APPLICATION OF DIGITAL CELLULAR RADIO FOR MOBILE LOCATION ESTIMATION


IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

NEW APPROACH TO PREDICT FECAL COLIFORM 

REMOVAL FOR STORMWATER BIOFILTER 

APPLICATIONS 

SAI HIN LAI1*, CHUN HOOI BU1, REN JIE CHIN2*, XIANG TING GOH3

AND FANG YENN TEO4 

1
Department of Civil Engineering, Faculty of Engineering, Universiti Malaya, 

50603 Kuala Lumpur, Malaysia. 
2
Department of Civil Engineering, Lee Kong Chian Faculty of Engineering and Science, 

Universiti Tunku Abdul Rahman, 43000 Kajang, Malaysia. 
3
Department of Parasitology, Faculty of Medicine, Universiti Malaya, 

50603 Kuala Lumpur, Malaysia. 
4
Department of Civil Engineering, Faculty of Engineering,  

University of Nottingham (Malaysia Campus), 43500 Semenyih, Malaysia. 

*
Corresponding author: laish@um.edu.my; chinrj@utar.edu.my 

(Received: 21st August 2021; Accepted: 26th March 2022; Published on-line: 4th July 2022) 

ABSTRACT: Fecal coliform removal using stormwater biofilters is an important aspect 

of stormwater management. A model that can provide an accurate prediction of fecal 

coliform removal is essential. Therefore, feedforward backpropagation neural network 

(FBNN) and adaptive neuro-fuzzy inference system (ANFIS) models were developed 

using a range of input features, namely grass type, the thickness of biofilter, and initial 

concentration of E. coli, while the estimated final concentration of E. coli was the output 

variable. The ANFIS model shows a better overall performance than the FBNN model, as 

it has a higher R2-value of 0.9874, lower MAE and RMSE values of 3.854 and 6.004 

respectively, and a smaller average percentage error of 14.2%. Hence, the proposed ANFIS 

model can be served as an advanced alternative to replace the need for laboratory work. 

ABSTRAK: Penyingkiran kolifom tinja menggunakan turas biologi (bioturas) air hujan 

merupakan aspek penting dalam pengurusan air hujan. Model yang dapat menunjukkan 

anggaran tepat tentang penyingkiran kolifom tinja adalah penting. Oleh itu, model 

rangkaian suapan neural perambatan belakang (FBNN) dan sistem adaptasi inferen neuro-

fuzi (ANFIS) telah dibentukkan menggunakan pelbagai ciri input, iaitu jenis rumput, 

ketebalan bioturas dan kepekatan awal E. coli, manakala anggaran kepekatan akhir bagi 

E. coli merupakan hasil pembolehubah. Model ANFIS menunjukkan peningkatan

keseluruhan yang lebih baik berbanding model FBNN, kerana ia mempunyai nilai R2 yang

lebih tinggi iaitu 0.9874, nilai MAE dan RMSE yang lebih rendah iaitu sebanyak 3.854

dan 6.004 masing-masing, dan ralat peratusan purata yang lebih kecil sebanyak 14.2%.

Oleh itu, model ANFIS yang dicadangkan boleh dijadikan alternatif awal bagi

menggantikan keperluan kerja makmal.

KEYWORDS:  artificial intelligence; biofilters; fecal coliform; neural network; 

stormwater 

1. INTRODUCTION

Biofiltration systems such as swale and bio-detention systems are increasingly popular

low-energy treatment technologies for improved stormwater management, e.g. increase of 

45



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

infiltration, reduction of peak flow, improvement of water quality and increase of 

surrounding aesthetic value. Stormwater biofilters can be defined as vegetated vertical 

infiltration systems that can achieve runoff volumes and contaminant load reductions for 

urban environments [1]. They have shown promising yet variable removal of fecal 

microorganisms [2-4]. Fecal coliform bacteria are a group of bacteria that are passed through 

the fecal excrement of humans, livestock, and wildlife and they are the indicator bacteria. 

The most common member of fecal coliform bacteria is Escherichia coli.  

Vegetation and filter media depth may affect the capability of stormwater biofilters in 

removing fecal coliform. In the past two decades, researchers from around the world have 

experimented with biofilters using different design elements to investigate the biofiltration 

system in removing fecal coliform [5-10]. These studies reported that vegetation type or 

filter media depth caused variable bacteria removal performance. A previous study in 

Australia reported that biofilters planted with native grasses (Paspalum 

conjugatum and Buchloe dactyloides) and shrubs (Melaleuca incana, Leptospermum 

continentale) showed improved E. coli removal, possibly due to reduced infiltration rates in 

vegetated biofilter systems. In addition, the leaf or seed extracts of L. continentale 

demonstrated potential antibacterial activity against E. coli [7]. With regards to filter media 

depth, it was reported that E. coli concentration decreased with increasing filter media depth 

[7]. Nevertheless, there is no data available for the influence of native vegetation and filter 

media depth on microbial removal by stormwater biofilters in Malaysia. Therefore, in this 

study, the effect of biofilter designs (i.e., vegetation type, media thicknesses), as well as the 

inflow concentration in fecal coliform removal, are investigated to fill the gap of knowledge. 

