Theoretical and Applied Climatology (2022) 147:87–98

https://doi.org/10.1007/s00704-021-03817-4

ORIGINAL PAPER

Simulation of rainfall‑runoff process using an artificial neural network

(ANN) and field plots data
Vahid Gholami1 · Hossein Sahour2

Received: 14 April 2021 / Accepted: 8 October 2021 / Published online: 17 October 2021
© The Author(s), under exclusive licence to Springer-Verlag GmbH Austria, part of Springer Nature 2021

Abstract
Rainfall-runoff modeling is necessary for many hydrological studies, such as estimating peak discharges and designing
hydraulic structures. The intensity and frequency of extreme climatic events necessitate the use of advanced approaches that
incorporate different climatic and landscape parameters for rainfall-runoff modeling. The majority of small basins around the
world lack hydrometric data. This study applied an artificial neural network (ANN) to simulate the rainfall-runoff process
using data from field sampling plots in conjunction with rainfall and hydrometric data. For this purpose, similarly sized
field plots were established among different land uses to determine the amounts of initial loss and infiltration during rainfall
occurrences at the Talar basin in the north of Iran. The modeling process was carried out using a multi-layer perceptron
network where the network inputs were rainfall time series, initial loss, soil antecedent moisture condition (A.M.C), and the
time to peak of the basin, and the output was runoff time series. The data from rain gauge and hydrometric stations and field
plots were collected for three consecutive months. The threefold exercises of training (R-sqr = 0.96, MSE = 0.005), cross-
validation (R-sqr = 0.95, MSE = 0.006), and test (R-sqr = 0.81, MSE = 0.05) have yielded favorable results. The modeling
results also indicated the significance of the cumulative rainfall data and initial loss in the modeling process. Results show
that runoff time series and flood hydrograph can be simulated using the optimal inputs and an appropriate neural network
structure for the basins without active hydrometric stations.

1 Introduction In the past few decades, the frequency and intensity of

extreme climatic events have increased (Sharafati and Pez-
Runoff is typically measured through hydrometric stations eshki 2020). Climatic changes in conjunction with land
established throughout the basins. Unfortunately, such prac- use changes have exacerbated flood risk (Zhao et al. 2018).
tices are constrained in many parts of the world by several Therefore, it is necessary to investigate the effect of climate
limitations such as the absence of stations in some parts of change on runoff.
basins, temporal gaps in collecting data, and data quality Considering the importance of river engineering and
issues. Moreover, measurements of the maximum instan- flood risk management for water resource management,
taneous discharge and flood hydrograph using hydromet- empirical models and methods have been widely applied
ric stations are time-consuming and expensive and it may to overcome the data availability issues (Fotovatikhah
not even be achievable in many large basins (Gholami et al. et  al. 2018). Numerous models and methods have been
2018). presented to simulate river discharge and river flow hydro-
graph (e.g., HEC-HMS or Hydrologic Modeling System,
Mike, regression methods, artificial intelligence). All those
models require appropriate input data and their accuracy
* Vahid Gholami is highly dependent on the accuracy of the input data.
Gholami.vahid@guilan.ac.ir
There are advantages and disadvantages to applying those
1
Department of Range and Watershed Management methods. Regression methods, for instance, showed lower
and Department of Water Engineering and Environment, accuracies in comparison to other methods such as artifi-
Faculty of Natural Resources, University of Guilan, cial intelligence-based models in several case studies (Ada-
Sowmeh Sara 1144, Guilan, Iran
mowski et al. 2012). In recent decades, the use of artificial
2
Marine Science Institute, The University of Texas at Austin, intelligence has become more common for hydrological
Port Aransas, TX 78373, USA

