See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/283286316

Prediction of soil type and standard penetration test (SPT) value in Khulna
City, Bangladesh using general regression neural network

Article  in  Quarterly Journal of Engineering Geology and Hydrogeology · October 2015

DOI: 10.1144/qjegh2014-108

CITATIONS READS

4 695

4 authors, including:

Grytan Sarkar Sumi Siddiqua

Khulna University of Engineering and Technology University of British Columbia - Okanagan
14 PUBLICATIONS   50 CITATIONS    57 PUBLICATIONS   204 CITATIONS   

SEE PROFILE SEE PROFILE

Md Rokonuzzaman
Khulna University of Engineering and Technology
5 PUBLICATIONS   12 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Experimental validation of a numerical model for the interaction of dip-slip normal fault ruptures, sand deposits, and raft foundations View project

Acid rock drainage from oil sands tailings View project

All content following this page was uploaded by Grytan Sarkar on 16 April 2016.

The user has requested enhancement of the downloaded file.

XXX10.1144/qjegh2014-108G. Sarkar et al.Soil type prediction and SPT value, Khulna, Bangladesh
2015

2014-108 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
research-articleResearch article10.1144/qjegh2014-108Prediction of soil type and standard penetration test (SPT) value in Khulna City, Bangladesh using general regression neural networkGrytan Sarkar, Sumi Siddiqua, Rajib Banik &, Md. Rokonuzzaman

Research article Quarterly Journal of Engineering Geology and Hydrogeology

Published online October 23, 2015 doi:10.1144/qjegh2014-108 | Vol. 48 | 2015 | pp. 190­–203

Prediction of soil type and standard penetration test (SPT)

value in Khulna City, Bangladesh using general regression
neural network
Grytan Sarkar, Sumi Siddiqua*, Rajib Banik & Md. Rokonuzzaman
The University of British Columbia, School of Engineering, 3333 University Way, Kelowna, BC V1V1V7, Canada
* Correspondence: sumi.siddiqua@ubc.ca

Abstract:  In this study, a general regression neural network (GRNN) is developed for predicting soil type and standard pen-
etration test (SPT) N (standard penetration resistance) values based on SPT test results. It focuses on soils mainly in Khulna
City, Bangladesh that comprise fine-grained alluvium deposits of mostly silt and clay with some organic content and sand.
A detailed geological and geotechnical investigation of the city and its surroundings was conducted to generalize the subsoil
condition of the study area based on soil type and SPT values. The investigation results showed that the city is divided into
four geological units and three geotechnical zones. To develop the GRNN model, more than 2326 field SPT values (N) have
been collected from 42 clusters containing 143 boreholes spread over an area of 37 km2. Two trained models were developed:
initially the borehole locations were trained with the soil types and after that the borehole location-soil types were trained
with the Nc values. The model prediction was compared with the borehole data and the results showed that the GRNN model
predicts well compared with the actual site investigation data. Therefore, this model can be used for future planning and
expansion of the city.
Received 05 January 2015; revised 11 August 2015; accepted 13 August 2015