Artificial intelligence (AI) appears as a popular tool in providing the solution to 

complex non-linear problems and its application on issues relevant to environmental and 

hydrological researches has been widely seen, i.e. application of artificial neural networks 

(ANN), fuzzy logic and adaptive neuro-fuzzy systems (ANFIS) for the solution of water/ 

wastewater and air pollution-related environmental problems [11], integration of ANN and 

genetic algorithms (GA) for water quality modelling [12], implementation of machine 

learning classification to detect simulated increases of de facto reuse and urban stormwater 

surges in surface water [13], performance prediction of stormwater biofilters in heavy metal 

removal and risk mitigation using multilinear regressions (MLR), neural network (NN), and 

random forest (RF)  [14], etc. Therefore, this study aims to introduce the use of feedforward 

backpropagation neural network (FBNN) and adaptive neuro-fuzzy inference system 

(ANFIS) to predict the final concentration of fecal coliform for different conditions of 

stormwater biofilters. The proposed model can serve as an advanced method to replace the 

need for laboratory work. This study is innovative as it adds value to the current 

development of AI applications in improving stormwater management systems. 

2.  MATERIALS AND METHODS 

2.1  Experimental Works 

In this study, four native plants, namely Cow grass (Axonopus compressus), Pearl grass 

(Axonopus compressus, dwarf), Philippine grass (Zoysia matrella), and Japanese grass 

(Microstegiumvimineum) were selected. The biofilter columns were set up as shown in Fig. 

1.  

46



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

 

Fig. 1: The setup of biofilter columns used in this study. 

The main constituent that forms the filter media was washed river sand. Four different 

depths were fixed in the sand columns, which are 150 mm, 250 mm, 350 mm and 450 mm. 

With respect to each depth, the recorded average hydraulic conductivity was 60.2, 47.1, 35.4 

and 25.4 mm/hr, respectively. The values fell within the ranges recommended by Urban 

Stormwater Management Manual of Malaysia (MSMA) [15].  

One month before conducting the experiments, the biofilter columns were planted with 

native vegetation for the plants to mature. An amount of 80 L of water was collected from 

a local pond to act as stormwater in the experiments.  

Four liters of water were poured into every biofilter column and the filtered water was 

collected. The analysis was carried out for the number of remaining indicator bacteria. A 

vacuum pump was then used to further filter the collected water sample through sterile 

nitrocellulose membrane filters. The membrane filters had characteristics of 0.45 µm pore 

size and 47 mm diameter. The membrane was transferred to a sterile petri dish with an 

absorbent pad (Millipore, Bedford, MA, USA) containing lauryl sulfate membrane medium 

(Oxoid, Hampshire, UK) agar plates. The plate was sealed with parafilm and incubated at 

30 °C for 4 hours to resuscitate the growth of bacteria before further incubation at 44.5 °C 

for 14 hours. Fecal coliform that formed yellow colonies were counted and expressed as 

colony-forming units (CFU) per 100 mL (CFU/100 mL). 

The water samples collected before and after the filtration were termed inflow and 

outflow concentration respectively. The removal efficiency of the biofilter columns can be 

obtained using Eq. (1).  

log 𝑟𝑒𝑚𝑜𝑣𝑎𝑙 =   𝑙𝑜𝑔10
𝐼𝑛𝑓𝑙𝑢𝑒𝑛𝑡 𝑝𝑎𝑡ℎ𝑜𝑔𝑒𝑛 𝑐𝑜𝑛𝑐𝑒𝑛𝑡𝑟𝑎𝑡𝑖𝑜𝑛

𝐸𝑓𝑓𝑙𝑢𝑒𝑛𝑡 𝑝𝑎𝑡ℎ𝑜𝑔𝑒𝑛 𝑐𝑜𝑛𝑐𝑒𝑛𝑡𝑟𝑎𝑡𝑖𝑜𝑛
      (1) 

2.2  Architecture of the Feedforward Backpropagation Neural Network (FBNN) 

Model 

Data used in this study can be retrieved from the authors’ previous work [10]. 

Feedforward backpropagation neural network (FBNN), as shown in Fig. 2, has been 

47



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

commonly used in different fields of applications, particularly in developing non-linear 

mathematical/prediction models [16-17]. 

 

Fig. 2. The general architecture of the FBNN model. 