13
Vol.:(0123456789)
88 V. Gholami, H. Sahour

simulations (Smith and Eli 1995; Dawson 1998; Sajikumar quality and availability, type of data, and users’ skills.
and Thandaveswara 1999; Cheng et al. 2002; Wilby et al. The temporal distribution of rainfalls, soil characteristics,
2003; Jain et al. 2004; Anctil and Rat 2005; Cheng et al. initial loss and infiltration, land use and land cover, soil
2005; Peters et al. 2006; Kisi et al. 2013; Wang et al. 2013; moisture content, and physiographic characteristics of the
Kisi 2015; Taormina and Chau 2015; Gholami et al. 2019; basin are important factors in runoff generation (Evans
Nourani et al. 2019a,b; Sharghi et al. 2018, 2019; Sahour 1995; He et al. 2011; Gholami et al. 2018). A combina-
et al. 2021). The artificial neural network (ANN) is one of tion of these factors determines the hydrologic response
the most commonly used artificial intelligence techniques of a basin to rainfall and ultimately determines the peak
that have been widely used in hydrological studies, espe- discharge and flood volume and shape of a hydrograph.
cially in the simulation of river discharge and hydrograph Unfortunately, many artificial intelligence-based models
and it proved to be a capable tool for rainfall-runoff mod- do not consider the main factors controlling the hydro-
eling (Luk et al. 2001; Dakhlaoui et al. 2012; Chau 2017; logical response of the basins to the rainfall occurrences.
Warwade et al. 2018), simulation of flow river hydrograph Therefore, we need to provide a comprehensive model
(Manson et al. 1996; Tokar and Johnson 1999; Tokar and by considering the effective parameters on the genera-
Markus 2000; Dibik and Solomatine 2001; Wu and Chau tion of runoff. Regarding the availability of input data for
2011; Isik et al. 2013; Alizadeh et al. 2017), simulation a model, the rainfall data is typically available through
of river water quality parameters (May et al. 2008; Wu meteorological stations in many parts of the world. Soil
et al. 2014), and simulations of the quantity and quality of and vegetation data can be generated through field stud-
groundwater (Dixon 2004; Samani et al. 2007; Motevalli ies or can be extracted from remote sensing data (Zhao
et al. 2019). The advantages of ANNs over other rainfall- et al. 2018). Physiographic data can be extracted from
runoff simulation methods are their high precision and time topographic maps. The key issue, however, is determining
efficiency of the method. Moreover, it is more efficient than the amount of infiltration and initial loss for different land
traditional modeling methods such as regression methods. use and land cover types.
Wu and Chau (2011) simulated the runoff using ANN cou- Previous studies have shown that even in similar land use
pled with singular spectrum analysis and the results showed types and soil hydrologic groups, the amounts of initial loss
the high performance of ANN for estimating runoff val- and runoff, in the area under different agricultural practices,
ues. Minns and Hall (1996) used an ANN to simulate flood vary from one place to another (Kirkby et al. 2005; Keim
hydrograph and peak discharge. The results showed the et al. 2006; Farajzadeh and Khaleghi 2020). Therefore, it is
performance of ANN for simulation of peak discharge was necessary to determine the precise value of the initial loss
better than that for simulation of flood hydrograph (due to in different parts of the basins to accurately simulate the
optimization of network based on peak discharge). Accord- rainfall-runoff process. One of the most common methods
ing to Minns and Hall (1996), the selection of the target in determining the initial loss, runoff, and sediment gen-
function is a decisive factor in the error rate of the simu- eration is implementing sampling plots in the fields. Using
lated hydrograph and is depending on the purpose of the field sampling plots with the same size, we can measure the
study. If the purpose is to determine the peak discharge or amount of infiltration, initial loss, runoff generation, and
to design a hydraulic structure, the target function should antecedent soil moisture conditions under different slopes
be the peak discharge. If the goal is to determine the flow and land use and land cover conditions. These measurements
volume, the target function should be the volume of the dis- can eventually be applied as inputs in the modeling process.
charge or the shape of the hydrograph (Varvani et al. 2019). Several studies have been carried out on the measurement
In some cases, optimization of the models is performed of runoff and sediments using field plots (Wischmeier and
with different target data in successive repetitions to achieve Smith 1958, 1978; Renard et al. 1996; Lane and Nearing
maximum matching between the observed and simulated 1989; Zhang et al. 1996; Ascough et al.1997; Nearing et al.
hydrographs (Farajzadeh and Khaleghi 2020). 1999; Joel et al. 2002; Licznar and Nearingb 2003; Las
Using ANN, Kalteh and Hjorth (2009) simulated run- Heras et al. 2010; Gholami et al. 2018). This study aimed
off values in a watershed in northern Iran. Their results to simulate the rainfall-runoff process using an ANN and
indicated the high performance of ANN in the simula- field sampling plots to reconstruct statistical gaps. For this
tion of temporal gaps in runoff time series. However, the purpose, we applied and evaluated a comprehensive method-
number of inputs in their study was limited, and they did ology by simulation of the spatial and temporal distribution
not consider many crucial factors such as soil properties, of rainfall and considering the soil properties, vegetation
vegetation, and infiltration (Cattan et al. 2009; Sun et al. cover, and physiographic parameters of the basin. Further,
2016). There are different methods of artificial intelli- the proposed model can be a methodology to study the effect
gence for modeling, the application and results of which of climate changes and changes and extreme events in flood
depend on the factors such as nature of research, data hazard and risk analysis.

13
Simulation of rainfall‑runoff process using an artificial neural network (ANN) and field plots… 89

2 Study area 3 Methodology

To evaluate the performance of the methodology presented 3.1 Runoff measurement

in this study, field studies were carried out through the
deployment of field plots as well as meteorological and The first step in the modeling process is the precise determi-
hydrometric data of the Talar basin in the north of Iran. nation of the input and output parameters.
The study basin has three rain gauges and one hydromet- In general, the factors affecting runoff generation include
ric station. The effects of land use changes and spatial rainfall intensity, previous soil moisture, vegetation condi-
and temporal changes in rainfall are also significant in tions, soil properties, and topographic characteristics of
the study area. Finally, the validated methodology can be the land. In the present study, the input parameters were
used for the other watersheds with similar settings. The incremental rainfall (hourly), cumulative rainfall, anteced-
field data collection was carried out for 3 months from ent soil moisture condition (A.M.C), time to peak discharge,
September to November 2017. The Talar river basin in the and initial loss and the output data was hourly runoff dis-
north of Iran with an area of about 1825 ­km2 is extending charge or flood hydrograph. Incremental rainfall shows the
from 52° 35′ to 53° 23′ E longitudes and 35° 42′ to 36° temporal distribution of precipitation. Cumulative rainfall
17′ N latitude (Fig. 1). The elevation of the Talar basin can be effective in investigating the cumulative effect of
varies from 270 to 3800 m. Depending on the location of precipitation in reducing soil permeability and runoff gen-
the rain gauge stations, the mean annual precipitation has eration. A.M.C reflex the initial amount of moisture in the
been reported from 400 to 840 mm/year. The main type of soil, which is inversely related to permeability. Initial loss
precipitation in the study area is rainfall. The land use and represents the combination of vegetation conditions and soil
land cover of the Talar basin can be classified into forest, properties (soil texture). Time to peak discharge is a physi-
rangelands, agricultural lands, and urban areas. There are ographic parameter that is dependent on parameters such as
three rain gauge stations within or near the basin (Golaf- the basin shape, the basin slope, and the rainfall intensity.
shan, Sangedeh, and Paland) and one hydrometric station The river discharge before rainfall occurrences was con-
(Shirgah hydrometry station) at the outlet of the Talar sidered the base flow. To estimate the runoff, the base flow
basin. According to the recorded hydrographs (Shirgah was excluded from the total recorded. The hourly runoff
hydrometry station), the average time of concentration for discharge data for three consecutive months was the model
the Talar basin is 33 h. output.

