Computers and Electronics in Agriculture 147 (2018) 158–165

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Soil microbial dynamics prediction using machine learning regression T

methods
⁎
Sunil Kr. Jhaa,b, , Zulfiqar Ahmadc,d
a
Department for Management of Science and Technology Development, Ton Duc Thang University, Ho Chi Minh City, Viet Nam
b
Faculty of Applied Sciences, Ton Duc Thang University, Ho Chi Minh City, Viet Nam
c
State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan, Hubei 430072, China
d
Department of Environmental Sciences, University of California, Riverside, CA 92521, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Soil microbial dynamics is significant for the soil productivity. The present study explores the application of
ANN machine learning based regression methods in the prediction of selected soil microbial dynamics, including
SVR bacterial population (BP), phosphate solubilization (PS), and enzyme activities. An experiment was designed in a
SC-FIS salt medium with rock phosphate inoculated with the solubilizing microorganism to measure the PS, BP, and 1-
WM-FIS
Aminocyclopropane-1-carboxylate (ACC) deaminase activity at a different temperature, pH, and incubation
Soil microbial dynamics prediction
period. The artificial neural network (ANN), support vector regression (SVR), Wang and Mendel’s (WM) - fuzzy
inference systems (FIS), and subtractive clustering (SC)-FIS methods have been applied in the estimation of PS,
BP, and ACC deaminase activity using the experimental conditions. The performance of four regression methods
has been evaluated in the terms of the coefficient of determination (R2), root mean square error (RMSE), and
correlation coefficient (ρ). The SC-FIS method has better performance than the rest three methods in the pre-
diction of each of the soil microbial dynamics (R2 of 0.99 in the prediction of PS).

1. Introduction multifunctional support to the soil ecosystem by increasing the crop

The soil is a vital system of life. It contains water, nutrients, mi-
croorganism, and inorganic and organic materials, etc. and provides the
living ecosystem of all organisms and plants. The condition of soil
changes over time due to continuous degradation and variations in its
composition (Doran et al., 1999; Karlen et al., 2003; Ghosh et al.,
2017). A proper balance of physical, chemical and biological con-
stituents is essential for maintaining the productivity of soil for agri-
culture (Schoenholtz et al., 2000; Paz et al., 2016). The estimation of
the constituents of soil is necessary for monitoring of strength of soil.
Microorganisms are a major part of biological constituents of soil and
have a greater contribution to the strength of soil than physical and
chemical constituents (Winding et al., 2005; Hermans et al., 2017).
Microorganisms exhibit high activity towards any change in soil com-
position and get used to environmental conditions (Lange et al., 2015;
Žifčáková et al., 2016). This adaptation hypothetically results in var-
iations in microbial populations and their activities. Consequently, it is
an indication of changes in soil condition that can be used in dis-
crimination in soil fitness (Van Bruggen and Semenov, 2000; Doran and
Zeiss, 2000; Hermans et al., 2017). Microorganisms have
production by controlling the weed, insects, and other infections, in-
creasing water and nutrients holding capacity, reutilizing nutrients, and
nitrogen-fixing, etc. (Finlay, 2004; Barrios, 2007; Adnan et al., 2017).
The accurate evaluation of soil microbial conditions and fitness of
soil is one of the significant research objectives at the present (Doran
et al., 1999; Liu et al., 2006; Armenise et al., 2013). The phosphate
solubilization (PS), bacterial population (BP), and 1-Aminocyclopro-
pane-1-carboxylate (ACC) deaminase activity are substantial soil mi-
crobial condition indicators. Consequently, their estimation is sig-
nificant in soil fitness monitoring. The prediction of PS, BP, and ACC
deaminase activity using mathematical models results in lower esti-
mation accuracy. The artificial intelligence (AI) methods have ad-
vanced computing capability and better efficiency, therefore widely
adopted in modeling applications
Since last few years, some statistical and AI based methods have
been implemented successfully in soil parameter modeling and classi-
fication (Coopersmith et al., 2014; Shiri et al., 2017a, 2017b; Sirsat
et al., 2017; Wu et al., 2018) and modeling of soil microbial dynamics
(Hughes et al., 2001; Haider et al., 2008; Tajik et al., 2012; Mukhlisin
et al., 2012; Taghavifar and Mardani, 2014; Ebrahimi et al., 2017). K-

⁎
Corresponding author.
E-mail address: drsuniljha@tdt.edu.vn (S.K. Jha).

