULC maps of the study area. For LULC classification in 1990 and 2000, 370 GCPs were identified by using a topographic map, defined relationships of NDVI and LULC class and information collected from elder people. For classifying the image of 2010, 420 GCPs were identified from the SPOT image and the information obtained from the interviewees.

LULC classification

In this study, five machine learning (ML) classifiers that have been used in recent studies (e.g., Zurqani et al. 2020; Akanksha et al. 2021; Shetty et al. 2021) and one traditional parametric classifier were compared to select the best classifier for LULC mapping of Akaki catchment. Those ML classifiers used in this study are: (i) classification and regression tree (CART), (ii) random forest (RF), (iii) support vector machine (SVM), (iv) minimum distance (MD), and (v) Naïve Bayes (NB). From the traditional methods, Maximum likelihood classifier (MLC) was also considered. MLC is the most common parametric classifier found in many image processing packages. Equal number of training data and image band combinations were used for each classifier. For a quantitative comparison of the performance of the classifiers, the classification results were evaluated using a confusion matrix. Such comparison helps to evaluate the agreement and details of disagreements between the classified results and reference data in terms of the overall classification accuracy, kappa coefficient, producer’s accuracy, and user’s accuracy (Rientjes et al. 2011; Yimer et al. 2020; Assefa et al. 2021).

The overall accuracy was calculated by dividing the correctly classified pixels by the total number of ground truth pixels. The user's accuracy informs the user how well the map represents the reality on the ground, and the producer's accuracy measures how well a certain area is classified (Chughtai et al. 2021). The user's accuracy corresponds to an error of commission (inclusion) and producer's accuracy corresponds to an error of omission (exclusion). Individual accuracies of more than 70%, overall accuracy of at least 85%, and the Kappa coefficient of 75% indicate acceptable accuracies (Ramita et al. 2009; Phan et al. 2021).

HEC-HMS model setup

The initial values of the HEC-HMS model parameters are summarized in Table 1. These initial values were estimated by combining information from the allowable range value of parameters (USACE, 2018), basin characteristics, analysis of observed streamflow data, and literature (Haile et al. 2016). Rainfall-runoff generation in the Akaki catchment is expected to show large spatial variation mainly due to: (i) significant spatiotemporal rainfall variations, (ii) heterogeneous LULC characteristics, and (iii) presence of reservoirs. Hence, we divided the catchment into 15 sub-basins (Fig. 2) by considering the location of flow gauging stations and reservoirs, distribution of rain gauges, size of drainage area, major tributaries and the dominant LULC classes. The largest sub-basin drainage area is 205.3 km2 (W4580), and the smallest sub-basin drainage area is 39 km2 (W6940).

Table 1 Initial model parameter values of Akaki catchment for the 15 sub-basins
Fig. 2
figure 2

HEC-HMS basin model representation of the Akaki catchment

HEC-HMS uses separate models to represent each component of the rainfall-runoff process (USACE 2018). In this study, the following seven HEC-HMS methods were used for simulation, calibration, and validation. SCS-CN method was selected to calculate losses due to its applicability to evaluate the impact LULC change since its parameter can be determined from LULC and other readily available data. Clark unit hydrograph was used to transform excess precipitation into a direct runoff. This method was selected since it requires only two input data, it is easy to use and it has been used by many researchers. Muskingum method was applied for flow routing in the channels. The recession method was used for modeling of base flow. This method has been used often in literature to explain the drainage from the natural storage in watersheds. Simple surface method was applied to model losses due to surface storages and depreciations. Simple canopy method was also applied in the study to model loss due to trees and other vegetation.

The storage and routing effect of the reservoirs were represented in HEC-HMS model of Akaki. The outflow structure routing method was selected in this study since this method was designed to model reservoirs with several uncontrolled outlet structures like spillway, intake, low-level bottom outlet structures and evaporations from the surface of reservoirs. Diversion of water from the reservoirs to the city was represented as a loss in the model.

HEC-HMS model sensitivity

Sensitivity analysis is a critical component of rainfall-runoff modeling that helps to identify the influential parameters to guide model calibration. For this study, several model simulations were performed first using the initial values of the parameters, and then sequentially changing the values of one parameter by increments from ± 10% to ± 100% while keeping the values of other parameters fixed. For each simulation, the values of the objective functions (Nash Sutcliffe efficiency (NSE), relative volume error (RVE), and coefficient of determination (R2)) were estimated, and then these estimated values of the objective functions were plotted against the parameter values. A steep line plot of values of a parameter against the objective function shows that the model is highly sensitive to that parameter.