(A) (B)

Fig. 1  Location of the a Talar river basin and b rain gauge stations and the influence area of the rain gauge stations

13
90 V. Gholami, H. Sahour

For accurate measurement of runoff and initial loss in the basin), for all rainfall increments, the coefficients for
each land cover type, the sampling plots with dimensions of the influence area of each station were applied in rain-
1.8 × 22 m were set up in the study area (Fig. 2). For this pur- fall values. Thus, the spatial and temporal distribution of
pose, three similarly sized sampling plots were established rainfall over the basin was simulated in hourly intervals as
in each land use type, except forest with only two sampling inputs for the rainfall-runoff model. Further, both incre-
plots due to space limitations. The plots were established mental (hourly) and cumulative rainfall data were applied
in the places with the same slopes. These plots were estab- to determine the optimal inputs for the ANN model. Using
lished by the minimum disturbance of the soil. The initial cumulative rainfall as input, we could estimate the amount
height of the metal plots was 40 cm, in which 10 cm of them of infiltration and soil moisture.
was placed into the soil. At the outlet of each plot, there was The most important parameters affecting the runoff
a pipe connected to a tank with a capacity of 220 l. After generation are soil moisture content, vegetation type, and
each rainfall event, the amount of runoff inside the tank was soil characteristics (Poesen and Hooke 1997; Renard et al.
measured. Then, the runoff for the plots was estimated by 1996; Rosa et al. 1999; Peakup and Marks 2000; Sahour
dividing the volume of runoff inside the tank over the area et al. 2014, 2016). To estimate these parameters, the initial
of the plots. Moreover, the total loss (S) was estimated by loss was applied in the model. The initial loss and runoff
subtracting the runoff from the rainfall values. Finally, 20% generation are the functions of all these factors. Since almost
of the total loss (0.2S) was determined as the initial loss of all rainfalls over roads and urban areas (with no gardens and
the study area. The initial loss reflects the characteristics of green spaces) turn into the runoff, these areas were estimated
soil or soil hydrologic groups, land use, vegetation cover, in the GIS environment and were considered impenetrable
and soil moisture content (Kirkby et al. 2005; Keim et al. surfaces. However, because the value of this parameter is
2006). Therefore, by applying the estimated initial loss val- not changing over time, it was not used as an optimal input
ues, soil properties, soil cover, and soil moisture content can for the ANN model.
be quantitatively involved in the modeling process. As we mentioned before, measurements were carried out
on the plots after any rainfall events. Since three plots were
3.2 Modeling of runoff and flood hydrograph using set up for each site, the average of the three plots was con-
ANN sidered the final measurements. Antecedent soil moisture
was also estimated based on the sum of precipitations which
In the intervals between rainfalls and during sunny days, occurred during the past 5 days (Gholami et al. 2018). The
the discharge of the river includes only the base flow of physiographic features of the basin have a significant effect
the river; as a result, the runoff discharge equals zero. The on the shape of the hydrograph (He et al. 2011; Khaleghi
studies were carried out during the 3 months of the spring et al. 2011). The most important of them is the area and the
season when snowfall was not reported at high altitudes time of concentration of the basin. Parameters such as the
of the basin. Rainfall was simulated based on the data area of the basin, which is constant over time, cannot be con-
obtained from Golafshan, Sangdeh, and Paland gauge sta- sidered an appropriate input to simulate the rainfall-runoff
tions. The rainfall data were simulated in hourly intervals process since the values of these parameters are fixed. The
based on the area of influence for each rain gauge station time of concentration of the basin is a function of several
(Fig. 3). The area of influence for the stations was deter- factors such as the slope of the streamflow channels, the
mined by the Thiessen method in the GIS environment area, and rainfall intensity. Since the study is carried out
(Gholami et al. 2015; Tehrani et al. 2018). To simulate in one basin, the rainfall intensity determines changes in
the area of influence (non-uniform rainfall distribution in parameters such as time of concentration, lag time, and time

Fig. 2  Field plots in different

land use and land cover estab-
lished in the study area. a Field
plots in the rangelands. b Field
plots in forestlands

(A) (B)

13
Simulation of rainfall‑runoff process using an artificial neural network (ANN) and field plots… 91

Fig. 3  Land use of the Talar

river basin and the influence
area of the rain gauge stations

Table 1  The mean runoff and initial loss values obtained from the data), and test data (25% of the recorded data). Finally,
field plots the network training and testing were performed, and a
No Sub-basin Land use Area ­(km2) Initial The mean of trial–error method was applied to establish the optimal
loss initial loss network structure. In this study, only one of the compo-
(mm) (mm) nents (including the inputs, the number of neurons, train-
1 Paland Rangeland 502.4 11 11.4 ing method, transform function, and learning technique)
Forest 76.32 20 was manipulated each time, and then, the model was evalu-
Dry farming 132.7 8.4 ated by comparing the observed hydrograph with the sim-
2 Golafshan Rangeland - 11 19.8 ulated hydrograph (Kalteh and Hjorth 2009; Kisi et al.
Forest 220 20 2013). Moreover, statistical coefficients such as coefficient
Dry farming 1.97 9.5 of determination (R-sqr), mean square error (MSE), and
3 Sangdeh Rangeland 517.4 11 14.1 minimum and maximum absolute error (MAE) were used
Forest 326.9 20 to evaluate the modeling process. Finally, by using the
Dry farming 48.8 9.5 trial–error method and evaluating the model, the optimal
4 The mean of initial loss in the Talar basin (mm) 15 network structure for simulating the rainfall-runoff pro-
cess, including optimal inputs, optimal transfer function,
optimal training technique, and the number of neurons
to peak of a flood hydrograph. Therefore, for the ease of were selected. After determining the optimal network, the
measurement, the time to peak was extracted based on the performance of the ANN model was evaluated through the
hydrographs recorded at the basin outlet. This parameter was comparison of the observed and simulated hydrographs, as
reconstructed in hourly intervals in terms of time to peak. well as the measurement of errors. Accuracy assessment
In this study, a multi-layer perceptron network (MLP) in the simulation of flood hydrographs was carried out on
was applied. The ANN model inputs included quantities of both peak discharge and runoff volume. The effects of the
rainfall as increments, cumulative rainfall, influence area ANN inputs (optimal inputs selection) on the accuracy of
of the gauge stations, initial loss, soil antecedent moisture the rainfall-runoff simulation were also evaluated through
(A.M.C), and time to peak (Table 1). The model output the trial–error method and by addition and subtraction of
was the hourly runoff discharges. The input data were individual variables from the set of inputs and sensitivity
partitioned into three subsets, training data (65% of the analysis of the input variables. Therefore, using a tested
recorded data), cross-validation data (10% of the recorded network and reliable inputs, we could simulate the gaps

