Run-off modelling of pervious and impervious areas using couple SWAT and a novel machine learning model in cross-rivers state Nigeria

The urban area as an exceptional ecological system acts as integrated region that unified the residents and their nearby environment (Mahdy & Rizk 2023). Urbanisation significantly transforms the original environment, conveying intense changes during the conversion from natural to anthropic landscapes (Roccati et al. 2018). Along with urban expansion and construction, the initial land cover, such as grasslands and forests, is increasing substitutes by larger extents of impervious surfaces (Pouyat & Pavao-Zuckerman 2017). The alterations of the land cover resulted in a substantial decrement in urban rainwater infiltration capacity and interception (Shao et al. 2018; Aliasghar et al. 2022). Furthermore, changes in local climate extremely affect storm drainage within urban and nearby areas, prompting urban areas to become vulnerable to short-duration extreme rainfall and high intensity (Douglas et al. 2018). These dynamics causes augmented risks from rainfall as well as runoff, and storm-related flood hazards within urban regions. The impacts could be mitigated through the improvement of awareness, early warning, prediction, and mapping. Flood-prone areas prediction employing land use maps, precipitation data and digital elevation model (DEM), and flood hazard maps are critical for better city planning (Darabi et al. 2018; Konin et al. 2022; Nejatian et al. 2023).

Generally, in Nigeria at the local and regional scale, the majority of the study on runoff was conducted using conventional approaches. Even though the conventional approaches have allowed investigation of areas' rainfall–runoff in numerous urban statements, these approaches do have limitations (Jehanzaib et al. 2022). First, they are frequently imprecise because of the little part of the runoff areas that are sampled. Second, the limitation is that suitable data are hard to gather because of the cost and data gathering times related to surface water runoff volumes. The third limitation of the conventional approaches is the logistic implication. The sample size limitations, timeliness, expenditure, and access can conceivably be addressed at a sequence of scales by utilising the geospatial-modelling approach (runoff modelling) in the study area.

Previous studies have utilised various types of models for assessing runoff within urban areas. These models comprise Illinois Urban Drainage Area Simulator (Survey et al. 1974), Hydrological Simulation Program-FORTRAN (Al-Abed & Al-Sharif 2008), Technical Release 55 (TR-55) (Scs 1986), Storm Water Management Model (Peterson & Wicks 2006), Hydrologic Engineering Center-River Analysis System (Zhang & Johnson 2017), and Urban Flood Cell Model (MODCEL) (Miguez & de Magalhaes 2010). Furthermore, multicriteria decision analysis tools for geographical information system (GIS) based, like the analytic hierarchy process (Malczewski 2006), have been widely used. However, because of the particularity and complexity of urban planning tasks, such as crisscrossed underground drainage networks, complex urban terrain, and imposed restrictions by urban management, it is difficult to obtain enough data for hydrological monitoring (Jacobson 2011). This escalated the difficulty in conducting rainfall–runoff research, while predicting flood-prone areas are equally facing many challenges (Kourgialas & Karatzas 2011). Hence, techniques relating to hydrological models as well as 3S technology (comprising GIS, remote sensing (RS), and global positioning system, were adopted for investigating urban rainfall–runoff and related issues (Saadi et al. 2022). All these techniques help to comprehend spatial scales, and patterns of the urban landscape, which influences the processes of rainfall–runoff in the area.

The selection of a hydrological model capable of simulating the impacts of climate change on a region's water resources system, considering the diverse conditions of agricultural management and water resources, will greatly influence the assessment of these effects and enhance decision-making processes for organizations and stakeholders (Ualiyeva et al. 2022). The findings of the conducted research indicate that the utilization of the soil and water assessment tool (SWAT) model, which has the capability to concurrently simulate the interdependent influences of hydrological variables and agricultural management practices (e.g., runoff, evaporation, transpiration, and groundwater) as well as plant-related variables (e.g., yield), within basins characterized by complex land use conditions and diverse soil types, represent a viable and appropriate choice.