Terzaghi (1996) stated that ‘The behavior of a soil and rock in the most of the region of this area contains fine-grained soil with some
field depends not only on the significant properties of the individ- organic deposits, which, having lower bearing capacity, are not
ual constituents of the soil mass, but also on those properties that good for shallow foundations. However, Khulna City is progress-
are due to the arrangement of the particles within the mass’. Also, ing with several development projects including construction of
soil does not always show homogeneous and isotropic properties high-rise buildings, oil storage tanks, long span bridges, harbours,
comparable with other engineering materials such as steel, con- port structures, flood protection embankments and barrages.
crete and timber. Generally, soil properties vary from place to Because all structural loads are transferred to the soil, the soil char-
place with greater uncertainty owing to the complexity of soil for- acteristics play a vital role in the safety of these structures.
mation (Jaksa 1995). The complex geotechnical behaviour and Therefore, this generalized geological and geotechnical investiga-
spatial variability of soil are a challenge for a simplified geotechni- tion will provide a preliminary idea to the planner during planning
cal model. An alternative approach using artificial neural networks of the city and to the engineer for selecting types of foundations.
(ANN) is developed based on field data to determine the structure This paper draws on data from Consultancy Research and
and parameters of the model. The process is suitable to model Testing Services (CRTS), Khulna to present a generalized descrip-
complex problems where the relationships between the model tion and potential model for the subsoil conditions of the Khulna
variables and parameters are unknown (Hubick & Hubick 1992). City area in Bangladesh. CRTS is an organization that has under-
This technique is extensively used in the field of geotechnical taken an investigation of the subsurface soil conditions in several
engineering for site characterization (Kurup & Griffin 2006), prop- locations of this region. To analyse the data, site investigation
erties of geomaterials (Goh 1995; Ozer et al. 2008; Park & Kim reports from 42 clusters (containing more than one borehole) over
2011), bearing capacity of piles (Kurup & Griffin 2006), liquefac- a 37 km2 area throughout the city were collected. These 42 clusters
tion (Agrawal et al. 1997; Ural & Saka 1998; Goh 2002; Kim & contain 143 boreholes with the boreholes located close to each
Kim 2006) and slope stability (Ni et al. 1996; Neaupane & Achet other in each cluster (about 5 m apart in each). The latitude and
2004; Zhao 2008). longitude of the 143 boreholes were used to plot borehole locations
In this work, a detailed study of the geological and geomorpho- within the existing Khulna City corporation map. SPT values col-
logical behaviour of Khulna City and its environs was conducted. lected from these boreholes were corrected based on the water
The city is situated in the southern region of Bangladesh bounded table, overburden pressure and types of sampler. The corrected
by rivers. The soils formed in these areas are mainly alluvial standard penetration resistance (Nc) values are widely used for
deposits. Geologically, the area can be grouped into four units and field testing to determine subsoil bearing capacity, settlement and
soils in this region are mostly composed of silt, clay, sand and liquefaction potential, and to characterize the soil profiles. Nc val-
organic soil with peat. For a particular depth, soil found in the ues are also correlated with many other geotechnical properties
region differs from that at other locations and therefore the study such as angle of internal friction, shear wave velocity and cone tip
area was further divided into three geotechnical zones. The gener- resistance (Samui & Sitharam 2010).
alized investigation results showed that the soils in the northern In the current study, an ANN, explicitly the general regression
and southern part are mainly composed of clay with silt and neural network (GRNN) model, was developed to predict the
organic deposits, the middle eastern part comprises sand with soil profile and Nc value over the investigated area. Using the GRNN
some clay, and a intermediate type of soil is found in the middle and the data from the boreholes, information about the soil composi-
western part, which is mainly silt with clay and sand. It is seen that tion can be obtained within an acceptable engineering approxima-

© 2015 The Author(s). Published by The Geological Society of London. All rights reserved. For permissions: http://www.geolsoc.org.uk/permissions.
Publishing disclaimer: www.geolsoc.org.uk/pub_ethics
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 191