13
92 V. Gholami, H. Sahour

(missing values) for runoff data obtained from hydrometric As we mentioned earlier, a trial–error method was
stations. applied for training the ANN and the result for each step
was evaluated by comparing the simulated hydrograph
with the observed hydrograph. The hyperbolic tangent
4 Results transfer function was eventually used as the optimum
transfer function. This method has been applied in sev-
As we mentioned before, rainfall data from rain gauge sta- eral previous studies. The Levenberg–Marquardt (LM)
tions were collected during the 3-month study period and algorithm was selected as an optimal learning technique
precipitation events were used as continuous increments (Gholami et al. 2015). After evaluating the inputs, the
as inputs for the ANN model. Precipitation values were cumulative rainfall (three stations), the mean initial loss,
ranged between 0 and 52.9 mm and the maximum rainfall and time to peak were determined as optimal inputs and
intensity was recorded 8.8 mm in an hour at the Paland the incremental rainfall was removed from the sets of input
station. The weighted coefficients based on the Thiessen variables. The results of the sensitivity analysis showed
method for the Golafshan, Sangdeh, and Paland stations, that the network performance was strongly dependent on
were 0.04, 0.38, and 0.58 respectively. During the 3-month the optimal inputs. Also, the correlation analysis showed
study period (Sep 2017 to Nov 2017), there were thir- that there was no significant correlation between incre-
teen rainfall events with recorded runoff in the study area. mental rainfall and runoff (Table 3).
There were seven other rainfall events, in which the rain- The accuracy assessment showed high precision for the
fall amounts were less than the initial losses; as a result, no training step (Fig. 4, R-sqr = 0.96). The influence areas of
runoff was generated in the basin. The maximum recorded rain gauge stations were analyzed using a trial–error method
peak discharge in the outlet of the basin was 30.4 ­m 3/s in the training stage. The results showed that the influence
including 7.53 ­m3/s base flow and 22.87 ­m3/s runoff flow. area of the rain gauge station is not significant in the mod-
The land use and the influence area for each station eling process in our case. However, this parameter has an
are presented in Fig. 3. The mean initial loss for agricul- important role in reconstructing the local distribution of
tural lands, rangelands, and forest with the antecedent soil rainfall. Table 4 shows the results of the accuracy assess-
moisture condition of II were 9.5, 11, and 20 mm respec- ment for the training and cross-validation steps. Finally, a
tively (Table 1). Finally, the estimated weighted average multi-layer perceptron network with the hyperbolic tangent
initial loss of all land use types was 15 mm. These values transfer function, the LM learning technique, 1000 epochs,
were subtracted from cumulative rainfall from each hourly and one neuron was selected as the optimal neural network
increment and were applied as input for the ANN model structure. Further, cumulative rainfall, initial loss, and time
(Table 2). to peak were selected as the optimal inputs for the ANN
Due to mismanagement and inappropriate agricultural model. The testing step was carried out after training the
practices in the study area such as plowing in the direc- network and the results are presented in Table 5.
tion of slope, the precise determination of the initial loss As we mentioned before, testing data included 25% of
using an empirical formula is challenging. For this rea- the total observed data that were randomly selected. For the
son, significant runoffs were recorded during the rainfall testing step, the simulated hydrographs were compared to
events in which the amount of that rainfall was far less the observed hydrographs (Fig. 5). In addition to the hydro-
than the initial loss estimated by empirical formulas such graphs, the simulated hourly runoff data were also compared
as the curve number (CN) method. For example, the initial to the recorded runoff in Fig. 6. The results for the test stage
losses estimated by empirical formulas for dry farming and are presented in Table 5 which shows the performance of the
rangelands are 12.7 and 17 respectively (for soil moisture ANN (R-sqr = 0.81) for rainfall-runoff modeling.
condition of II). However, the observed values using filed
plots data for rangelands and dry farming were 8 and 11
respectively. 5 Discussion
Among the physiographic parameters, the time to peak
was considered as one of the model inputs. Finally, the Rainfall is the most important input in a rainfall-runoff
time to peak, according to Table 2, was estimated for each model and both intensity and amount of rainfall should be
rainfall event by measuring the starting time of the rainfall considered in the modeling process. However, the intensity
and the peaking time for the runoff. The time to peak of of the rainfall is not uniform in all over a large basin. There-
the Talar basin varies with the amount and intensity of fore, the non-uniformity of the rainfall should be considered
rainfalls and peak discharges. In the study basin, the time quantitatively in the rainfall-runoff model. In the present
to peak was ranging from 23 to 35 h. study, for the better performance of the model, the spatial
and temporal distribution of rainfall in hourly increments

13
Simulation of rainfall‑runoff process using an artificial neural network (ANN) and field plots… 93

Table 2  Network inputs and outputs for simulating rainfall-runoff process for one of the rainfall events
Time Discharge P sum P sum P sum IL IL IL Golaf- Tp (hr)
Flood ­(Q3/s) Paland (mm) Sangdeh (mm) Golafshan (mm) Paland (mm) Sangdeh (mm) shan (mm)