https://doi.org/10.1016/j.compag.2018.02.024
Received 23 October 2017; Received in revised form 18 February 2018; Accepted 19 February 2018
0168-1699/ © 2018 Elsevier B.V. All rights reserved.
S.K. Jha, Z. Ahmad
nearest-neighbor and boosted perceptron algorithm (Coopersmith et al.,
2014), gene expression programming (GEP), and artificial neural net-
work (ANN) (Shiri et al., 2017a, 2017b; Wu et al., 2018; Haider et al.,
2008; Tajik et al., 2012; Mukhlisin et al., 2012; Taghavifar and
Mardani, 2014; Ebrahimi et al. 2017), support vector machine (SVM)
(Shiri et al., 2017a, 2017b; Sirsat et al., 2017; Wu et al., 2018), neuro-
fuzzy (Shiri et al., 2017a, 2017b), random forest (RF) (Sirsat et al.,
2017; Shiri et al., 2017a, 2017b), regression analysis (Tajik et al., 2012;
Ebrahimi et al. 2017) are the commonly used methods in soil parameter
modeling. Amongst earlier stated methods, ANN and fuzzy methods
result in better accuracy in the modeling of nonlinear problems, like
modeling soil cation exchange (Shiri et al., 2017a, 2017b) and water
capacity (Shiri et al., 2017a, 2017b), and estimation of biosurfactant
production (Ahmad et al., 2016). Both methods have some pros and
cons like for ANN method, it is difficult to understand the rules of the
model due to weights and error adjustment in each iteration but easy in
learning from the data. Fuzzy logic methods use if-then rules and it is
easy to interpret them. Meanwhile, the combination of ANN and fuzzy
method based fuzzy inference system (FIS) results in better decision
rules in modeling of a system (Shiri et al., 2017a, 2017b). FIS methods
are more efficient in several research applications. Specifically, two
categories of FIS methods have been implemented in modeling, in-
cluding the Wang and Mendel’s (WM) rule-based FIS, and subtractive
clustering (SC)-FIS (Eftekhari and Katebi, 2008; Yang et al., 2010;
Lohani et al., 2014).
Though we have not noticed the application of FIS methods in the
prediction of soil microbial dynamics. With this motivation, we have
implemented the WM-FIS and SC-FIS methods in the prediction of soil
microbial dynamics. Also, the performance of FIS methods has been
compared with the ANN and support vector regression (SVR) methods
in terms of root mean square error (RMSE), correlation coefficient in
between the actual and predicted values of the soil microbial dynamics
and the coefficient of determination (R2). The FIS methods create a
linguistic regression relation in between soil microbial dynamics and
experimental condition and results in better performance than the ANN,
and SVR methods.
2. Analysis methods
The regression analysis method is used to establish a relationship
between the dependent (soil microbial dynamics) and independent
variables (experimental conditions). Regression methods are categor-
ized into linear and nonlinear methods, including linear regression
method, logistic regression method, polynomial regression method, etc.
(Kleinbaum et al., 2013). In the present study, four AI based non-linear
regression methods, including ANN, SVR, WM-FIS, and SC-FIS have
been implemented. The non-linear regression models have been for-
mulated for PS, BP, and ACC deaminase activity, independently using
the earlier mentioned methods and their performance is compared in
order to obtain optimal prediction accuracy. A short description and
implementation details of four regression methods are as follows.
2.1. Artificial neural network (ANN)
ANN is an organization of artificial neurons in different layers
(input, hidden, and output) (Theodoridis and Koutroumbas, 2006). The
basic parameters of artificial neurons, including weight, threshold, and
activation functions were optimized during the training of ANN method
when used for regression analysis. Back-propagation feed forward ar-
tificial neural network (BPNN) is selected in the prediction of soil mi-
crobial dynamics and experimental conditions. The Levenberg-Mar-
quardt algorithm is used in training of BPNN method. A 3 × 10 × 1
structure of BPNN method with sigmoid activation function in the
hidden layer is implemented in R (R Core Team, 2017) using the nnet
package (Venables and Ripley, 2002) in the prediction of PS, BP, and
ACC deaminase activity using the experimental parameters, incubation
159
Computers and Electronics in Agriculture 147 (2018) 158–165
period, temperature and pH as inputs. The minimum error is achieved
at 29, 23 and 20 iterations for PS, BP, and ACC deaminase activity,
respectively. The details of the method can be seen in Ref. (Theodoridis
and Koutroumbas, 2006; Venables and Ripley, 2002).
2.2. Support vector regression (SVR)
Support vector machine (SVM) is termed as support vector regres-
sion (SVR) when used for the regression analysis (Theodoridis and
Koutroumbas, 2006; Bishop, 2006). In the SVR method the actual va-
lues of PS, BP, and ACC deaminase activity are assumed as output and