The net values at each hidden neuron (with first pattern inputs and random weight and 

bias) are presented as [16]: 

𝑛𝑒𝑡𝑣𝑎𝑙𝑢𝑒𝑗 = ∑ 𝑤𝑖𝑗 𝑥𝑗 + 𝑏𝑗
𝑛
𝑖=1   (2) 

where 𝑛𝑒𝑡𝑣𝑎𝑙𝑢𝑒𝑗  is net input to node i in hidden or output layer, 𝑥𝑗  are the inputs to node i 

(or output of the previous layer), 𝑤𝑖𝑗 are the weights representing the power of the 

relationship between the ith node and jth node, n is the number of nodes and 𝑏𝑗  is the bias 

related to node j.   

The transfer function is required to activate the neurons. In this research study, the 

sigmoid function is chosen as the activation function. 

ℎ𝑗 =
1

1+𝑒
−𝑛𝑒𝑡𝑣𝑎𝑙𝑢𝑒𝑗

   (3) 

where ℎ𝑗  is the output node of j and is an element of the inputs to the nodes in the next layer.  

The net values at the output layer and output neuron values are calculated by Eqs. (4) 

and (5) respectively. 

𝑛𝑒𝑡𝑣𝑎𝑙𝑢𝑒 = ∑ 𝑤𝑘 ℎ𝑘 + 𝐵
𝑚
𝑘=1   (4) 

𝑜𝑢𝑡𝑝𝑢𝑡 𝑛𝑒𝑢𝑟𝑜𝑛 𝑣𝑎𝑙𝑢𝑒𝑠 =
1

1+𝑒 −𝑛𝑒𝑡𝑣𝑎𝑙𝑢𝑒
   (5) 

where B is the bias.  

In this study, the inputs for the FBNN model were grass type, the thickness of the 

biofilter, and the initial concentration of E. coli. The expected resulting output of the model 

is the final concentration of E. coli. The architecture of the feedforward backpropagation 

neural network (FBNN) for the final E. coli concentration prediction is shown in Fig. 3. 

In addition, data sorting is one of the crucial procedures in developing any FBNN 

model. This is to ensure the smoothness of the overall process and to obtain a model with a 

respectively high level of accuracy. A proper size of training-testing data is required so that 

the model can learn enough possible input-output patterns [18-19]. There is no fixed 

guideline while setting the training to testing ratio. However, it was normally suggested to 

set the training dataset within the range of 60% to 80% while the remaining 20% to 40% 

becomes the testing dataset [20-21]. Since this study is considered as very first attempt to 

introduce the application of artificial intelligence in predicting the final concentration of 

48



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

fecal coliform with respect to different conditions of stormwater biofilters, the upper limit 

of 80% is selected so that the developed model will be provided with the most possible 

input-output patterns.  

 

Fig. 3: The architecture of the FBNN model. 

The number of hidden layers and the transfer function were set as one and a sigmoid 

function, respectively [21]. This is mainly due to its performance achievement in the 

prediction and forecasting model. Meanwhile, the training algorithm is selected as 

Levenberg-Marquardt (trainlm), since it is suited for function fitting (nonlinear regression) 

problems [22-24]. 

While designing the architecture of the FBNN, the determination of the number of 

hidden neurons is one of the main challenging tasks. This is due to the sensitivity of the 

networks to the number of hidden neurons. Underfitting problems may appear if there are 

too few neurons while overfitting issues may arise if there are too many neurons. Therefore, 

it is important to choose a proper number of neurons [25]. For this study, the hidden neurons 

were set within the ranges of 2 to 19. 

2.3  Architecture of Adaptive Neuro-Fuzzy Inference System (ANFIS) Model 

The integration of different techniques to form a hybrid AI model becomes the main 

trend of the development of AI applications. Adaptive neuro-fuzzy inference system 

(ANFIS) is a technique that integrates both neural networks and fuzzy logic principles 

within a single framework. This may strengthen the ability of the model to reach a higher 

level of accuracy [26-28]. A basic ANFIS architecture is presented in Fig. 4 [29-31].  

 
Fig. 4: The general architecture of ANFIS. 

49



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

Two rules were used in the method of “If-Then” for Takagi-Sugeno fuzzy model, as 

shown in the following: 

Rule 1: If x is 𝐴1 and y is 𝐵1, then 𝑓1 = 𝑝1𝑥 + 𝑞1𝑦 + 𝑟1    (6) 

Rule 2: If x is 𝐴2 and y is 𝐵2, then 𝑓2 = 𝑝2𝑥 + 𝑞2𝑦 + 𝑟2 (7) 

where 𝐴1, 𝐴2, 𝐵1 and 𝐵2 are the membership functions for each input x and y (part of the 
premises), f1and f2 are the outputs within the fuzzy region specified by the fuzzy rule, while 
𝑝1, 𝑞1, 𝑟1, 𝑝2, 𝑞2 and 𝑟2 are linear parameters in part- Then (consequent part) of Takagi-
Sugeno fuzzy inference model [26]. 