This approach represents an alternate methodology for flow forecasting that relies on data-driven techniques, namely, on machine learning algorithms. These models have the capability to acquire knowledge of nonlinear associations among variables and discern the connection between input and output elements of a process without necessitating a comprehensive understanding of its physical attributes (Chen et al. 2023). The application of machine learning in diverse hydrological forecasting scenarios has undergone extensive evaluation and has been extensively documented in recent years. Predicting streamflow in ungauged watersheds, when gauged data are lacking, poses a formidable challenge irrespective of the model employed. The application of machine learning in ungauged watersheds is not feasible due to the requirement of observed flow data for training the model. Watershed models, such as the SWAT, have the potential to be utilized in ungauged watersheds. However, the lack of observational data poses uncertainty regarding their effectiveness and suitability in such contexts. Alternatively, in cases when there are watersheds in close proximity to the research area that have existing streamflow data, it is possible to employ regionalization procedures to utilize watershed models or empirical methodologies.

Essenfelder & Giupponi (2020) presented a methodological framework aimed at simulating the flow contributions of interbasin water transfers in situations where observational data are lacking. This framework involves the integration of machine learning techniques with hydrologic modelling. The methodology presented in this study utilizes a hydrologic model, specifically the SWAT, to replicate the rainfall–runoff process within a watershed. In addition, a machine learning algorithm is employed to simulate the decision-making process involved in interbasin water transfer. To enhance the accuracy of daily flow forecast in unmonitored watersheds, a hybrid model was devised by integrating the SWAT-ANN couple model (Noori & Kalin 2016). The initial step was the simulation of daily streamflow using the SWAT. Subsequently, the simulated baseflow and stormflow from SWAT were utilized as inputs for the artificial neural network (ANN) model. Senent-Aparicio et al. (2020) integrated the SWAT model with the chloride mass balance approach to enhance the simulation of streamflow in basins with high-permeability bedrock that receive interbasin groundwater flow. The proposed methodological framework encompasses the original conceptualization of the system and the development of a novel formulation. This framework offers a reliable approach for addressing comparable basins where the baseflow component is predominantly influenced by the interbasin groundwater flow.

Recent hydrological studies have been conducted in various regions of the world (Marshall et al. 2004, 2018; Vaze et al. 2010; Guo et al. 2018). The majority of these studies focused on rainfall–runoff modelling at the level of the community or a small catchment, as opposed to the larger area. In Nigeria's Cross River State, pervious and impervious areas were modelled for rainfall–runoff in this study, complementing the work of others. In this study, run-up modelling with the SWAT model and GIS/RS systems were used to extract details of the land cover and analyse the rainfall–runoff using logistic regression and other descriptive statistical methods. This study employed the emotional artificial neural network (EANN) model as a novel application of machine learning to improve the obtained SWAT. Due to the EANN model's capabilities in simulating the rainfall–runoff phenomenon, the SWAT-EANN coupled model has been employed to evaluate urban flooding. This is essential for quantitatively identifying regions with higher and lower runoff. This identification will enhance urban planning and development in an effort to prevent the recurrence of floods and other related catastrophes.

In runoff modelling, black-box and white-box models are two approaches, each with their own advantages and disadvantages. Black-box models, also known as empirical or data-driven models, are based on the statistical relationship between input and output data and do not require in-depth knowledge of the physical processes involved. Black-box models have the benefit of being simple and easy to implement. In cases where the mechanisms governing the runoff process are not well understood or when data are limited, they can provide reasonably accurate results. In addition, black-box models are renowned for their adaptability, as they can be applied to various regions and runoff scenarios. However, black-box models lack physical interpretation because they do not explicitly consider the mechanics of the underlying process. This can be considered a disadvantage when insight into the contributing factors and mechanisms influencing runoff is required. In addition, black-box models may have difficulty generalizing well beyond the data range used for calibration (Tajfar et al. 2023).

White-box models, also referred to as physically based models, explicitly represent the underlying process mechanics and incorporate knowledge of the catchment's physical properties. This allows for a deeper understanding of the factors that influence runoff and the evaluation of hypothetical scenarios. When examining the effects of land use changes or climate variability on runoff, white-box models can be especially beneficial. In addition, white-box models have the ability to handle extreme or rare events that may not be adequately represented by empirical relationships captured by black-box models. However, white-box models require more specific catchment information, such as soil properties, rainfall characteristics, and land use patterns, which is not always readily available. In addition, the complexity and computational requirements of white-box models can be a limiting factor, particularly when simplified approximations are required due to limited data or resources (Bhuvaneswari et al. 2022).

By using the SWAT-EANN couple model, a more precise runoff modelling of pervious and impervious areas prediction was attempted in this study.