experimental condition, incubation period, temperature and pH as in-
puts. The kernel function (linear, Gaussian, sigmoid, radial basis, and
polynomial, etc.) computes the inner product of input feature vectors in
high dimensional space and reduces the time and space complexity of
the SVR method. The Radial basis kernel function is used in SVR
method in the prediction of PS, BP, and ACC deaminase activity in-
dependently using the e1071 package in R (Meyer et al., 2017). The
kernel function has similar values of parameters such as cost = 1,
γ = 0.33 and ξ = 0.1 for SVR models in the prediction of PS, BP, and
ACC deaminase activity. Though, the number of support vectors varies
(34, 19, and 29 for three SVR models, respectively).
2.3. Fuzzy inference system (FIS)
FIS creates a map in between the dependent and independent
variables using the fuzzy logic. In which, initially, a partial membership
is assigned to each input, thereafter complete membership is for-
mulated by composing the partial truths using certain if-then rules (Klir
and Yuan, 1995). Several clustering methods (subtractive, mountain, c-
means, etc.) have been used in the implementation of FIS. Amongst
them, subtractive clustering based FIS and Wang and Mendel’s rule-
based general FIS have been implemented in the prediction of soil
microbial dynamics. A short description of both methods is as follows.
2.3.1. Wang and Mendel’s fuzzy inference system (WM-FIS)
WM-FIS is the most basic type of fuzzy inference system. The
method is described in (Wang and Mendel, 1992). WM-FIS is im-
plemented using the frbs package in R (Riza et al., 2015). The method
divides the independent and dependent variable spaces into several
fuzzy regions and assign a partial membership, generates if-then rules
and assign a degree to each rule, then creates a joint rule for final
mapping of a test input. The basic steps and implementation details of
the method are shown in Fig. 1(a). The details of the method can be
seen in Ref. (Wang and Mendel, 1992).
2.3.2. Subtractive clustering fuzzy inference system (SC-FIS)
In SC-FIS, initially, SC finds fuzzy clusters according to a Gaussian
measure. Data points close to a point of maximum potential (center) are
assigned to the first cluster. The rest of the data points are assigned to
the second cluster and the process is repeated till all data points are
clustered. After that fuzzy if-then rules are generated using previous
clusters and the dependent variable is predicted using rules (Chiu,
1994). A schematic representation of basic steps and implementation
details of SC-FIS method is shown in Fig. 1(b). The method is im-
plemented using the frbs package in R (Riza et al., 2015).
3. Materials and methods
3.1. Sampling of soil
The wheat plants were uprooted at tillering stage and stored in
polythene bags. The non-rhizospheric soil was detached by agitating
the roots and thereafter the soil strictly adhering to the roots was col-
lected as a rhizospheric soil sample. A composite sample was prepared
by using seven rhizospheric samples of wheat plants. For the simplicity,
S.K. Jha, Z. Ahmad
Fig. 1. An algorithm table of (a) WM-FIS and (b) SC-FIS methods.
the composite sample is termed as WRS (wheat root rhizospheric
samples). The measured values of soil microbial dynamics including,
BP, PS and ACC deaminase activity using the Taguchi design at a dif-
ferent value of three experimental conditions, including the incubation
period, temperature and pH are summarized in Table 1.
3.2. Analytical measurements
3.2.1. PS activity measurements
The amount of solubilization of rock phosphate in NBRIP broth
(glucose 10 g, rock phosphate 5 g, MgCl2·6H2O 5 g, MgSO4·7H2O 0.25 g,
KCl 0.2 g, (NH4)2SO4 0.1 g/L) media is used to define the PS activity of
microorganism in the soil sample (Chen et al., 2006). The procedure is
completed by dissolving 0.1 g rhizospheric soil samples in 100 ml
Tryptic soy broth at 100 RPM, 28 °C and 18 h incubation, subsequently,
turbidity is used to monitor the development of microorganism. In the
160
Computers and Electronics in Agriculture 147 (2018) 158–165
next step, 20 µL of inoculums was included in 50 ml modified NBRIP
broth for 72 h of incubation then NBRIP broth inoculated media was
separated. Finally, the P content in samples was measured at 410 nm
(Olsen, 1982).
3.2.2. BP measurement
The spread plate method using Marine Agar sterilized at 121 °C,
1 atm for 20 min is adopted in the counting of rhizospheric bacteria.
The medium is diluted in 34 g/L sterile sodium chloride. Plates have
been assembled at temperature 25–40 °C and incubation period of
3–12 days. The population of bacteria was measured in colony-forming
units (CFU) per ml (Leonard et al., 2000).
3.2.3. ACC deaminase activity measurement
1-Aminocyclopropane-1-carboxylate (ACC) deaminase activity
measures the amount of α- ketobutyrate produced by the cleavage of
S.K. Jha, Z. Ahmad Computers and Electronics in Agriculture 147 (2018) 158–165