ANFIS architecture consists of five layers excluding the input layer (layer 0). The 

description of each layer is shown as follows [26]: 

1. Layer 0: It is an input layer that has n nodes, where n is the number of inputs to the 
system. 

2. Layer 1: It is the fuzzification layer. Every node in this layer adapts to a function 
parameter. The output from each node is a degree of membership value that is given 

by the input of the membership functions. The typical membership function is shown 

below: 

µ𝐴(𝑥) =
1

1+|
𝑥−𝑐𝑖

𝑎𝑖
|
2𝑏𝑖

   
(8) 

where 𝑎𝑖, 𝑏𝑖 and 𝑐𝑖 are parameters for the function. The parameters in this layer are 
defined as premise parameters. 

3. Layer 2: Every node in this layer is a fixed or nonadaptive node. The output is the 
product of all the incoming signals. Each node in this layer represents the fire 

strength for each rule. T-norm operator with general performance, such as the AND, 

is used to obtain the output: 

𝑂2𝑖 = 𝑤𝑖 = µ𝐴𝑖 (𝑥) ∗ µ𝐵𝑖 (𝑦), 𝑖 = 1, 2   (9) 

4. Layer 3: It is the normalization layer. Each node in this layer is fixed. Each node is 
a calculation of the ratio between the i-th rules firing strength and the sum of all 

rules’ firing strengths. The result is known as the normalized firing strength. The 

strength of all rules is normalized by: 

𝑂3𝑖 = �̄�𝑖 =
𝑤𝑖

∑ 𝑤𝑖𝑖
  (10) 

5. Layer 4: It is a layer of adaptive nodes. Every node in this layer is an adaptive node 
to output, with a node function defined as: 

𝑂4𝑖 = �̄�𝑖 𝑓𝑖 = �̄�𝑖 (𝑝𝑖 𝑥 + 𝑞𝑖 𝑦 + 𝑟𝑖 )  (11) 

where �̄�𝑖 is the normalized firing strength from the third layer and (𝑝𝑖 𝑥 + 𝑞𝑖 𝑦 + 𝑟𝑖 ) 
is a parameter in the node. The parameters in this layer are referred to as consequent 

parameters. It is assumed in Eq. (11) that all the universe of discourse for all input 

variables can be defined using the selected type of the membership functions, and 

the final output is computed using the regression parameters for each rule R. The 

regression parameters are the premise parameters in Eq. (11) which define the shape 

of the selected type of the membership function for each input variable. 

Consequently, the training process aims at tuning the premise and consequence 

parameters to achieve the desired output. 

50



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

6. Layer 5: It is an output layer whose function is the summation of net outputs of the 
nodes in the fourth layer using the formula as shown:  

∑ 𝑤𝑖 𝑓𝑖𝑖 =
∑ 𝑤𝑖𝑓𝑖𝑖

∑ 𝑤𝑖𝑖
  (12) 

While determining the input and output features as well as sorting data for ANFIS 

model development, a procedure that is similar to the FBNN model development was 

followed. The input features were made up of grass type, the thickness of stormwater, and 

the initial concentration of E. coli, while the output feature was the final concentration of E. 

coli, as depicted in Fig. 5. The number of membership function (mf) was set as three. In 

order to tune the patterns of the ANFIS network, the hybrid optimization method, which is 

the combination of backpropagation and least square-type approaches, was selected. The 

models were trained using different input membership functions, i.e. trimf, trapmf, gbellmf, 

gaussmf, gauss2mf, pimf, dsigmf and psigmf, and output membership function, i.e. constant 

and linear membership function. 

 

Fig. 5: The architecture of the ANFIS model. 

2.4  Model Performance Evaluation 

The commonly used analyses for model performance evaluation are coefficient of 

determination (R2), mean absolute error (MAE), mean squared error (MSE), root mean 

squared error (RMSE), and percentage error (% error).  These are the relevant important 

indicators to show the suitability of the developed model in predicting the final 

concentration of E. coli. 

R2 = (
𝑛 ∑ 𝑥𝑖𝑦𝑖−∑ 𝑥𝑖 ∑ 𝑦𝑖

√𝑛 ∑ 𝑥𝑖
2−(∑ 𝑥𝑖)²√𝑛 ∑ 𝑦𝑖

2−(∑ 𝑦𝑖)²

)

2

  (13) 

MAE =
∑|𝑦𝑖 − 𝑥𝑖 |

𝑛
 (14) 

51



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

RMSE = √
1

𝑛
∑(𝑦𝑖 − 𝑥𝑖 )² (15) 

Percentage error =
|𝑇𝑟𝑢𝑒 𝑣𝑎𝑙𝑢𝑒 − 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑣𝑎𝑙𝑢𝑒|

𝑇𝑟𝑢𝑒 𝑣𝑎𝑙𝑢𝑒
× 100% (16) 

where n is the number of data pairs, x is the observed variable, y is the predicted variable. 

3.   RESULTS AND DISCUSSION 

3.1  Experimental Works 

As shown in Fig. 6, the removal capability of Cow grass on a 450 mm depth river sand 

column together with 150 mm of topsoil and 100 mm of gravel drainage layer revealed the 

highest fecal coliform mean log removal (2.4 log) compared to other biofilter columns. It 

agrees with the findings in Barrett et al. [32] and Chandrasena et al. [5] which indicated that 

vegetated biofilters improved FC removal.   