HEC-HMS model calibration and validation

The semi-distributed HEC-HMS model of Akaki was manually calibrated and validated at Mutinicha, Big Akaki and Little Akaki stations. This was done by iteratively changing the model parameters’ values until a reasonable match was achieved between the simulated and observed streamflow. This was repeated at the catchment outlet just upstream of Aba-Samuel reservoir by transferring the streamflow data from Big Akaki and Little Akaki gauging stations by catchment area ratio. In the Akaki catchment, the attention given to streamflow monitoring declined over recent years. As a result, access to streamflow data is limited up to 2004. Hence, the historical streamflow data (1990 to 2004) at the gauging stations were used for model initialization (1990–1992), calibration (1993–1999), and validation (2000–2004). This modeling period covers both high (1999) and low flow (2002) conditions in the catchment.

The performance of the HEC-HMS model was evaluated through visual inspection of various aspects of the simulated and observed hydrographs, and a set of objective functions that measure the goodness of fit between various aspects of the two hydrographs. The objective functions which were used in this study are Nash–Sutcliffe model efficiency (NSE), coefficient of determination (R2), and Relative Volumetric Error (RVE) as shown below.

$$NSE=1-\frac{\sum_{\mathrm{I}=1}^{\mathrm{n}}{{(\mathrm{Q}}_{\mathrm{obs},\mathrm{i}}-{\mathrm{Q}}_{\mathrm{sim},\mathrm{i}})}^{2}}{\sum_{\mathrm{i}=1}^{\mathrm{n}}({{\mathrm{Q}}_{\mathrm{obs},\mathrm{ i}}-{\overline{\mathrm{Q}} }_{\mathrm{obs},\mathrm{ i}})}^{2}}$$
$${R}^{2}=\frac{\sum_{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{Q}}_{\mathrm{obs},\mathrm{i}}-{\overline{\mathrm{Q}} }_{\mathrm{obs},\mathrm{i}})}^{2}-\sum_{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{Q}}_{\mathrm{sim},}-{\overline{\mathrm{Q}} }_{\mathrm{sim}.\mathrm{i}})}^{2}}{\sum_{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{Q}}_{\mathrm{obs},\mathrm{i}}-{\overline{\mathrm{Q}} }_{\mathrm{obs},\mathrm{i}})}^{2}}$$
$$RVE=\frac{\sum_{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{Q}}_{\mathrm{sim},\mathrm{i}}-{\mathrm{Q}}_{\mathrm{obs},\mathrm{i}})}{\sum_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{Q}}_{\mathrm{sim},\mathrm{i}}}\times 100$$

where, \({Q}_{obs,i}\) is the observed streamflow at time step i, \({Q}_{sim,i}\) is the simulated streamflow at the same time step, n is the number of observations, and the overbar indicates the mean of the streamflow over the calibration or validation period.

NSE is used to assess the model's ability to reproduce the pattern of the observed hydrograph. Its lower limit is − ∞ with and upper limit of 1 (that indicates a perfect match). Values of NSE between 0.6 and 1.0 show acceptable levels of performance, whereas NSE = 0 indicates that the mean observed value is a better predictor than the simulated value, indicating poor model performance. The values of RVE range between − ∞ to + ∞ with values of zero being the target performance in terms of capturing the observed streamflow volume. RVE values between − 5 and 5% indicate the model performance is very good, values between −10 and −5%, and 5 and 10% indicate acceptable (good performance) volumetric error and values outside the range of -10 and 10% indicate unacceptable model performance. R2 is sensitive to extremes in the hydrograph but insensitive to additive and proportional differences. Its values range from 0 to 1, with one indicating perfect agreement between the extremes of the observed and simulated discharge.

Hydrological impact of LULC changes

HEC-HMS model allows evaluation of LULC change impact by changing the curve number (CN), percentage of impervious (IP) and initial abstractions (IA) as per the LULC cover of 1990, 2000, 2010, and 2020 as well as other basin parameters. The model parameters such as CN, IP and IA were estimated as a function of land use, soil type, and antecedent moisture conditions. For hydrological impact evaluation, two sets of LULC maps were produced using the traditional MLC and CART classifiers. The simulations were carried out using 7 years of climate data (1993 to 1999) for all land uses scenarios. We did not vary the climate data to isolate the impact of climate change and variability on changes in annual and seasonal streamflow caused by LULC change. Figure 3 shows the conceptual framework of the present study, and hence summarizes the steps that we followed in our research.