Table 1
Experimental values of phosphate solubilization (PS), bacterial population (BP), and ACC activity.

No. Incubation period in days Temperature in oC pH PS in μg ml−1 BP in CFU ml−1 ACC activity in nM α-ketobutyrate mg−1 h−1)

1. 3 25 6 10.52 2.00E+05 1240.05

2. 3 25 7 10.52 1.00E+07 657.23
3. 3 25 8 123.24 4.00E+07 1311.63
4. 3 25 9 169.29 3.00E+05 430.01
5. 3 30 6 146.27 3.00E+05 821.97
6. 3 30 7 151.53 6.00E+03 1399.11
7. 3 30 8 6.88 1.00E+02 3.81
8. 3 30 9 170.39 4.00E+06 848.10
9. 3 35 6 142.10 2.00E+05 1063.96
10. 3 35 7 79.82 1.00E+05 252.78
11. 3 35 8 120.83 3.00E+04 269.82
12. 3 35 9 74.56 3.00E+06 859.46
13. 3 40 6 3.77 1.00E+02 2.34
14. 3 40 7 124.56 4.00E+05 393.66
15. 3 40 8 0.13 1.00E+02 7.02
16. 3 40 9 5.13 1.00E+02 3.37
17. 6 25 6 131.79 5.00E+05 94.86
18. 6 25 7 103.72 1.00E+06 422.06
19. 6 25 8 113.59 2.00E+06 440.24
20. 6 25 9 71.71 2.00E+06 531.12
21. 6 30 6 119.51 1.00E+04 305.04
22. 6 30 7 9.29 1.00E+02 1.15
23. 6 30 8 119.95 7.00E+04 366.39
24. 6 30 9 133.11 1.00E+05 401.61
25. 6 35 6 122.58 1.00E+05 61.91
26. 6 35 7 117.10 1.00E+04 138.03
27. 6 35 8 8.68 1.00E+02 8.60
28. 6 35 9 91.66 2.00E+04 295.95
29. 6 40 6 138.59 7.00E+04 170.98
30. 6 40 7 0.74 1.00E+02 0.46
31. 6 40 8 12500.00 3.00E+05 433.42
32. 6 40 9 111.40 1.00E+05 336.85
33. 9 25 6 50.87 1.00E+07 587.93
34. 9 25 7 123.90 5.00E+05 309.58
35. 9 25 8 124.78 1.00E+04 199.38
36. 9 25 9 122.36 2.00E+04 255.05
37. 9 30 6 128.28 9.00E+03 677.68
38. 9 30 7 270.39 9.00E+03 551.57
39. 9 30 8 116.66 2.00E+04 328.90
40. 9 30 9 100.00 1.00E+05 906.04
41. 9 35 6 102.85 1.00E+05 215.29
42. 9 35 7 111.62 2.00E+05 308.45
43. 9 35 8 95.61 2.00E+05 408.42
44. 9 35 9 105.92 3.00E+05 1118.49
45. 9 40 6 9.86 1.00E+02 6.61
46. 9 40 7 117.54 7.00E+03 516.35
47. 9 40 8 96.92 1.00E+05 128.94
48. 9 40 9 138.37 1.00E+05 257.32
49. 12 25 6 132.67 2.00E+04 1155.98
50. 12 25 7 0.24 1.00E+02 0.53
51. 12 25 8 4.64 1.00E+02 2.88
52. 12 25 9 0.42 1.00E+02 1.00
53. 12 30 6 113.59 1.00E+04 311.86
54. 12 30 7 122.58 1.00E+05 267.55
55. 12 30 8 119.73 9.00E+04 472.05
56. 12 30 9 0.52 1.00E+02 1.84
57. 12 35 6 112.71 2.00E+06 351.63
58. 12 35 7 117.10 3.00E+04 431.15
59. 12 35 8 125.21 4.00E+05 343.67
60. 12 35 9 137.50 1.00E+06 168.71
61. 12 40 6 1600.00 1.60+06 20.61
62. 12 40 7 3700.00 3.70E+06 8.77
63. 12 40 8 9500.00 9.50E+06 13.81
64. 12 40 9 3400.0 3.40E+06 13.81

ACC. ACC deaminase activity was evaluated according to Ref. (Penrose
and Glick, 2003). Tris-HCl (0.1 M, pH 8.5, 4 °C) was used to prepare a
standard solution of α-ketobutyrate. Specifically, absorbance curves of
an experimental sample and a standard α-ketobutyrate (0.1–1 nmole at
540 nm) were compared to determine the quantity of α-ketobutyrate.
Protein estimation was used to measure the specific activity of the
cultures (Lowry et al., 1951)
4. Prediction outcomes of regression methods
Figs. 2–4 presents the plots of predicted values of PS, BP, and ACC
deaminase activity vs. the actual values using their ANN, SVR, WM-FIS
and SC-FIS methods, respectively. The quantitative accuracy of four
methods has been evaluated in terms of thee evaluation metrics (Steel
and Torrie, 1960), including (i) Pearson correlation coefficient (ρ), (ii)

161
S.K. Jha, Z. Ahmad Computers and Electronics in Agriculture 147 (2018) 158–165

Actual PS ANN predicted PS SVR predicted PS WM predicted PS SBC predicted PS

10000

1000
Phosphate solubilization (PS)) on log scale

100

10

1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63

0.1

Measurement sample number

Fig. 2. ANN, SVR, WM-FIS and SC-FIS methods in the prediction of phosphate solubilization.

Fig. 3. ANN, SVR, WM-FIS and SC-FIS methods in the prediction of bacterial population.

root mean square error (RMSE), and (iii) coefficient of determination actual, model predicted values of dependent variable, and total number
(R2) (Table 2). The Pearson correlation coefficient (ρ) is defined as of observation, respectively.
n n n n n 2 n
ρ = ∑i Ai Pi −B / C × D ..B = ∑i = 1 Ai ∑i = 1 Pi / n, C = ∑i = 1 Ai2 −(∑i = 1 Ai ) / n , RMSE and R2 are calculated as RMSE = ∑i (Ai −Pi )2/n and
n n 2
and D = ∑i = 1 Pi2−(∑i = 1 Pi ) / n·Ai ,Pi and n . Ai , Pi and n represents the R2 = 1− ∑ (Ai −Pi )2 / ∑ (Ai −A )2 , respectively. Table 3 summarizes some

162
S.K. Jha, Z. Ahmad Computers and Electronics in Agriculture 147 (2018) 158–165

Fig. 4. ANN, SVR, WM-FIS and SC-FIS methods in prediction of ACC activity.