 

Fig. 6: Mean log removal of fecal coliform of four vegetations vs four filter media. 

Planting of grasses, sedges, and shrubs in bioretention systems not only fulfils an 

esthetic purpose but also improves pathogen removal [33-35]. In the present study, Cow 

grass was found to be more suitable to use in stormwater biofilters because the survival rate 

of Cow grass was the highest compared to Pearl, Philippine, and Japanese grass. The bigger 

root mass of Cow grass improved the removal rates of FC due to its effect on biofilter 

retention time. It is hardy and able to grow with minimal to no fertilizer. The physical 

appearance of Pearl grass is similar to Cow grass but it has shorter, rounder, and thicker 

leaves. Pearl grass needs more water compared to Cow grass to grow and is less hardy. As 

for Philippine grass, it needs regular trimming about once every 2 weeks. Meanwhile, 

Japanese grass blades are softer, shorter, and compact. They grow rather slowly, require 

frequent watering, and in dry soil, they tend to die off.  

52



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

Apart from vegetation type, the physiochemical nature of filter media in biofilters plays 

a significant role in microbe removal [36-37]. In this study, the major component that 

constituted the filter media was washed river sand, offering an effective and low-cost means 

of treating stormwater. The finding is in-line with the previous studies which reported that 

sand filters showed satisfying outputs on fecal coliform removal in stormwater treatment 

[5]. Barrett et al. [32] reported that Austin sand filters achieved high FC (85%) and E. coli 

removal (97.1%) in an experiment to test for biofiltration performance. 

In terms of depth, the biofilter with 150 mm media depth exhibited an inconsistent 

performance in removing fecal coliform. This may be mainly because the preferential flow 

was more prone to occur in lower media depth due to the intermittent wet-drying cycle. It 

was found that, in order to achieve > 1 log fecal coliform removal, 250 mm should be the 

minimum media depth required. On the other hand, the mean FC removal at 350 mm depth 

filter was slightly higher than the 250 mm depth filter but it exceeded 2 log for all types of 

biofilter columns at 450 mm depth filter.  

3.2  Models Development  

3.2.1 FBNN Models  

The proposed FBNN models, which were developed with a range of hidden neurons 

from 2 to 19, were evaluated using the selected statistical analyses. However, since there 

were so many developed models, only the selected models were presented in this paper. 

Table 1 depicts the performance of all the developed FBNN models from the aspect of 

R2, MAE, RMSE and average percentage error (% error).  Based on the common theory, a 

higher R2-value indicates that the model has a higher ability to explain all the variance 

within the model. In this case, the highest R2-value is 0.9285, as shown in model IV. 

Table 1: Statistical performance of the selected FBNN models with respect to the different 

number of neurons 

Model Number of 

neurons 

R2 MAE RMSE 

I 14 0.8646 6.014 9.926 

II 15 0.0650 27.850 65.771 

III 16 0.8910 5.107 9.194 

IV 17 0.9285 4.204 7.322 

V 18 0.9249 4.512 7.423 

VI 19 0.0630 57.487 122.191 

While evaluating the performance of a FBNN model, a smaller error value is always 

favorable. This is because the smaller the calculated value, the better the accuracy of the 

estimated output. From the aspect of MAE and RMSE, model IV achieves the lowest values, 

displaying a value of 4.204 and 7.322, respectively. 

All the above-discussed aspects indicate that model IV is the best-performed model. 

The appropriateness of model IV to predict the final E. coli concentration is further verified 

using the average percentage error. Percentage error is another common indicator. Fig. 7 

contains the average percentage error of the selected FBNN models. Model IV shows the 

lowest average percentage error (27.5%), indicating that it has the highest level of accuracy 

among the examined models. 

53



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

 
Fig. 7: The average error of the developed FBNN models. 

Overall, model IV appears as the most suitable FBNN model to predict the final E. coli 

concentration for the stormwater biofilters application. This is because it has the highest R2-

value (0.9285), the lowest values of MAE (4.204) and RMSE (7.322), and the smallest 

average percentage error (27.5%).  

3.2.2 ANFIS Models  

A total number of 16 ANFIS models were developed, and their respective performances 

are contained in Table 2. In terms of R2-value, if the value is near 1, indicating that the 

observed values and the predicted values have a strong linear relationship. In other words, 

the observed values and predicted values are almost similar if the R2-value approximates 1. 

Referring to Table 2, the highest R2-value is 0.9874.  

On the other hand, the models with the output linear membership function (model II, 

IV, VI, VIII, X, XII, XIV, and XVI), in general, show a better performance than the models 

with constant output membership function. Therefore, it can be deduced that the output 

linear membership function is more suitable for the development of the ANFIS model to 

deal with the stormwater biofilters application. In this case, model IV displays the lowest 

MAE and RMSE values.    