There were six main data utilised in this research, which include land use mapped, DEM, precipitation data, temperature, wind and humidity data, and land cover of the study area obtained from Google Earth. Furthermore, the software and hardware used were also described. More information about the data and their sources were discussed as follows:

• (a)

  Land use map: The map was generated by using Landsat 5 thematic mapper, Landsat 7 enhanced thematic mapper plus (ETM+), and Landsat 8 Operational Land Imager (OLI), for 1998, 2008, and 2018, respectively. This process was successfully realised by using the maximum likelihood classification technique in ArcGIS 10.4 software to classify pervious and impervious land cover areas. All the Landsat images have 30 m spatial resolution, identified as a medium spatial resolution satellite image. Similarly, the images were obtained from the United State Geological Survey (USGS).

• (b)

  DEM: DEM data were used among the input of the SWAT model like a land use map. The DEM was obtained from the National Aeronautics and Space Administration (NASA) website due to the issue of compatibility. The DEM has 14 spectral bands derived from the visible onto the wavelength region of thermal infrared. It has a high spatial resolution of 50–300 feet and is obtained from ASTER images. ASTER is among the instruments of five Earth-observing launched on ‘December. 18, 1999, on NASA's Terra satellite’. The DEM data used in this study were downloaded in February 2019.

• (c)

  Temperature, wind, and humidity data were obtained from the same source.

Software that were used for the study include Envi 5.3, Microsoft Excel, and ArcGIS 10.4 for data handling that involves database creation, runoff modelling (SWAT model an extension of ArcGIS software), and analysis, PYTHON 3.9 for couple SWAT-EANN model.

This is to map and examine the land cover changes according to pervious and impervious locations to analyse the difference in the study area. Thus, some processes were observed to achieve the set objective, which includes pre-processing, processing of the satellite images, mapping, and estimation of the impervious and pervious landscape of the study area to enable the execution of runoff modelling (Figure 2). This runoff modelling is necessary to allow urban planning and expansion that can prevent the recurrence of floods and related disasters.

The pre-processing part includes geometric correction, atmospheric correction, and transformation of the digital number to radiance. All the information is important in particular atmospheric correction reduces the distortions on the satellite images as a result of the effects of cloud cover and shadow within the atmosphere. Equally, the geometric correction was done to make the image represent reality, similar to the geo-referencing technique. Likewise, the conversion of satellite image digital numbers to radiance was realised, as this is important before further processing can be implemented. The stated procedures were achieved using Envi 5.3 image processing software (Chen et al. 2019).

When an image is finally converted from a digital number to radiance, it eases the processing procedure and permits analysis. Hence, the processing of the image in this stage was done by applying supervised classification techniques. Maximum likelihood classification was used to distinguish between impervious and pervious areas of the study site. More so, water bodies layer were included in conducting the classification in order to have excellent land use maps for runoff modelling. The produced land use maps are among the essential input of the SWAT model.

This study achieved the development of a runoff model based on an impervious surface analysis via the SWAT model. Therefore, this realisation was possible as SWAT required several sets of spatial as well as temporal parameters (input data). SWAT a semi-distributed model has to process combine and analyse the data spatially using GIS tools. Hence, to enable the usage of the SWAT model, it was attached within GIS software as a free added extension ArcSWAT (2012) for ArcGIS (Kangsabanik & Murmu 2017).

This study used reliable online sources and Nigerian government agencies such as the Nigerian Meteorological Agency (NIMET). While the online sources include The National Aeronautics and Space Administration (NASA), DIVA-GIS (https://www.diva-gis.org) and USGS among others (Čerkasova et al., 2018). The SWAT needs monthly precipitation values, maximum as well as minimum temperature, relative humidity, solar radiation, and wind speed for modelling several physical processes. Precipitation (rainfall data) and temperature data were downloaded from the NIMET. Hence, 30 years’ data of monthly rainfall (250 mm) and likewise 30 years of 20-degree daily maximum, as well as minimum temperature data (1998–2018) were also employed.

To identify and categorize regions containing pervious and impervious surfaces, it is necessary to employ a land cover classification technique such as maximum likelihood classification or supervised classification. During the process of land cover classification, a predetermined set of categories was established to differentiate between impervious surfaces and pervious surface areas. To perform the land cover classification with precision, it was necessary to carry out field sampling for the purpose of establishing training sites and reference points. A selection of land cover categories was determined to carry out the analysis of impervious surfaces. The categorization scheme was initially developed with the purpose of ultimately condensing the various classifications into two primary classes: impervious surfaces and pervious surfaces. Bare soil was incorporated into the field sampling and classification procedure to maximize the representation of this class within the image.