Table 2 values. Consequently, a higher value of ρ indicated better performance

Performance measures of ANN, SVR, WM-FIS and SC-FIS Methods. of the prediction model. The coefficient of determination (R2) measures
the confident strength in between the actual values and model pre-
Methods Predicted Evaluation measures
microbial dicted values of PS, BC, and deaminase activity. The value of R2 lies in
dynamics Correlation Root mean Coefficient of between 0 and 1, R2 → 1 signifies better performance of the prediction
coefficient (ρ) square error determination (R2) model. The value of R2 provides an idea about the unpredictability of
(RMSE) dependent variables.
ANN 1. PS 0.81 775.04 0.66 The value of RMSE indicates the precision of any predictive model.
2. BP 0.08 8359733.00 0.00 The lower value of RMSE in between the actual and predicted values
3. ACCA 0.47 331.13 0.22 signifies higher precision of the model. The large difference in between
SVR 1. PS 0.04 1335.36 0.00 the actual and model predicted values results in a higher value of RMSE
2. BP 0.58 4907319.00 0.32 and hence lower precision of the prediction model. The actual and
3. ACCA 0.52 310.98 0.27 model predicted values of PS by using the ANN, SVR, WM-FIS, and SC-
WM-FIS 1. PS 0.78 1020.00 0.60 FIS is shown in Fig. 2 on a log scale. Based on the visual inspection, it is
2. BP 0.50 4660000.00 0.25 obvious that the SC-FIS method predicted values of PS are adjacent to
3. ACCA 0.54 363.00 0.29
their actual measured values with the exception of sample number 40.
SC-FIS 1. PS 1.00 45.28 1.00 The most deviation is observed for the ANN model predicted values
2. BP 0.97 1350000.00 0.94 of PS. The SVR and WM-FIS models predicted values lie in between
3. ACCA 0.72 271.00 0.52
actual and SC-FIS model predicted values of PS. The SVR predicted
values are closer to the actual values of PS than the ANN predicted
combinations of experimental independent variables for which the best values and WM-FIS predicted values are closer to the actual values than
prediction results are achieved. the ANN and SVR predicted values of PS. This fact is further established
in quantitative evaluation metrics of four methods (Table 2). The SC-FIS
method has the maximum values of R2 = 0.9998 and ρ = 0.9994 , and the
5. Discussion minimum value of RMSE = 45.28, while the ANN method has the
minimum value of R2 = 0.66 and ρ = 0.81, and the maximum value of
The outcomes of ANN, SVR, WM-FIS, and SC-FIS methods in the RMSE = 775.04.
prediction of PS, BP, ACC deaminase activity and their performance Furthermore, the best accuracy of SC-FIS is obtained in the pre-
assessment in terms of ρ, RMSE, and R2 have been presented in diction of PS than the BP and ACC deaminase activity according to
Figs. 2–4, Table 2. visual inspection of Figs. 2–4, and ρ, RMSE and R2 value in Table 2.
The value of correlation coefficient ρ lies in between −1 and +1. Fig. 3 presents the predicted results of BP by using the ANN, SVR, WM-
For ρ → 1 specifies a perfect relationship between the actual values of FIS, and SC-FIS methods. The SC-FIS method predictions are the best
PS, BC, and ACC deaminase activity and their model predicted values. ρ matching with the actual values of BP except for fewer samples than the
characterizes the relationship between actual and model predicted rest three methods. The quantitative analysis (Table 2) also approve the

163
S.K. Jha, Z. Ahmad Computers and Electronics in Agriculture 147 (2018) 158–165

Table 3
Best combinations of experimental variables in the prediction of each of soil microbial dynamics with respect to SC-FIS method.

Input parameters and their value Soil microbial Actual value SC-FIS predicted WM-FIS predicted SVR predicted ANN predicted value
dynamics types value value value
Incubation period in Temperature (°C) pH
days