Table 2: Performance in terms of R2, MAE and RMSE for the developed ANFIS models. 

Models NMFs MFTI MFTO R2 MAE RMSE 

I 3 trimf constant 0.7994 11.397 23.885 

II 3 trimf linear 0.9112 7.820 15.866 

III 3 trapmf constant 0.7994 11.397 23.885 

IV 3 trapmf linear 0.9874 3.854 6.004 

V 3 gbellmf constant 0.8231 11.260 22.428 

VI 3 gbellmf linear 0.9217 7.665 14.892 

VII 3 gaussmf constant 0.7909 11.451 24.394 

VIII 3 gaussmf linear 0.9215 7.706 14.908 

IX 3 gauss2mf constant 0.7784 11.471 25.126 

X 3 gauss2mf linear 0.9216 7.626 14.910 

XI 3 pimf constant 0.7714 11.597 25.534 

XII 3 pimf linear 0.9214 7.688 14.927 

XIII 3 dsigmf constant 0.7709 12.328 25.502 

XIV 3 dsigmf linear 0.9215 7.630 14.915 

XV 3 psigmf constant 0.7726 12.253 25.406 

XVI 3 psigmf linear 0.9218 7.629 14.892 

0

50

100

150

200

250

300

350

400

450

500

I II III IV V VI

A
v
e
ra

g
e
 P

e
rc

e
n
ta

g
e
 E

rr
o

r 
(%

)

FBNN model

54



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

From the perspective of average percentage error, model IV has achieved the lowest 

value if compared with other investigated models, recording at 14.2%, as shown in Fig. 8. 

In other words, it can achieve an average accuracy of around 86%. In general, no guideline 

was set for the range of acceptable error in certain engineering applications. However, a 

smaller error is always preferable. As this is the first attempt to introduce the use of ANFIS 

in predicting the final E. coli concentration for stormwater biofilters application, such an 

average accuracy is encouraging. 

 
Fig. 8. Average percentage error of the ANFIS model. 

In short, model IV is the model with the best performance while evaluating through the 

series of analytical analyses. It exhibits the highest R2-value of 0.9874, the lowest MAE and 

RMSE of 3.854 and 6.004 respectively, and the smallest average percentage error of 14.2%.  

3.2.3 AI Models Comparison 

This study investigates both FBNN and ANFIS as the advanced methods to predict the 

final E. coli concentrations. After conducting the performance evaluation through a series 

of statistical analyses, the best-performed model for each approach was identified. Table 3 

shows the comparison of the selected model in terms of R2, MAE, RMSE and average 

percentage error. 

Table 3: Comparison between the best-performed FBNN and ANFIS model 

Model R2 MAE RMSE Average percentage error 

FBNN model IV 0.9285 4.204 7.322 27.5 

ANFIS model IV 0.9874 3.854 6.004 14.2 

Among all the examined statistical indicators, ANFIS model IV achieved a better 

performance than that of FBNN model IV. Overall, it shows an improvement from FBNN 

to ANFIS. The most significant enhancement can be seen from the aspect of the average 

percentage error. The value has been reduced from 27.5% to 14.2%, showing an 

improvement of around 50%.  

The architecture of the best-performed ANFIS model is therefore described as follows: 

• Network inputs: Grass-type, the thickness of biofilters, initial E. coli concentration 

0

5

10

15

20

25

30

35

40

I II III IV V VI VII VIII IX X XI XII XIII XIV XV XVI

A
v
e
ra

g
e
 P

e
rc

e
n
ta

g
e
 E

rr
o

r 
(%

)

ANFIS model

55



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

• Network output: Final E. coli concentration 

• Number of membership functions: 3 

• Input membership function: trapmf 

• Output membership function: Linear 

• Optimization method: Hybrid 

4.   CONCLUSIONS 

The main purpose of this study was to develop an artificial intelligence (AI) model to 

serve as an alternative to predict the final E. coli concentration in the application of 

stormwater biofilters in stormwater management practices. Both feedforward 

backpropagation neural network (FBNN) and adaptive neuro-fuzzy inference system 

(ANFIS) models have seen their application in different fields of study, especially while 

dealing with non-linear regression problems. Both techniques are appropriate for this task 

because it has the capability to learn the relationships between input-output variables for a 

complex physical relationship and hence provide an output with a considerably high level 

of accuracy. 

In this study, it is found that a single-layer feedforward backpropagation neural network 

(FBNN) with 17 hidden neurons to be the most suitable model for the final E. coli 

concentration prediction. Meanwhile, the ANFIS model with the number of membership 

function of 3, input trapmf membership function and output linear membership function has 

shown the best performance among the examined models. The selection of the models was 

supported by the results of a range of statistical analyses. The selected FBNN model and 

ANFIS model were then further compared using the same series of statistical analyses to 

investigate their appropriateness to achieve the main goal of this study.  