After completing the classification of land cover, a decision was taken to adopt a different approach for categorizing areas covered by water. This involved refraining from collecting field samples to determine the exact locations of water. The process of collecting field samples was carried out via a rigorous random sampling technique. The attainment of restricted time and convenient access to destinations within the urbanized region may become unachievable. The samples were collected in a deliberate manner with the intention of ensuring their representativeness in relation to the class they were intended to describe. While the sampling format employed was not random, it is important to note that the samples taken were spatially representative within the designated study area. The availability of samples within the urban area imposed limitations on the collection process. Samples will be collected in areas that are open to the public, excluding private gardens and similar spaces. Every sample exhibited homogeneity and had a diameter of 3 m, which corresponds to 12 pixels. To enhance the robustness of the samples, a criterion was established mandating that the distance between each sample's location and any other class must be at least 5 m. The sampling procedure was conducted utilizing a portable GPS receiver, employing the World Geodetic Surveys (WGS84) as the designated coordinate system.

In the training method of the EANN model shown in Figure 3, the numerical values of III, II, I, and IV, respectively, indicate the neural elements of input weights, net function, activation function, and output unit, and it is similar to the classic ANN value. VI, V, VII, and VIII, respectively, indicate the weight of the net output, the Hh hormone gland, and the input or output hormone units, respectively, while XI, X, IX, and XII indicate the net hormone unit, the hormone activation function, the net function, and static weight as inputs (Sharghi et al. 2018).

There are three significant results demonstrated in this section, which answered the set objectives of this study. These comprise (1) mapping and examining the land cover changes based on pervious and impervious areas to analyse the difference and used them as an input in the SWAT model. (2) Development of a runoff model based on an impervious surface analysis using the SWAT-EANN model; and (3) analyses of the runoff model using descriptive analyses. These revelations provided by this research could serve as a guideline for predicting urban runoff, which can help in urban management to prevent the occurrence of further flood events.

The spatial extents of the three land covers (impervious, pervious, and water body) are presented in Figure 4, whereas Figure 5 demonstrates their changes by proportion (%) every 10 years (1998, 2008, and 2018). The impervious area (urban) was increasing from 39% in 1998 to 86% in 2018, and this was anticipated to occur due to urban expansion and developments related to urbanisation. While on the contrary, the pervious areas (forest) have been decreasing from 79% in 1998 to 38% in 2018. All these aforementioned decreases were due to human developments such as agricultural (oil palm plantation) and industrial activities.

Figure 6 depicts changes in pervious (light green), impervious (red), and water body (blue) land cover features across three epochs (1988, 2008, and 2018) of land cover maps. In 1998, Figure 6(a) reveals the locations dominated by pervious land cover, along with a few impervious locations in the northeastern region of the state. In contrast, in 2008 (Figure 6(b)), impervious areas expanded, displacing pervious and waterbody areas. Moreover, in 2018 (Figure 6(c)), there were significant changes as the impervious land cover continued to increase, as the histogram below demonstrates. The zoomed-out area of the changes within the study site (i, ii, and iii) is presented to justify the alterations, especially in 2018.

The water body was included in the results to have an excellent and precise presentation of the land use/land cover changes in the study area. The marine areas or water bodies were also decreasing from 3% in 1998 to 1% in 2018. Therefore, the decreasing and increasing patterns occurred generally because of climate change impacts. The outcome of these results was used as inputs for runoff modelling using SWAT-EANN. Similarly, the results can be for urban planning and development.

As demonstrated in Figure 9, the runoff in 2012 revealed the highest (178 mm) (more detailed in the problem statement), followed by 160 mm in 2007. While 152 mm was recorded in 2014 and 134 mm in 2003 and 2010, 123 mm of runoff was the highest among others. The lowest runoff was revealed in 2000 (47 mm), 2008 (63 mm), and 2017 (65 mm) among others. Figure 9 shows that 2012 and 2008 have the highest average proportion of runoff, and also shows that 2007 has the highest total annual runoff. The reason for this fluctuation of the runoffs was the impacts of climate change that influence the quantity of the precipitation within the study area – likewise, soil type (clay and mud) is one of the attributing factors of this fluctuation.