6 25 8 PS 113.59 113.56 0.13 117.28 54.76

12 25 9 0.1305 0.29 0.42 114.68 112.95

9 25 7 BP 500000.00 500000.00 22200.00 1751934.33 3016238.26

12 25 6 20000.00 20000.00 1.62 2036694.30 −16139575.02

3 35 8 ACCA 269.82 269.86 832.85 232.21 208.42

6 25 7 422.06 422.00 747.41 384.21 615.73
12 25 9 0.9989 1.00 8.31 320.52 198.03

better performance of SC-FIS method than the WM-FIS, SVR, and ANN
methods (the maximum values of R2 = 0.9356 and ρ = 0.9673, and a
minimum value of RMSE = 1350000). Fig. 4 presents the visual de-
monstration of actual values and predicted values of ACC deaminase
activity by using ANN, SVR, WM-FIS and SC-FIS methods. Like, PS and
BP, most of the SC-FIS predicted values of ACC deaminase activity are
closer to their actual values.
Furthermore, SC-FIS method has a better prediction accuracy for
ACC deaminase activity than the ANN, SVR, and WM-FIS methods. The
maximum values of R2 and ρ , and a minimum value of RMSE for SC-FIS
method than the other three methods in prediction of ACC deaminase
activity further proves this fact (Table 2). In the modeling of PS, BP and
ACC deaminase activity using the ANN, SVR, WM-FIS and SC-FIS
methods, it is obvious that the SC-FIS method has the best prediction
accuracy than rest three methods. Furthermore, it was noticed that
there are some combinations of input experimental variables for which
the SC-FIS predicted values of PS, BP, and ACC deaminase activity are
exactly or approximately equal to their actual values. Best combina-
tions of such experimental parameters for each of the PS, BP, and ACC
deaminase activity are summarized in Table 3. For instance, the ex-
perimental conditions, incubation days equal to 6 at temperature 25 °C
and pH 8 results in 100% accuracy of SC-FIS method in the prediction
of PS (actual value 113.59 and predicted value 113.56 of PS).
The experimental conditions, incubation days equal to 9 at tem-
perature 25 °C and pH 7, and incubation days equal to 12 at tempera-
ture 25 °C and pH 6 results in 100% accuracy of SC-FIS method in the
prediction of BP (Table 3). Likewise, the experimental conditions, in-
cubation days equal to 3 at temperature 35 °C and pH 8, incubation
days equal to 6 at temperature 25 °C and pH 7, and incubation days
equal to 12 at temperature 25 °C and pH 9, results in 100% prediction
accuracy of SC-FIS method in the prediction of ACC deaminase activity.
The optimal value of experimental conditions in combinations (Table 3)
can be used further in the precise monitoring of soil microbial dynamics
and soil fitness for better productivity. The estimation results of ACC
deaminase activity can be used to check the growth of a plant.
The fuzzy methods have better performance in adjusting the va-
gueness in the data set, therefore, result in better performance in pre-
dictive modeling. This fact is established in some previous studies based
on the application of WM-FIS method in storage time prediction of pork
(Barbon et al., 2016), human-machine system (Zhang and Xia et al.,
2017), and mental models analysis (Garcia-Nunes et al., 2017). More-
over, the WM-FIS method results in better performance than the ANN
and SVM methods (Barbon et al., 2016). The SC-FIS method also results
in better performance in the modeling of soil cation exchange capacity
(Keshavarzi et al., 2012), performance prediction of the road header
(Yazdani-Chamzini et al. 2013), and fault detection and classification
(Chudasama et al., 2016). Nevertheless, the WM-FIS and other FIS
methods have been not implemented in the prediction of soil microbial
dynamics estimation. Consequently, two FIS methods have been im-
plemented in the present study in predictive modeling of soil microbial
dynamics and their performance is compared with ANN and SVR
methods. The better prediction accuracy of SC-FIS than WM-FIS is due
to the fact that the earlier creates the minimum number of fuzzy if-then
rules on the basis of similarity in the data.
The distribution of experimental conditions and soil microbial dy-
namics may affect the performance of FIS methods and transfer learning
can be implemented in such situation (Zhang and Zuo et al., 2016;
Zhang and Yang et al., 2017). The performance of FIS methods in
predictive modeling of soil microbial dynamics can be improved further
using the transfer learning , extreme learning (Zhang and Zhang, 2016),
and subspace learning (Zhang and Liu et al., 2017) in future studies.
6. Conclusion
The study presents a comprehensive performance analysis of ANN,
SVR, WM-FIS and SC-FIS methods in the predictive molding of soil
microbial dynamics, including PS, BP, and ACC-deaminase activity
using varying experimental parameters (incubation period, tempera-
ture, and pH). The SC- FIS method has the best and ANN has the worst
prediction performance on the basis of ρ, RMSE, and R2. An average
performance of WM-FIS and SVR methods has been noticed. The PS has
been predicted accurately than the BP and ACC-deaminase activity.
Acknowledgements
The authors acknowledge anonymous reviewers for their valuable
comments and suggestions.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in the
online version, at http://dx.doi.org/10.1016/j.compag.2018.02.024.
References
Adnan, M., Shah, Z., Fahad, S., Arif, M., Alam, M., Khan, I.A., Mian, I.A., Basir, A., Ullah,
H., Arshad, M., Saud, S., 2017. Phosphate-solubilizing bacteria nullify the antag-
onistic effect of soil calcification on bioavailability of phosphorus in alkaline soils.
Sci. Rep. 7, 16131.
Ahmad, Z., Crowley, D., Marina, N., Jha, S.K., 2016. Estimation of biosurfactant yield
produced by Klebseilla sp. FKOD36 bacteria using artificial neural network approach.
Measurement 81, 163–173.
Armenise, E., Redmile-Gordon, M.A., Stellacci, A.M., Ciccarese, A., Rubino, P., 2013.
Developing a soil quality index to compare soil fitness for agricultural use under
different managements in the Mediterranean environment. Soil Tillage Res. 130,
91–98.
Barbon, A.P.A., Barbon Jr, S., Mantovani, R.G., Fuzyi, E.M., Peres, L.M., Bridi, A.M., 2016.
Storage time prediction of pork by computational intelligence. Comput. Electron.
Agric. 127, 368–375.
Barrios, E., 2007. Soil biota, ecosystem services and land productivity. Ecol. Econ. 64,
269–285.
Bishop, C.M., 2006. Pattern Recognition and Machine Learning. Springer, New York,
USA.
Chen, Y.P., Rekha, P.D., Arun, A.B., Shen, F.T., Lai, W.A., Young, C.C., 2006. Phosphate
solubilizing bacteria from subtropical soil and their tricalcium phosphate solubilizing
abilities. Appl. Soil Ecol. 34 (1), 33–41.
Chiu, S., 1994. Fuzzy model identification based on cluster estimation. J. Intell. Fuzzy