In conclusion, the ANFIS model appears as the more suitable model for the final E. coli 

concentration prediction after comparing it with the selected FBNN model. In short, ANFIS 

is an effective tool to provide a more accurate simulation of the non-linear behavior between 

the final E. coli concentration and the factors affecting it. With such a model, it allows the 

user to determine the final E. coli concentration and thereby the removal percentage of the 

biofilters application by inserting the relevant input parameters into the model. Since this 

study is seen as the first attempt to implement the artificial intelligence techniques in 

predicting the final coliform concentration under different stormwater biofilter conditions, 

two basic techniques (BPNN and ANFIS) were chosen for the model development. To 

further enhance the model performance in terms of accuracy and effectiveness, the 

integration of optimization algorithms such as genetic algorithm (GA), ant colony algorithm 

(ACO), etc. to the proposed model can be performed. 

ACKNOWLEDGEMENTS  

This work was supported by University of Malaya under Postgraduate Research Grant (PPP) 

(Grant No. 4576). The authors would like to thank the Faculty of Engineering, University 

of Malaya for providing space to conduct the biofilter column experiments and Department 

of Parasitology, Faculty of Medicine, University of Malaya for the laboratory facilities and 

technical assistance provided for their experiments. 

 

56



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

REFERENCES  

[1]     Payne EG, Fletcher TD, Cook PL, Deletic A, Hatt BE. (2014) Processes and drivers of nitrogen 

removal in stormwater biofiltration. Crit. Rev. Environ. Sci. Technol., 44: 796-846. 

[2]     Kanda R, Kishimoto N, Hinobayashi J, Hashimoto T, Tanaka S, Murakami Y. (2017) Influence 

of temperature and COD loading on biological nitrification-denitrification process using a 

trickling filter: An empirical modeling approach. Int. J. Environ. Res., 11: 71-82. 

[3]     Samhan SA, Al-Sa'ed RM, Mahmoud NJ. (2007) Removal of pathogenic microorganisms in 

pilot‐scale UASB‐septic tanks and Albireh urban wastewater treatment plant in Palestine. 
Water Int., 32: 798-809. 

[4]     Wang R, Li X, Feng Y, Tariq F, Li K, Wei Y, . . . Chen L. (2019) Removal of formaldehyde 

from the air with a suspended growth bioreactor. Int. J. Environ. Res., 

doi:https://doi.org/10.1007/s41742-019-00228-2 

[5]     Chandrasena GI, Pham T, Payne EG, Deletic A, McCarthy DT. (2014) E. coli removal in 

laboratory scale stormwater biofilters: Influence of vegetation and submerged zone. J. 

Hydrol., 519: 814-822 

[6]     Li Y, McCarthy DT, Deletic A. (2016) Escherichia coli removal in copper-zeolite-integrated 

stormwater biofilters: Effect of vegetation, operational time, intermittent drying weather. Ecol. 

Eng., 90: 234-243. 

[7]     Chandrasena GI, Shirdashtzadeh M, Li Y, Deletic A, Hathaway, JM, Mccarthy DT. (2017) 

Retention and survival of E. coli in stormwater biofilters: Role of vegetation, rhizosphere 

microorganisms and antimicrobial filter media. Ecol. Eng., 102: 166-177. 

[8]     Galbraith P, Henry R, McCarthy DT. (2019) Rise of the killer plants: investigating the 

antimicrobial activity of Australian plants to enhance biofilter-mediated pathogen removal. J. 

Biol. Eng., 13: 52. 

[9]     Hunt WF, Smith JT, Jadlocki SJ, Hathaway JM, Eubanks PR. (2008) Pollutant removal and 

peak flow mitigation by a bioretention cell in urban Charlotte, N.C. J. Environ. Eng., 134: 

403-408. 

[10]   Bu CH, Lai SH, Goh XT, Chong WT, Chin RJ. (2021) Influence of filter media depth and 

vegetation on Faecal Coliform removal by stormwater biofilters. Water Environ. J., 35: 181-

189. 

[11]   Yetilmezsoy K, Ozkaya B, Cakmakci M. (2011) Artificial intelligence-based prediction 

models for environmental engineering. Neural Netw. World, 3: 193-218. 

[12]   Chau, K-w. (2006) A review on integration of artificial intelligence into water quality 

modelling. Mar. Pollut. Bull., 52: 726-733. 

[13]   Thompson KA, Dickenson ER. (2021) Using machine learning classification to detect 

simulated increases of de facto reuse and urban stormwater surges in surface water. Water 

Res., 204: 117556. 

[14]   Fang H, Jamali B, Deletic A, Zhang K. (2021) Machine learning approaches for predicting the 

performance of stormwater biofilters in heavy metal removal and risk mitigation. Water Res., 

200: 117273. 

[15]   DID Malaysia (2012) Urban stormwater management manual for Malaysia 2nd Edition. Kuala 

Lumpur, Department of Irrigation and Drainage Malaysia. 