Fig. 3
figure 3

Conceptual framework of the study

Result and discussion

Assessment of classifiers accuracy

Table 2 shows the accuracy assessment result of the six classifiers for the Akaki catchment. The highest overall accuracy (95.2%) was observed using CART classifier. The producer accuracy ranges from 73.9% for bare land to 99.5% for the water body using this classifier. This indicates that about 74% of the bare land pixels on the ground are identified as bare land on the produced LULC map while nearly all the water body pixels on the ground are classified as water bodies on the map. The user’s accuracy ranges from 70.94% for bare land to 97.40% for the forest. Thus, 70.94% of the bare land in the classified map occurs on the ground, and 97.94% of the forest on the classified map occurs on the ground. The result shows the presence of almost perfect agreement between the classified map pixels and the reference data as indicated by a kappa coefficient of 0.921. Therefore, the classification accuracies are within acceptable ranges compared to the recommended individual accuracies of more than 70%, the overall accuracy of at least 85%, and the kappa coefficient of 75% (Kelsey et al. 2018; Phan et al. 2021).

Table 2 Accuracy assessment results for CART, RF, MLC, SVM, MD and NB

The second-highest overall accuracy was observed (94.61%) for the RF classifier. For RF, the producer’s accuracy ranges from 74.18% for grassland to 98.73% for the forest. The user's accuracy ranges from 67.94 for bare land to 97.40% for the forest. The result shows the presence of good agreement between the classified map pixels and the reference data as indicated by kappa coefficients of 0.90%. However, the RF classifier resulted in the user’s accuracy below the recommended value of 70%.

In the SVM classifier, the third-highest overall accuracy (92.49%) was registered. The producer accuracy of SVM is relatively low (77.41%) for urban area whereas the user’s accuracy is lowest for waterbody (95.02%). SVM resulted in a good agreement between the classified map pixels and the reference data as indicated by the kappa coefficients of 0.856%. However, typically this classifier classifies the grassland and bare land as agricultural areas.

The fourth-highest overall accuracy (81.5%) was produced by the MLC classifier. It resulted in producer’s and user’s accuracies that were below the acceptable value (< 70%). Therefore, the classifier is not suitable for LULC classification in the study area.

The MD classifier did not provide acceptable accuracies as the individual accuracies are below than 70%, the overall accuracy is less than 85%, and the kappa coefficient is less than 75%.

As a result, based on the accuracy assessment result and visual inspection of the classified map, the CART classifier outperforms all the other classifiers. Therefore, CART was chosen to detect long-term LULC changes between 1990 and 2020 for the catchment. However, we also assessed the impact of LULC change on streamflow using the traditional MLC and CART classifiers.

LULC classification

Figure 4 shows the LULC maps of the study area for the year 1990, 2000, 2010, and 2020 using the CART classifier. The highlands in the north and northeast of Addis Ababa city were dominantly covered by forests with considerable bare land. Most of the central part of the catchment was highly urbanized because the catchment hosts Addis Ababa, Ethiopia's capital and largest city. Agricultural land was the dominant and widespread land cover class in the catchment. The major agricultural practices in this area were rain-fed agriculture. However, irrigated agriculture was also detected along the Akaki Rivers though it was not widespread.

Fig. 4
figure 4

Land use land cover maps of Akaki catchment from (1990–2020) using CART classifier

The LULC detection map of Akaki shows that the LULC has changed dramatically in the past three decades. The agricultural land was the dominant LULC class in the catchment. In 1990, it was representing 76.26% (1109.96 km2) of the total catchment area. However, this class was substantially reduced to 52.98% (771.12 km2) in 2020. Similarly bare land faces a large decline from 7.57% (110.18 km2) in 1990 to 3.23% (47.01 km2) in 2020. However, the urban area experienced the most significant increase over the study period. The area has grown from 8.05% (117.16 km2) in 1990 to 29.21% (425.15 km2) in 2020. Similarly, the forest area increased from 4.65% (67.68 km2) in 1990 to 10.09% (146.85 km2) in 2020. This suggests the presence of tree-planting campaigns in the catchment. The change in the water body and grassland cover was not very significant in the catchment.