1 0 0 0 0 15 15 15 33
2 0.55 0 0 0.1 15 15 14.9 32
3 0.83 0 0 0.5 15 15 14.4 31
4 1.1 0 0 4.4 15 15 10 30
5 1.38 0 0 7.9 15 15 2.1 29
6 1.65 0 0 9 15 15 0 28
7 1.93 0 0 10 15 15 0 27
8 1.93 1.5 0.8 11.5 13.5 14.2 0 26
9 3.87 5.2 11.4 15.1 8.3 2.8 0 25
10 5.37 8.8 16.3 17.2 0 0 0 24
11 6.77 13.4 17.8 18.1 0 0 0 23
12 8.87 18.2 18.6 19.8 0 0 0 22
13 10.87 23.7 21.3 21.9 0 0 0 21
14 14.97 31.7 23.4 24.6 0 0 0 20
15 19.17 36.8 24.5 26.9 0 0 0 19
16 22.87 36.9 27.3 28.7 0 0 0 18
17 22.87 36.9 30.2 31.3 0 0 0 17
18 21.57 36.9 31.2 32.1 0 0 0 16
19 20.37 36.9 31.8 32.9 0 0 0 15
20 19.07 36.9 32.4 34.6 0 0 0 14
21 17.77 36.9 34.8 35 0 0 0 13
22 16.47 36.9 34.8 35 0 0 0 12
23 15.27 36.9 34.8 35 0 0 0 11
24 13.97 36.9 34.8 35 0 0 0 10
25 12.67 36.9 34.8 35 0 0 0 9
26 11.47 36.9 34.8 35 0 0 0 8
27 10.17 36.9 34.8 35 0 0 0 7
28 8.97 36.9 34.8 35 0 0 0 6
29 7.67 36.9 34.8 35 0 0 0 5
30 6.47 36.9 34.8 35 0 0 0 4
31 5.17 36.9 34.8 35 0 0 0 3
32 5.17 36.9 34.8 35 0 0 0 2
33 4.07 36.9 34.8 35 0 0 0 1
34 3.97 36.9 34.8 35 0 0 0 0
35 3.87 36.9 34.8 35 0 0 0  − 1
36 3.47 36.9 34.8 35 0 0 0  − 2
37 3.07 36.9 34.8 35 0 0 0  − 3
38 2.87 36.9 34.8 35 0 0 0  − 4
39 2.67 36.9 34.8 35 0 0 0  − 5
40 2.67 36.9 34.8 35 0 0 0  − 6
41 2.67 36.9 34.8 35 0 0 0  − 7
42 2.67 36.9 34.8 35 0 0 0  − 8
43 2.57 36.9 34.8 35 0 0 0  − 9
44 2.57 36.9 34.8 35 0 0 0  − 10
45 2.57 36.9 34.8 35 0 0 0  − 11
46 2.47 36.9 34.8 35 0 0 0  − 12
47 2.46 36.9 34.8 35 0 0 0  − 13
48 2.43 36.9 34.8 35 0 0 0  − 14
49 0 0 0 0 0 15 15  − 15

13
94 V. Gholami, H. Sahour

Table 3  The correlation coefficient between the input variables and Table 4  Evaluation of the ANN model at the training and cross-vali-
runoff values dation stages for modeling the rainfall-runoff process
Parameter Pearson correlation with Runoff Criteria Training: runoff esti- Cross-validation
mation
Accumulative rainfall (Sangdeh) 0.47*
Accumulative rainfall (Paland) 0.48* MSE 0.005 0.006
Accumulative rainfall (Golafshan) 0.54* Min Abs error 0.07 0.08
Rainfall (Sangdeh) 0.16 Max Abs error 6.5 9.4
Rainfall (Paland) 0.17 R-sqr 0.96 0.94
Rainfall (Golafshan) 0.19
Initial loss (tree stations)  − 0.52*
Time to peak  − 0.48*
estimation of model inputs especially for days or basins
Soil antecedent moisture 0.1
with missing hydrometric data. Results show that exclud-
Rainfall × station’s area of influence The same as rainfall*
ing the initial loss as a model input causes a significant
*
 Significant error in the runoff estimation. Time to peak is also another
important parameter related to lag time and time of con-
centration of the basin; therefore, the average time to peak
was used as input. The rainfall values are not the only vari- of observed hydrographs is considered a network input.
able affecting the runoff generation. Infiltration and distribu- According to the results, excluding this parameter can
tion of the rainfall within a time scale are other important negatively affect the efficiency of the ANN model by caus-
parameters. Therefore, cumulative rainfall in hourly intervals ing an error in the estimation of runoff and the recession
was selected as the most suitable input for the rainfall-runoff curve. In the modeling process, the maximum and mini-
model, based on the correlation test and trial–error method mum values of the inputs should be used in the training
in ANN. and testing steps. Therefore, sunny days should be used in
The initial loss varies from one land use type to another, the input and output data. If we did not use sunny days (the
even if they are located in the area with the same slopes. On days with zero rainfall values), the tested network would
the other hand, simulating the vegetation cover and anteced- overestimate the runoff values for the days without rainfall
ent soil moisture condition on an hourly basis is challenging. or snow. We performed and compared two cases: modeling
As a result, applying the initial loss as one of the model
Table 5  Evaluation of the ANN Criteria Test: runoff
inputs which reflects both the soil properties and vegetation model at the testing stage for the estimation
condition was the best option. The initial loss values were rainfall-runoff modeling process
subtracted from each rainfall and were used as one of the MSE 0.05
model inputs. Min Abs error 0.09
Using the field sampling plots, we were able to accu- Max Abs error 10.3
rately determine the runoff height and volume, total loss, R-sqr 0.81
and finally initial loss. What is challenging in the pro-
cess of rainfall-runoff modeling is selection and accurate

Fig. 4  Comparison between the simulated and observed (measured) hydrographs at the Talar basin outlet to evaluate the performance of the
ANN model in the training stage (Rsqr = 0.96)

13
Simulation of rainfall‑runoff process using an artificial neural network (ANN) and field plots… 95

Fig. 5  Comparison between the simulated and observed (measured) hydrographs at the Talar basin outlet to evaluate the performance of the
ANN model in the test stage (Rsqr = 0.8)

without sunny days and modeling with sunny days. The over-training, and finally, the ANN model was evaluated
results showed that if the zero values (sunny days) are through the comparison of the simulated and observed
not applied to the modeling process, it will cause notable hydrographs. The correlation coefficients between network
errors in the simulation process. On the other hand, a con- inputs and trail-error method in the ANN and sensitivity
tinuous time series of rainfall-runoff data were applied in analysis of the inputs showed that the cumulative rainfall
the modeling process. Therefore, zero runoff values were and initial loss are the most important input parameters in
considered for sunny days for better performance of the the rainfall-runoff modeling process. The initial loss reflects
model and to generate a continuous modeled time series the role of vegetation, soil characteristics, and soil mois-
data. ture content. Therefore, increasing rainfall intensity due to
In the present study, the accuracy of the network at the climate change and land use manipulations through defor-
training stage was very high and the error rates are accept- estation is important in aggravating flood risk. According
able. Cross-validation was also performed to prevent to the results, the presented methodology can be used to