164
S.K. Jha, Z. Ahmad
Syst. 2 (3), 267–278.
Chudasama, K., Shah, V., Shah, S., 2016. Induction motor relaying scheme for external
faults detection and classification using subtractive clustering based Sugeno fuzzy
inference system. Electr. Power Compon. Syst. 44 (10), 1149–1162.
Coopersmith, E.J., Minsker, B.S., Wenzel, C.E., Gilmore, B.J., 2014. Machine learning
assessments of soil drying for agricultural planning. Comput. Electron. Agric. 104,
93–104.
Doran, J.W., Jones, A.J., Arshad, M.A., Gilley, J.E., 1999. Determinants of soil quality and
health. In: Lal, R. (Ed.), Soil Quality and Soil Erosion. CRC Press, USA, pp. 17–38.
Doran, J.W., Zeiss, M.R., 2000. Soil health and sustainability: managing the biotic com-
ponent of soil quality. Appl. Soil Ecol. 15, 3–11.
Ebrahimi, M., Sinegani, A.A.S., Sarikhani, M.R., Mohammadi, S.A., 2017. Comparison of
artificial neural network and multivariate regression models for prediction of
Azotobacteria population in soil under different land uses. Comput. Electron. Agric.
140, 409–421.
Eftekhari, M., Katebi, S.D., 2008. Extracting compact fuzzy rules for nonlinear system
modeling using subtractive clustering, GA and unscented filter. Appl. Math. Model.
32, 2634–2651.
Finlay, R.D., 2004. Mycorrhizal fungi and their multifunctional roles. Mycologist 18,
91–96.
Garcia-Nunes, P.I., Souza, R.M., da Silva, A.E.A., 2017. Mental models analysis and
comparison based on fuzzy rules: A case study of the protests of June and July 2013
in Brazil. IEEE Transact. Syst., Man, Cybernetics: Syst. 47 (8), 2021–2033.
Ghosh, P.K., Palsaniya, D.R., Kumar, T.K., 2017. Resource conservation technologies for
sustainable soil health management. In: Rakshit, A., Abhilash, P., Singh, H., Ghosh, S.
(Eds.), Adaptive Soil Management: From Theory to Practices. Springer, Singapore, pp.
161–187.
Haider, M.A., Pakshirajan, K., Singh, A., Chaudhry, S., 2008. Artificial neural network-
genetic algorithm approach to optimize media constituents for enhancing lipase
production by a soil microorganism. Appl. Biochem. Biotechnol. 144, 225–235.
Hermans, S.M., Buckley, H.L., Case, B.S., Curran-Cournane, F., Taylor, M., Lear, G., 2017.
Bacteria as emerging indicators of soil condition. Appl. Environ. Microbiol. 83
e02826-16.
Hughes, J.B., Hellmann, J.J., Ricketts, T.H., Bohannan, B.J., 2001. Counting the un-
countable: statistical approaches to estimating microbial diversity. Appl. Environ.
Microbiol. 67, 4399–4406.
Karlen, D.L., Ditzler, C.A., Andrews, S.S., 2003. Soil quality: why and how? Geoderma
114, 145–156.
Keshavarzi, A., Sarmadian, F., Rahmani, A., Ahmadi, A., Labbafi, R., Iqbal, M.A., 2012.
Fuzzy clustering analysis for modeling of soil cation exchange capacity. Austral. J.
Agricult. Eng. 3 (1), 27–33.
Kleinbaum, D.G., Kupper, L.L., Nizam, A., Rosenberg, E.S., 2013. Applied Regression
Analysis and Other Multivariable Methods. Cengage Learning.
Klir, G., Yuan, B., 1995. Fuzzy Sets and Fuzzy Logic (Vol. 4). Prentice Hall, New Jersey,
USA.
Lange, M., Eisenhauer, N., Sierra, C.A., Bessler, H., Engels, C., Griffiths, R.I., Mellado-
Vázquez, P.G., Malik, A.A., Roy, J., Scheu, S., Steinbeiss, S., 2015. Plant diversity
increases soil microbial activity and soil carbon storage. Nat. Commun. 6, 6707.
Leonard, N., Blancheton, J.P., Guiraud, J.P., 2000. Populations of heterotrophic bacteria
in an experimental recirculating aquaculture system. Aquacult. Eng. 22 (1), 109–120.
Liu, Z.F., Fu, B.J., Liu, G.H., Zhu, Y.G., 2006. Soil quality: concept, indicators and its
assessment [J]. Acta Ecologica Sinica 3, 901–913.
Lohani, A.K., Goel, N.K., Bhatia, K.K.S., 2014. Improving real time flood forecasting using
fuzzy inference system. J. Hydrol. 509, 25–41.
Lowry, O.H., Rosebrough, N.J., Farr, A.L., Randall, R.J., 1951. Protein measurement with
the Folin phenol reagent. J. Biol. Chem. 193 (1), 265–275.
Meyer, D., Dimitriadou, E., Hornik, K., Weingessel, A., Leisch, F., 2017. e1071: Misc.
Functions of the Department of Statistics, Probability Theory Group (Formerly:
E1071), TU Wien. R package version 1.6-8. https://CRAN.R-project.org/package=
e1071.
Mukhlisin, M., El-Shafie, A., Taha, M.R., 2012. Regularized versus non-regularized neural
network model for prediction of saturated soil-water content on weathered granite
soil formation. Neural Comput. Appl. 21, 543–553.
165
Computers and Electronics in Agriculture 147 (2018) 158–165