Figure 5 shows the LULC maps of Akaki produced using MLC for the year 1990, 2000, 2010, and 2020. The MLC classifier resulted in widespread bare lands and large coverage of grassland at the expense of agricultural land. This is in contrary to our field observation, and a large deviation from the results of the CART classifier. Such deviation of land cover coverage by MLC is consistent across the analysis period.

Fig. 5
figure 5

Land use land cover maps of Akaki catchment from (1990–2020) using MLC

HEC-HMS model calibration and validation

The observed and simulated streamflow hydrographs at Mutinicha, Big Akaki, and Little Akaki gauging stations, as well as just upstream of Aba-Samuel reservoir (catchment outlet), are shown in Figs. 6, 7, 8, and 9 for both the calibration and validation period. The model calibration and validation results show the presence of a good match between the rising limb of the simulated and observed hydrographs. The recession limb and the base flows are also well captured. Some peaks, however, are slightly overestimated while others are slightly underestimated. This could be related to rating curve uncertainties or observation intervals of water levels, but further investigation is required. The model also performed well in reproducing the observed streamflow during the independent (validation) period suggesting the model’s capability outside the calibration period.

Fig. 6
figure 6

Observed and simulated stream flow at Little Akaki gauging station (1993 to 1999 for calibration and 2000 to 2004 for validation)

Fig. 7
figure 7

Observed and simulated stream flow for the calibration and validation period at Mutinicha station (1993 to 1999 for calibration and 2000 to 2004 for validation)

Fig. 8
figure 8

Observed and simulated stream flow for the calibration between 1993 and 1999 and validation period between 2000 and 2004 at Big Akaki station

Fig. 9
figure 9

Observed and simulated stream flow for the calibration and validation period just upstream of Aba-Samuel reservoir (1993 to 1999 for calibration and 2000 to 2004 for validation)

The model performed well in terms of NSE, R2, and RVE objective functions (Table 3). For instance, according to the calibration at Little Akaki station, the observed and simulated streamflow volumes matched very well in terms of relative volume error (RVE = 0). Similarly, the calibration was under good range in terms of capturing the pattern of the observed hydrograph (NSE = 0.71). Overall, the model performance can be rated relatively very good in terms of reproducing the volume and pattern of the streamflow that was observed at the Little Akaki gauging station. The model was also found to perform very well in terms of NSE at Mutinicha, Big Akaki, and Aba-Samuel. However, the performance slightly deteriorated at Mutinicha in terms of RVE. Considering uncertainties related to the flood release of the Legedadi reservoir, the model performance can be considered acceptable at Mutinicha.

Table 3 Performance of HEC-HMS model during the calibration period for the four stations

Table 4 summarizes the performance of the HEC-HMS model for the validation period. According to the validation result the model well performed in the Akaki catchment in terms of NSE, R2, and RVE objective functions. For instance, according to the validation at Little Akaki station, the observed and simulated streamflow volumes matched well in terms of volume error (RVE = − 4.1%). Similarly, the validation was under a very good range in terms of reproducing the pattern of the observed hydrograph (NSE = 0.70). The validation results indicate that the model achieved a relatively very good fit between simulated and observed streamflow at the Little Akaki gauging station. The model performance was found to be very good in terms of NSE at all gauging stations. However, in terms of RVE, model performance at Big Akaki and Mutinicha slightly deteriorated as compared to the performance in the calibration period. Considering data limitations and the highly dynamic nature of the catchment, the model performance indicates that HEC-HMS model can be used for streamflow simulation in the study area.