Fig. 6  Comparison between the simulated discharges and observed (measured) discharge at the Talar basin outlet to evaluate the performance of
the ANN model in the test stage on a series of randomized data

13
96 V. Gholami, H. Sahour

simulate flood hydrograph and runoff discharge, and finally modeling and can simulate maximum peak discharge val-
to reconstruct the statistical gaps in hydrometric stations for ues with high performance. One of the major drawbacks
ungauged basins. However, the accuracy and reliability of of ANN-based models is their limitation in disclosing the
the model inputs are very important, and it will be easier to relationships between input and output due to the black-box
use our methodology in the small basins with the slight spa- nature of models. Additionally, the presented methodology
tial distribution of rainfall and land use variation. One of the requires establishing sampling plots for precise measure-
advantages of using this method over other rainfall-runoff ments of inputs that require labor works. The data acquisi-
models is the time efficiency of the operations and good per- tion for input parameters in the present study was accurately
formance with the limited length of the time series inputs. determined by field measurements and they had statistically
One of the most important issues in presenting rainfall- significant relationships with the output. Therefore, for the
runoff model or other hydrological simulation models is the future studies in rainfall-runoff modeling, it is suggested to
ability to accurately simulate extremes values. Based on the implement and evaluate the inputs of the present method in
results of the presented neural network model can simulate other artificial intelligence-based models.
the maximum and minimum values with high accuracy. This
implies the high capability of the model in predicting floods Acknowledgements  We would like to thank the Regional Water Com-
pany of Mazandaran for providing the rainfall and hydrometric data.
or low flows. Finally, the validated model is an effective
tool for simulating the rainfall-runoff process, reconstructing Author contribution  All authors contributed to the study conception
hydrographs or unrecorded river discharges, predicting flood and design. Material preparation and data collection and analysis were
risk, and studying different scenarios of climate and land use performed by V.Gholami and H.Sahour. The first draft of the manu-
changes on flood risk. script was written by V. Gholami and all authors commented on pre-
vious versions of the manuscript. All authors read and approved the
final manuscript.

6 Conclusion Availability of data and materials  All data and materials as well as
software application or custom code support our published claims and
Simulation of the rainfall-runoff process is very important comply with field standards.
in hydrological studies and river engineering. One of the
most important applications of the rainfall-runoff models is Declarations 
to estimate the peak discharge and flood volume to design
Ethical approval  Not applicable.
hydraulic structures, flood management plans, or river engi-
neering. We can also fill the gaps in hydrological time series Consent to participate  All authors consent to participate in this manu-
data by forecasting the values for the missing data and the script.
area without hydrometric stations. An artificial neural net-
Consent for publication  All authors consent to publish this manuscript.
work is an efficient tool in hydrological modeling. Several
technical studies have been carried out in the field of flow
Conflict of interest  The authors declare no competing interests.
hydrograph modeling or rainfall-runoff models using vari-
ous artificial intelligence techniques. However, the most
important issue in the modeling process is to determine
relevant inputs for the model and their accurate estimation
to properly reflect the process of modeling. In the present
References
study, cumulative rainfall inputs in hourly intervals, initial Adamowski J, Chan HF, Prasher SO, Sharda VN (2012) Comparison
loss, as a quantitative indicator reflecting the land cover and of multivariate adaptive regression splines with coupled wave-
vegetation conditions, and time to peak, were considered let transform artificial neural networks for runoff forecasting in
for the modeling process. The cumulative rainfall and ini- Himalayan micro-watersheds with limited data. J Hydroinform
14(3):731–744
tial loss had the highest impact in the simulation process Alizadeh MJ, Kavianpour MR, Kisi O, Nourani V (2017) A new
respectively. The effect of rainfall and initial loss was mostly approach for simulating and forecasting the rainfall-runoff process
on the runoff peak discharge and the shape of the hydro- within the next two months. J Hydrol 548:588–597
graph (runoff volume) respectively. In the modeling process, Anctil F, Rat A (2005) Evaluation of neural networks streamflow
forecasting on 47 watersheds. J Hydrol Eng ASCE 10(1):85–88.
this is also important to determine the spatial and temporal https://​doi.​org/​10.​1061/​(ASCE)​1084-​0699(2005)​10:​1(85)
variation of rainfall and the initial loss values. Therefore, Ascough JC, Baffaut C, Nearing MA, Liu BY (1997) The WEPP basin
selecting the appropriate inputs for the ANN model and model: I. Hydrology and erosion. Trans ASAE 40(4):921–933
an optimal network structure enables high precision simu- Cattan P, Ruy SM, Cabidoche YM, Findeling A, Desbois P, Char-
lier JB (2009) Effect on runoff of rainfall redistribution by the
lation of the flood hydrograph and runoff discharge. The
presented methodology performed well in rainfall-runoff