Olsen, S.R., 1982. Phosphorus. Methods of Soil Analysis 403–430.
Paz, C.G., Rodríguez, T.T., Behan‐Pelletier, V.M., Hill, S.B., Vidal‐Torrado, P., Cooper, M.,
2016. Encyclopedia Soil Sci.
Penrose, D.M., Glick, B.R., 2003. Methods for isolating and characterizing ACC deami-
nase-containing plant growth-promoting rhizobacteria. Physiol. Plant. 118 (1),
10–15.
R Core Team (2017). R: A language and environment for statistical computing. R
Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.
org.
Riza, L.S., Bergmeir, C., Herrera, F., Ben'itez, J.M., 2015. Frbs: Fuzzy rule-based systems
for classification and regression in R. J. Stat. Softw. 65 (6), 1–30. URL. http://www.
jstatsoft.org/v65/i06.
Schoenholtz, S.H., Van Miegroet, H., Burger, J.A., 2000. A review of chemical and phy-
sical properties as indicators of forest soil quality: challenges and opportunities. For.
Ecol. Manage. 138, 335–356.
Shiri, J., Keshavarzi, A., Kisi, O., Iturraran-Viveros, U., Bagherzadeh, A., Mousavi, R.,
Karimi, S., 2017a. Modeling soil cation exchange capacity using soil parameters:
assessing the heuristic models. Comput. Electron. Agric. 135, 242–251.
Shiri, J., Keshavarzi, A., Kisi, O., Karimi, S., 2017b. Using soil easily measured parameters
for estimating soil water capacity: soft computing approaches. Comput. Electron.
Agric. 141, 327–339.
Sirsat, M.S., Cernadas, E., Fernández-Delgado, M., Khan, R., 2017. Classification of
agricultural soil parameters in India. Comput. Electron. Agric. 135, 269–279.
Steel, R.G., Torrie, J., 1960. Principles and Procedures of Statistics. McGraw-Hill Book
Company, New York.
Taghavifar, H., Mardani, A., 2014. Use of artificial neural networks for estimation of
agricultural wheel traction force in soil bin. Neural Comput. Appl. 24, 1249–1258.
Tajik, S., Ayoubi, S., Nourbakhsh, F., 2012. Prediction of soil enzymes activity by digital
terrain analysis: comparing artificial neural network and multiple linear regression
models. Environ. Eng. Sci. 29, 798–806.
Theodoridis, S., Koutroumbas, K., 2006. Pattern Recognition, 2nd edition. Academic
Press, San Diego.
Van Bruggen, A.H.C., Semenov, A.M., 2000. In search of biological indicators for soil
health and disease suppression. Appl. Soil Ecol. 15, 13–24.
Venables, W.N., Ripley, B.D., 2002. Modern Applied Statistics With S, Fourth Edition.
Springer, New York.
Wang, L.X., Mendel, J.M., 1992. Generating fuzzy rules by learning from examples. IEEE
Transact. Syst., Man, Cybernet. 22 (6), 1414–1427.
Winding, A., Hund-Rinke, K., Rutgers, M., 2005. The use of microorganisms in ecological
soil classification and assessment concepts. Ecotoxicol. Environ. Saf. 62, 230–248.
Wu, W., Li, A.D., He, X.H., Ma, R., Liu, H.B., Lv, J.K., 2018. A comparison of support
vector machines, artificial neural network and classification tree for identifying soil
texture classes in southwest China. Comput. Electron. Agric. 144, 86–93.
Yang, X., Yuan, J., Yuan, J., Mao, H., 2010. An improved WM method based on PSO for
electric load forecasting. Expert Syst. Appl. 37 (12), 8036–8041.
Yazdani-Chamzini, A., Razani, M., Yakhchali, S.H., Zavadskas, E.K., Turskis, Z., 2013.
Developing a fuzzy model based on subtractive clustering for road header perfor-
mance prediction. Autom. Constr. 35, 111–120.
Zhang, J.H., Xia, J.J., Garibaldi, J.M., Groumpos, P.P., Wang, R.B., 2017a. Modeling and
control of operator functional state in a unified framework of fuzzy inference petri
nets. Comput. Methods Programs Biomed. 144, 147–163.
Zhang, L., Zuo, W., Zhang, D., 2016. LSDT: Latent sparse domain transfer learning for
visual adaptation. IEEE Trans. Image Process. 25 (3), 1177–1191.
Zhang, L., Yang, J., Zhang, D., 2017b. Domain class consistency based transfer learning
for image classification across domains. Inf. Sci. 418, 242–257.
Zhang, L., Zhang, D., 2016. Robust visual knowledge transfer via extreme learning ma-
chine-based domain adaptation. IEEE Trans. Image Process. 25 (10), 4959–4973.
Zhang, L., Liu, Y., Deng, P., 2017c. Odor recognition in multiple E-nose systems with
cross-domain discriminative subspace learning. IEEE Trans. Instrum. Meas. 66 (7),
1679–1692.
Žifčáková, L., Větrovský, T., Howe, A., Baldrian, P., 2016. Microbial activity in forest soil
reflects the changes in ecosystem properties between summer and winter. Environ.
Microbiol. 18, 288–301.