Table 4 Performance of HEC-HMS model for validation period at the four stations

Hydrological impact of LULC change

The model simulation for LULC change impact assessment was carried out to investigate the magnitude and direction of changes in streamflow. The HEC-HMS model was used to investigate the effects of LULC change in stream flow in the Akaki catchment over four different time periods (1990, 2000, 2010, and 2020). The assessment indicates the streamflow changed under each LULC change scenarios. Table 5 shows the mean annual streamflow responses to the past 30-year LULC changes at three river gauging stations and at the catchment outlet using two different LULC products (CART and MLC). The Mutinicha river catchment is located in the northeast direction of Addis Ababa city. The catchment includes two water supply reservoirs (Dire and Legedadi) and the mean annual streamflow was decreased by 5.57% and 2.56% using CART and MLC during the study period, respectively. This may be related to the fact that the bare land area in the highlands of the catchment was replaced by forest plantation. The Big Akaki and Little Akaki rivers’ catchments cover both the old and new parts of Addis Ababa city. The mean annual streamflow in those river catchments increased over the study period. This was due to very dynamic urban expansion in those catchments. For example, between the year 1990 and 2020 the mean annual streamflow of the Big Akaki river catchment was increased by 11.07% and 39.07% using CART and MLC classifiers, respectively (see Table 5). Similarly, the mean annual streamflow of the Little Akaki catchment was increased by 26.78% and 33.34% over the study period for LULC maps using CART and MLC classifiers, respectively.

Table 5 Summary of LULC change impacts on mean annual streamflow for major rivers outlets in Akaki catchment

Streamflow at the four sub-catchments of Akaki increased in the wet season (June–September). Using LULC maps generated by CART, the increment in streamflow was relatively small in 1900–2000 but becomes large from 2000 onwards (Table 6). Over the analysis period (1990–2000), Mutinicha experienced the smallest change in the wet season streamflow in the Akaki. For the other three catchments, there is a small difference in the magnitude of streamflow change over the analysis period, with Big Akaki experiencing the largest change. Noticeable urbanization is occurring in the Big Akaki catchment. Use of LULC maps generated by the MLC classifier exaggerates the magnitude of streamflow change because of LULC change. The largest percent exaggeration in streamflow increment occurred for the Little Akaki catchment.

Table 6 Summary of LULC change impacts on mean of main wet season (June–September) stream flow for major rivers outlets in Akaki catchment

The Big Akaki catchment saw the largest decline in streamflow during the dry season (Table 7). It was reduced by 36.09% and 28.94% when CART and MLC classifiers were used to generate the LULC product maps, respectively. This is in line with the LULC change result that shows that the urban areas expanded in the eastern and southeastern parts of the Big Akaki catchment. The percentage decline in dry season flow was lowest at Mutinicha. Similarly, streamflow during the dry season decreased in the Little Akaki river catchment. In this study, the effects of LULC change on streamflow were more pronounced in the central part of the study area compared to the entire catchment.

Table 7 Summary of LULC change impacts on mean of main dry season (December–March) stream flow for major rivers outlets in Akaki catchment


The purpose of this study was to assess the effects of LULC changes on streamflow in the Akaki catchment. Because of rapid urbanization, agricultural activity, the presence of artificial reservoirs, and the wide ranges of elevation variation, the catchment is complex. Preparing accurate LULC maps for change detection and hydrological modeling studies in a such catchment is critical. Our study showed the significance of comparing different classifiers for LULC mapping studies is important. However, there is still a scarcity of studies that provide a comparative analysis of LULC classifiers. Still, some studies are advocating for a comparison of the algorithms (e.g., Ghosh and Joshi. 2014; Loukika et al. 2021).

There are some efforts in literature that compare performance of land cover classifiers (Michelson et al 2000; Jamali 2019). However, we were not able to compare the findings of these studies since they did not compare the same set of classifiers. There is a strong need to reach agreement on which set of algorithms to include in any comparison study. Having a standard set of algorithms helps to compare findings across sites and draw a conclusion on the comparative advantages of the available algorithms. In our study, we compared six different classification algorithms and the result shows that CART performs best for LULC mapping of the Akaki catchment. As compared to Jamali (2019) who compared eight algorithms, we obtained a broader variation in the performance level of the five algorithms compared in this study.

The LULC maps of the study area were prepared at 10-year intervals, taking the importance of long-term changes into account. Image classification was performed using TM, ETM + , and OLI Landsat images. According to our findings, rapid LULC changes have occurred over the past 30 years, with the urban area expanding by 21% and the agricultural area shrinking by 24% during the study period. The effects of LULC changes in streamflow simulation were conducted using the conceptual semi-distributed HEC-HMS model. The Akaki catchment’s LULC change simulation result over the last 30 years reveals the presence of changes in streamflow. The rapid LULC change in the study area resulted in an increase in the mean annual streamflow by 13.03%, 39–64%. This increment is much larger than the annual streamflow change reported by the conversion of forest to agricultural land in the rural catchment of Gilgel Abay in Ethiopia (Rientjes et al. 2011). This suggests that the impacts of LULC change on streamflow are larger for urban than agricultural watersheds. The streamflow changes in the Akaki catchment were mainly attributed to the loss of agricultural area and the rapidly increasing urbanized areas. The sub-basin level simulation revealed the largest streamflow increase was located in the central parts of the catchment, i.e., in urbanized area.