13
Simulation of rainfall‑runoff process using an artificial neural network (ANN) and field plots… 97

impluvium-shaped canopy of banana cultivated on an Andosol Kirkby MJ, Bracken LJ, Shannon J (2005) The influence of rainfall
with a high infiltration rate. J Hydrol 368(1):251–261 distribution and morphological factors on runoff delivery from
Chau K (2017) Use of meta-heuristic techniques in rainfall-runoff mod- dryland catchments in SE Spain. CATENA 62:136–159. https://​
elling. Water 9(3):186. https://​doi.​org/​10.​3390/​w9030​186 doi.​org/​10.​1016/j.​catena.​2005.​05.​002
Cheng CT, Ou CP, Chau KW (2002) Combining a fuzzy optimal model Kisi O (2015) Discussion of improved particle swarm optimization–
with a genetic algorithm to solve multi-objective rainfall–runoff based artificial neural network for rainfall-runoff modeling by
model calibration. J Hydrol 268(1–4):72–86 Mohsen Asadnia, Lloyd HC Chua, XS Qin, and Amin Talei. J
Cheng CT, Wu XY, Chau KW (2005) Multiple criteria rainfall–runoff Hydrol Eng 20(9):07015009
model calibration using a parallel genetic algorithm in a cluster of Kisi O, Shiri J, Tombul M (2013) Modeling rainfall-runoff process
computers / Calage multi-critères en modélisation pluie–débit par using soft computing techniques. Comput Geosci 51:108–117
un algorithme génétique parallèle mis en œuvre par une grappe Lane LJ, Nearing MA (1989) USDA-water erosion prediction pro-
d’ordinateurs. Hydrologl Sci J 50(6):1–1087 ject-hillslope profile version. NSERL report No. 2. US depart-
Dakhlaoui H, Bargaoui Z, Bárdossy A (2012) Toward a more effi- ment of agriculture, Agriculture Research Service, W. Lafayette,
cient calibration schema for HBV rainfall–runoff model. J Hydrol IN
444:161–179 Las Heras MM, Nicolau JM, Martín LM, Bradford P, Wilcox BW
Dawson CW (1998) An artificial neural network approaches to rainfall (2010) Plot-scale effects on runoff and erosion along a slope deg-
runoff modeling. J Hydrol Sci 43(1):47–66 radation gradient. Water Resour Res 46(W04503):1–12. https://​
Dibik YB, Solomatine DP (2001) River flow forecasting using artificial doi.​org/​10.​1029/​2009W​R0078​75
neural network. J Phys Chem Earth 26:1–7 Licznar P, Nearingb MA (2003) Artificial neural networks of soil
Dixon B (2004) Prediction of groundwater vulnerability using an erosion and runoff prediction at the plot scale. CATENA
integrated GIS-based neuro-fuzy techniques. J Spa Hydrol 51(2003):89–114
14(12):1–38 Luk K, Ball JE, Sharma A (2001) An application of artificial neural net-
Evans R (1995) Some methods of directly assessing water erosion of works for rainfall forecasting. Math Comput Model 33:683–693
cultivated land: a comparison of measurements made on plots and Manson JC, Price RK, Ternme A (1996) A neural network model of rain-
in fields. Prog Phys Geogr 19:115–129. https://​doi.​org/​10.​1177/​ fall-runoff using radial basis functions. J Hydraulic Res 34:537–548
03091​33395​01900​106 May RJ, Dandy GC, Maier HR, Nixon JB (2008) Application of par-
Farajzadeh S, Khaleghi MR (2020) Evaluation of the efficiency of the tial mutual information variable selection to ANN forecasting of
rainfall simulator to achieve a regional model of erosion (case water quality in water distribution systems. Environ Model Softw
study: Toroq watershed in the east north of Iran). Acta Geophys 23(10e11):1289e1299
68(5):1477–1488. https://​doi.​org/​10.​1007/​s11600-​020-​00487-0 Minns AW, Hall MJ (1996) Artificial neural networks as rainfall runoff
Fotovatikhah F, Herrera M, Shamshirband SH, Chau SKW, Ardabili F, Models. Hydrol Sci J 43(1):399–417
Piran MJ (2018) Survey of computational intelligence as basis to Motevalli A, Pourghasemi HR, Hashemi H, Gholami V (2019) Assess-
big flood management: challenges, research directions and future ing the vulnerability of groundwater to salinization using GIS-
work. Eng Appl Comput Fluid Mech 12(1):411–437. https://​doi.​ based data-mining techniques in a coastal aquifer, Spatial mod-
org/​10.​1080/​19942​060.​2018.​14488​96 eling in GIS and R for earth and environmental sciences, 547–571
Gholami V, Darvari Z, Mohseni Saravi M (2015) Artificial neural net- Nearing MA, Govers G, Norton DL (1999) Variability in soil erosion
work technique for rainfall temporal distribution simulation (case data from replicated plots. Soil Sci Soc Am J 63(6):1829–1835
study: Kechik region). Caspian J Environ Sci 13(1):53–60 Nourani V, Davanlou Tajbakhsh A, Molajou A, Gokcekus H (2019a)
Gholami V, Booij MJ, Tehrani EN, Hadian MA (2018) Spatial soil Hybrid wavelet-M5 model tree for rainfall-runoff modeling. J
erosion estimation using an artificial neural network (ANN) and Hydrol Eng 24(5):04019012
field plot data. CATENA 163:210–218 Nourani V, Tajbakhsh AD, Molajou A (2019b) b) Data mining based
Gholami V, Torkaman J, Dalir P (2019) Simulation of precipitation on wavelet and decision tree for rainfall-runoff simulation. Hydrol
time series using tree-rings, earlywood vessel features, and arti- Res 50(1):75–84
ficial neural network. Theor Appl Climatol 137(3):1939–1948 Peakup G, Marks A (2000) Identifying large-scale erosion and depo-
He Y, Bárdossy A, Zehe E (2011) The role of catchment classifica- sition processes fromairbone gamma radiometrics and digital
tion in rainfall-runoff modeling. Hydrol Earth Syst Sc Discuss elevation models in a weathered landscape. Earth Surf Proc Land
8(3):6113–6153 25:535–557
Isik S, Kalin L, Schoonover J, Srivastava P, Lockaby BG (2013) Mod- Peters R, Schmitz G, Cullmann J (2006) Flood routing modelling with
eling effects of changing land use/cover on daily stream flow: an artificial neural networks. Adv Geosci 9:131–136
artificial neural network and curve number based hybrid approach. Poesen JWA, Hooke JM (1997) Erosion, flooding and channel manage-
J Hydrol 485:103–112. https://​doi.​org/​10.​1016/j.​jhydr​ol.​2012.​08.​ ment in the Mediterranean environments of southern Europe. Prog
032 Phys Geogr 21:157–199
Jain A, Sudheer KP, Srinivasulu S (2004) Identification of physical pro- Renard KG, Foster GR, Weesies GA, McCool DK, Yoder DC (1996)
cesses inherent in artificial neural network rainfall–runoff models. Predicting soil erosion by water: a guide to conservation planning
Hydrol Process 118(3):571–581 with the revised universal soil loss equation (RUSLE). Soil and
Joel A, Messing I, Seguel O, Casanova M (2002) Measurement of Water Conservation Society, Tucson, p 383
surface water runoff from plots of two different sizes. Hydrol Pro- Rosa D, de la Mayol F, Lozano S (1999) An expert system/neural
cesses 16:1467–1478. https://​doi.​org/​10.​1002/​hyp.​356 network model (impelERO) for evaluating agricultural soil ero-
Kalteh AM, Hjorth P (2009) Imputation of missing values in a precipi- sion in Andalucia region, southern Spain. Agri Ecosyst Environ
tation-runoff process database. Hydrol Res 40(4):420–432 73(3):211–226
Keim RF, Skaugset AE, Weile M (2006) Storage of water on vegetation Sahour H, Mokhtari A, Tehrani EN (2014) Effects of land use/land
under simulated rainfall of varying intensity. Adv Water Resour cover changes on surface runoff (a case study in Siahroud Water-
29(7):974–986 shed, Iran). Elixir Remote Sensing 74:26867–26870
Khaleghi MR, Gholami V, Ghodusi J, Hosseini H (2011) Efficiency Sahour H, Mokhtari A, Ghahfarokh SS (2016) Rainfall-runoff mod-
of the geomorphologic instantaneous unit hydrograph method in eling using remotely sensed data and the hydrologic modeling
flood hydrograph simulation. CATENA 87(2):163–171 system (HEC-HMS). Ecol Envir Conserv 22(4):1735–1745