The developed model was established using logistic regression (linear), which shows an excellent agreement between the in situ data and the SWAT-EANN model predicted. R² of 0.98, 0.99, 0.97, 0.99, 0.98, and 0.98 are shown in Figure 10. The RMSE also proved a very good accuracy with errors <1 mm of rainfall for the entire year (1998–2018). Table 2 presents the RMSE for the whole year of rainfall. The results revealed precise accuracy that can be enhanced in future research.

The results in Table 2 proved a very significant revelation as the entire observation shows less than 1 mm error as mentioned earlier. However, the differences in the results of the accuracy in every single year are due to the variation of the runoff proportion, which was influenced by some climatic factors as shown in Figure 10.

The initial results presented in this section provide an answer to the question of what the purpose of this study was. The land use map was one of the fundamental inputs that was required to accomplish goals 2 and 3. The findings will also serve as a starting point for urban planning, particularly with regard to the control of contractions and floods. Urban planning is an area that will benefit from the findings. The relationship between urbanization and forestation will be better understood by those who are responsible for urban planning. Researchers in the future will be able to build on the foundational work that has already been done to further investigate urban dynamics in relation to ‘pervious’ and ‘impervious’ areas.

In the area that is the subject of this investigation, the rainfall–runoff modelling that was carried out in this study is the first of its kind to ever be used. This is as a result of the fact that the majority of the previously published studies failed to implement the SWAT-EANN model and instead concentrated on a relatively limited catchment area. This is the reason for this finding. Nevertheless, this particular investigation covered the entire state. As a result, the results of this runoff model can be utilized to broaden the scope of further research. It can also be used for climate change mitigation strategies, which is goal 13 of the sustainable development goals. This is related to urban statements and other environmental issues such as floods and drought.

The modelling of rainfall and runoff that was carried out for this study is considered to be the first of its kind to be applied in the study area. The reason for this is that many of the published studies only focused on a small catchment, and they did not use the SWAT-EANN model. In contrast, the scope of this study included the entire state. As a consequence of this, the findings of this runoff model can be utilized to broaden the scope of the subsequent research. In addition, it can be utilized for the development of strategies to mitigate the effects of climate change (goal 13 of the sustainable development goals), which are connected to urban statements and other environmental issues (such as floods and droughts).

The exceptional ecological system of the urban area integrates residents and their surroundings. From natural to anthropic landscapes, urbanisation changes the environment drastically. Along with urban expansion and construction, grasslands and forests are replaced by impervious surfaces. The land cover changes reduced urban rainwater infiltration and interception. Climate change greatly impacts storm drainage in urban and nearby areas, making urban areas vulnerable to short-duration, high-intensity rainfall. These dynamics increase rainfall-, runoff-, and storm-related flood risks in urban areas. Awareness, early warning, prediction, and mapping can reduce impacts. Better city planning requires flood-prone area prediction using land use maps, precipitation data, DEMs, and flood hazard maps.

This study presented and discussed the results, according to the objectives, guided by a structured methodology. The mapping and estimating of the pervious, impervious, water body, and land cover were demonstrated by this study area (for three epochs, 1998, 2008, and 2018). While the runoff was modelled, SWAT-EANN was used to answer the second objective. Finally, accuracy assessment and R² were used to analyse runoff modelling. The achieved results shall contribute to the related industries to improve urban planning, management, and development in Cross River state and elsewhere across the globe dynamics.

The year 2012 had the highest runoff (178 mm), followed by 160 mm in 2007. Runoff was highest in 2014 at 123 mm, followed by 2003 and 2010 at 152 and 134 mm. The lowest runoff was 47 mm in 200, 63 mm in 2008, and 65 mm in 2017. 2012 and 2008 have the highest average runoff proportion, while 2007 has the highest annual runoff. Climate change affects precipitation in the study area, and soil type (clay and mud) contributes to this fluctuation. The in situ and predicted data are in excellent agreement with the developed couple SWAT-EANN model. R² values of 0.98, 0.99, 0.97, 0.99, 0.98, and 0.98 were recorded, respectively. The RMSE also demonstrated a high degree of precision, with errors 1 mm of precipitation for the entire period of 1998–2018. The results revealed a level of precision that can be enhanced by future investigation.

The authors are thankful to Deanship of Scientific Research and under the supervision of the Science and Engineering Research Centre at Najran University for funding this work under the Research Centers Funding program grant code (NU/RCP/SERC/12/5) and also acknowledge funding and study facilities used at D & U Geospatial Councult Limited, Kano, Nigeria. Finally, the authors wish to thank Dr Baiya and Mr Umar Musa Umar for their academic advice.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.