Other researchers were also investigated the effects of LULC change on surface runoff in various parts of the world. Sun et al. (2013), for example, investigated the long-term LULC change (1992–2009) on surface runoff in Beijing, Chania. For the LULC map preparation, they used an SVM classifier and Landsat TM/ETM + imagery. Their findings show that surface runoff increased by 30% and 35% for the entire catchment and the urbanized catchment, respectively. The increased urbanization in the area was responsible for the increase in surface runoff. From 1982 to 2003, Li and Wang (2013) studied the effects of urbanization in the Dardenne Creek watershed in Missouuni. The findings of their study revealed that LULC have changed significantly. The study's findings revealed a rapid increase in urban areas in the watershed, from 3.4% in 1982 to 27.3% in 2003. Because of the rapid change in LULC, surface runoff increased by more than 70%. Astuti et al. (2019) investigated the effect of LULC change on surface runoff in increasingly urbanized East Java, Indonesia. The SWAT model was used in their study to assess the effects of LULC on surface runoff. According to the findings of their study, LULC caused an 8% change in runoff. Bulti and Abebe (2020) also investigated the effect of LULC change on surface runoff in the urban–rural catchment of Adama, Ethiopia. Their findings indicate that the urbanization of Adama city increases surface runoff. The findings of the previous studies are similar to our findings that annual streamflow increments were highly correlated with urban expansion.


The effects of LULC changes on streamflow in the Akaki catchment, which hosts the rapidly growing city of Addis Ababa, Ethiopia, were investigated in this study. Images from Landsat TM, ETM + , and OLI sensors were used to create the LULC maps. For producing the LULC maps, five non-parametric machine learning and one parametric classifier were compared. The accuracy assessment result reveals small to large differences in classifier performance. The following conclusions are drawn based on the results of this study:

  • Accuracy assessments from mapping LULC of the Akaki catchment indicated that the performance of CART and RF classifiers was high whereas that of MD and NB were relatively low.

  • The finding showed that the Akaki catchment experienced significant LULC changes during the study period (i.e., 1990–2020). There were six major LULC classes identified: urban, forest, agriculture, waterbody, grassland, and bare land. Among these LULC classes, urban area has increased significantly in the last two decades (2000–2010 and 2010–2020) at the expense of agricultural area. Similarly, forest cover increased in the catchment during the study period, at the expense of bare land. The increase in forest cover was primarily due to increased plantation activity in the catchment during the study period.

  • The conceptual semi-distributed HEC-HMS model simulated discharge was compared to the observed streamflow at four different hydrological stations for both calibration and validation periods between 1993–1999 and 2000–2005, respectively. The model performed well, with NSE and RVE ranging between 0.69–0.77 and 0.00–5.40, respectively. The calibrated HEC-HMS model was used to assess the effects of LULC change on streamflow.

  • Our analysis revealed that the Akaki catchment experienced significant changes in streamflow due to LULC change over the last three decades. The catchment's mean annual and wet season streamflow increased, while the dry season streamflow decreased.

  • However, the LULC maps generated by the MLC classifier did not satisfactorily reflect LULC features on the ground. The classification error propagated to the simulated streamflow, exaggerating the effects of LULC changes on streamflow. However, the CART classifier results were found to be accurate and valid for the study area. As a result, future research should avoid arbitrarily selecting LULC classifiers for impact evaluation studies. Classifier selection should be based on a comparison of multiple classifiers to determine which one performs best.

There are numerous effects of LULC change due to urban expansion on hydrology. These are as follows: (i) watersheds lose their ability to hold and retain water due to increased impervious surface; (ii) decreased infiltration capacity and groundwater potentials of the area; and (iii) increased stormwater and the frequency of extreme hydrological events cause more intense local flooding. As a result, similar research in other rapidly developing cities is critical for understanding LULC changes caused by urbanization, as it aids in quantifying and recognizing such impacts.