[16]   Chin RJ, Lai SH, Shaliza I, Wan Zurina WJ, Ahmed Elshafie AH. (2019) New approach to 

mimic rheological actual shear rate under wall slip condition. Eng. Comput., 35: 1409-1418. 

[17]   Deng B, Chin RJ, Tang Y, Jiang C, Lai SH. (2019) New approach to predict the motion 

characteristics of single bubbles in still water. Appl. Sci., 9: 3981. 

[18]   Alimissis A, Philippopoulos K, Tzanis CG, Deligiorgi D. (2018) Spatial estimation of urban 

air pollution with the use of artificial neural network models. Atmos. Environ., 191: 205-213. 

[19]   Zhang Y, Chen H, Yang B, Fu S, Yu J, Wang Z. (2018) Prediction of phosphate concentrate 

grade based on artificial neural network modeling. Results Phys., 11: 625-628. 

[20]   Akter T, Desai S. (2018) Developing a predictive model for nanoimprint lithography 

usingartificial neural networks. Mater. Des., 160: 836-848. 

[21]   Chin RJ, Lai SH, Shaliza I, Wan Zurina WJ, Elshafie A. (2019) Rheological wall slip velocity 

prediction model based on artificial neural network. J. Exp. Theor. Artif. Intell., 31: 659-676. 

57



IIUM Engineering Journal, Vol. 23, No. 2, 2022 Lai et al. 
https://doi.org/10.31436/iiumej.v23i2.2173 

 

 

[22]   Pham DT, Sagiroglu S. (2001) Training multilayered perceptrons for pattern recognition: a 

comparative study of four training algorithms. Int. J. Mach. Tools Manuf., 41: 419-430. 

[23]   Demuth H, Beale M. (2014) Neural Network Toolbox for Use with MATLAB User's Guide 

Version 4. Natick, MA, The MathWorks Inc. 

[24]   Sharma B, Venugopalan K. (2014) Comparison of Neural Network Training Functions for 

Hematoma Classification in Brain CT Images. IOSR J. Comput. Eng., 16: 31-35. 

[25]   Alsmadi MK, Omar KB, Noah SA (2009) Back propagation algorithm: The best algorithm 

among the multi-layer perceptron. Int. J. Comput. Sci. Netw., 9: 378-383. 

[26]   Chin RJ, Lai SH, Shaliza I, Wan Zurina WJ, Elshafie A. (2020) ANFIS-based model for 

predicting actual shear rate associated with wall slip phenomenon. Soft Comput., 24: 9639-

9649. 

[27]   Ghasemi E, Kalhori H, Bagherpour R. (2016) A new hybrid ANFIS–PSO model for prediction 

of peak particle velocity due to bench blasting. Eng. Comput., 32: 607-614. 

[28]   Korotkikh V, Korotkikh G (2008) On principles in engineering of distributed computing 

systems. Soft Comput., 12: 201-206. 

[29]   Jang JS (1993) Adaptive network-based Fuzzy Inference System. IEEE Trans. Syst. Man 

Cybern. Syst., 23: 665-685. 

[30]   Jang JS, Sun CT. (1997) Neuro-Fuzzy and soft computing: A computing approach to learning 

and machine intelligence, Englewood Cliffs, New Jersey, Prentice Hall. 

[31]   Jang JS, Sun CT, Mizutani E. (1997) A computational approach to learning and machine 

intelligence. Neuro-Fuzzy and Soft Computing. Upper Saddle River, New Jersey,  Prentice-

Hall. 

[32]   Barrett ME, Limouzin M, Lawler DF. (2013) Effects of media and plant selection on 

biofiltration performance. J. Environ. Eng., 139: 462-470. 

[33]   Read J, Fletcher TD, Wevill T, Deletic A. (2009) Plant traits that enhance pollutant removal 

from stormwater in biofiltration systems. Int. J. Phytoremediation, 12: 34-53. 

[34]   Kim MH, Sung CY, Li MC (2012). Bioretention for stormwater quality improvement in Texas: 

removal effectiveness of Escherichia coli. Sep. Purif. Technol., 84: 120-124. 

[35]   Li YL, Deletice A, Alcazar L, Bratieres K, Fletcher TD, McCarthy DT. (2012) Removal of 

Clostridium perfringens, Escherichia coli and F-RNA coliphages by stormwater biofilters. 

Ecol. Eng., 49: 137-145. 

[36]   Hermawan AA, Chang JW, Pasbakhsh P, Hart F, Talei A. (2018) Halloysite nanotubes as a 

fine grained material for heavy metal ions removal in tropical biofiltration systems. Appl. Clay 

Sci., 160: 106-115. 

[37]   Sileshi R, Pitt R, Clark S. (2016) Prediction of flow rates through various stormwater biofilter 

media mixtures. Proceedings of the World Environmental and Water Resources Congress, 

West Palm Beach. 

 

58