13
98 V. Gholami, H. Sahour

Sahour H, Gholami V, Vazifedan M, Saeedi S (2021) Machine learning Varvani J, Gholami KMR, V, (2019) Investigation of the relationship
applications for water-induced soil erosion modeling and map- between sediment graph and hydrograph of flood events (case
ping. Soil Tillage Res 211:105032. https://​doi.​org/​10.​1016/j.​still.​ study: Gharachay river tributaries, Arak, Iran). Water Resour
2021.​105032 46(6):883–893. https://​doi.​org/​10.​1134/​S0097​80781​90602​04
Sajikumar N, Thandaveswara BS (1999) A nonlinear rainfall–runoff Wang W, Xu D, Chau K, Chen S (2013) Improved annual rainfall-
model using an artificial neural network. J Hydrol 216:32–55 runoff forecasting using PSO–SVM model based on EEMD. J
Samani N, Gohari-Moghadam M, Safavi AA (2007) A simple neu- Hydroinform 15(4):1377–1390
ral network model for the determination of aquifer parameters. Warwade P, Tiwari S, Ranjan S, Chandniha SK, Adamowski J (2018)
J Hydrol 340:1–11. https://​doi.​org/​10.​1007/​s10584-​005-​5922-3 Spatio-temporal variation of rainfall over Bihar State, India. J
Sharafati A, Pezeshki E (2020) Strategy to assess the uncertainty Water Land Dev 36(1):183–197
of a climate change impact on extreme hydrological events in Wilby RL, Abrahart RJ, Dawson CW (2003) Detection of conceptual
the semi-arid Dehbar catchment in Iran. Theor Appl Climatol. model rainfall–runoff processes inside an artificial neural network.
139(1–2):389–402. https://​doi.​org/​10.​1007/​s00704-​019-​02979-6 J Hydrol Sci 48(2):163–181
Sharghi E, Nourani V, Najafi H, Molajou A (2018) Emotional ANN Wischmeier WH, Smith DD (1978) Predicting rainfall erosion losses:
(EANN) and wavelet-ANN (WANN) approaches for Markovian a guide to conservation planning. Agric. Handbook No. 282. US
and seasonal based modeling of rainfall-runoff process. Water Department of Agriculture, Washington, DC
Resour Manag 32(10):3441–3456 Wischmeier WH, Smith DD (1958) Rainfall energy and its relationship
Sharghi E, Nourani V, Molajou A, Najafi H (2019) Conjunction of to soil loss. Trans-Am Geophys Union 39(2):285–291
emotional ANN (EANN) and wavelet transform for rainfall-runoff Wu CL, Chau KW (2011) Rainfall–runoff modeling using artificial
modeling. J Hydroinform 21(1):136–152 neural network coupled with singular spectrum analysis. J Hydrol
Smith J, Eli RN (1995) Neural network models of the rainfall–runoff 399(3–4):394–409
process, J. Water Resour. Plan Manag ASCE 121:499–508 Wu W, Dandy GC, Maier HR (2014) Protocol for developing ANN
Sun J, Yu X, Li H, Yu Ch, Wang H, Tu Z, Liang H (2016) Simulated models and its application to the assessment of the quality of
erosion using soils from vegetated slopes in the Jiufeng Moun- the ANN model development process in drinking water quality
tains, China. Catena 136:128–134 modeling. Environ Model Softw 54(2014):108–127
Taormina R, Chau KW (2015) Data-driven input variable selec- Zhang XC, Nearing MA, Risse LM, McGregor KC (1996) Evaluation
tion for rainfall–runoff modeling using binary-coded particle of WEPP runoff and soil loss predictions using natural runoff plot
swarm optimization and Extreme Learning Machines. J Hydrol data. Trans ASAE 39(3):855–863
529:1617–1632 Zhao CP, Huang H, Li Z, Chen M (2018) Drought monitoring of South-
Tehrani EN, Sahour H, Booij MJ (2018) Trend analysis of hydro- western China using insufficient GRACE data for the long-term
climatic variables in the north of Iran. Theoret Appl Climatol mean reference frame under global change. Climate 31(17):6897–
136:85–97 6911. https://​doi.​org/​10.​1175/​JCLI-D-​17-​0869.1
Tokar AS, Johnson PA (1999) Rainfall-runoff modeling using artificial
neural networks. J Hydrol Eng ASCE 4(3):232–239 Publisher’s note Springer Nature remains neutral with regard to
Tokar AS, Markus M (2000) Precipitation runoff modeling using arti- jurisdictional claims in published maps and institutional affiliations.
ficial neural network and conceptual models. J Hydrol Eng ASCE
5(2):156–161

13