tion. In addition, the Nc values of subsoil of the city can be obtained ary between the Indian and Eurasian plates, whereas the southern and
from the analysis. The model consists of two sets of data: one for southwestern parts of the basin including Khulna City are not prone
training (80% of total boreholes in 28 clusters) and the other for test- to large magnitude earthquakes (HBRI-BSTI 1993). Bangladesh is
ing (the other unseen 20% of boreholes from 14 clusters). Although divided into three seismic zones: Zone 1, Zone 2 and Zone 3, where
the boreholes in each cluster are located very close together, the peak horizontal ground accelerations are 0.25, 0.15 and 0.075 g,
properties of soil in each cluster were not exactly same. Therefore, respectively (HBRI-BSTI 1993). Khulna City and its surrounding
instead of one borehole from each cluster all boreholes were consid- area is located in Zone 3 where peak ground acceleration is 0.075 g.
ered in training. However, for testing purpose, one borehole from
each cluster was selected. The testing boreholes were selected ran- Geotechnical units of the study area
domly from within the study area. Additionally, three boreholes
were selected at the outer periphery of the training boreholes to The study area has been divided into four geotechnical units based
observe the prediction capability, when the boreholes are located at on the physical properties of subsurface soils of the city encoun-
almost the outside range of the training boreholes. Two intercon- tered in the boreholes to a depth of about 45 m (Figs 3 and 4): unit
nected trained models were developed: initially the borehole loca- 1, sandy soil; unit 2, silty soil; unit 3, clayey soil; unit 4, organic
tions were trained with the soil types and after that the borehole soil and peat. Unit 1 is mainly composed of sand with silt having
location-soil types were trained with the Nc values. The trained Nc values of 6–47. Unit 2 comprises silt with clay having Nc values
model was then tested for the selected boreholes described above. of 1–15. Unit 3 consists of clay with silt having Nc of 1–17. Unit 4
The main advantage of this model is its ability to determine the soil includes organic clay and peat having Nc values of 1–6.
type and Nc values by knowing the position of the boreholes from
the reference boreholes within the trained zone. The main drawback Geotechnical zones of the study area
of this model is that it cannot predict soil information outside the
trained zone. However, this model can be used for small-scale pro- Khulna City is divided into three geotechnical zones based on the
jects and provides preliminary information about the subsoil condi- horizontal and vertical variation of the geological units and geo-
tions, which assists professionals to predict bearing capacity, suitable technical properties of the soil (Fig. 5). The geotechnical zone 1
foundation layer and stability of the soil at a given depth. covers the middle eastern part of the city along the bank of Rupsha
River. The subsurface soil within this zone mainly comprises sand
Geology of Khulna City and silts with some clay and organic deposit. The soil from 2 to
30 m below the ground surface is suitable for both shallow and deep
Khulna is the third largest metropolitan and second port city in foundations, where the Nc varies from 9 to 47. The geotechnical
Bangladesh. It is located in the southwestern region of the country, zone 2 occupies the middle western part of the city along the bank
which is bounded by latitudes 22°46’0"–22°58’0"N and longi- of the Mayur River. The subsurface soil below 10 m from the
tudes 89°28’0"–89°37’0"E (Fig. 1). The city has an elevation ground surface mainly consists of fine to dense sand and is suitable
above mean sea level of 9 m in the north to 2 m towards the SW. It for deep foundations, where Nc varies from 8 to 25. Soil in the
covers an area of 37 km2; however, the city and its outskirts are northern and southern part (geotechnical zone 3) of the city mainly
expanding continuously owing to rapid urbanization. The popula-
comprises soft clay to stiff clay with some silt and organic soil,
tion of Khulna City is over 1.5 million.
where Nc varies from 1 to 12. Soils in this zone are not suitable for
Geomorphologically, the city area is characterized by the
shallow or deep foundation without treatment and/or improvement.
Ganges tidal floodplains in the south and deltaic plain in the north
with low relief and crisscrossed by rivers (Roy et al. 2005). The
Methods
city is surrounded by four tidal rivers: the Rupsha, Bhairab, Mayur
and Hatia. It is bounded by the Rupsha in the SE and by the Bhairab Data collection and GRNN model
in the NE (Fig. 2). The Mayur and Hatia flow from NW to SW
along the western boundary of the city. Several marshes and Forty-two subsoil investigation reports containing data from
swamps are also located in and around the city. The subsurface of 143 boreholes in different parts of the city area were collected
the city area consists of alluvium to a depth of 40 m or more, com- from CRTS. Most of these boreholes were drilled to a depth of
posed of sand, silt, clay, organic clay and peat in varied proportion. 15.25–21.35 m and a few were drilled to a depth of 30.5 and 45 m.
Alam (1990) also mentioned that the surface geology of the city Drilling was carried out using the wash boring method where the N
consists of deltaic deposits, which are composed of tidal deltaic values were recorded at 1.52 m intervals and soil samples were
deposits, deltaic silt deposits and mangrove swamp deposits. collected. The field N values were corrected for overburden pres-
sure (CN) using the Peck et al. (1974) equation before being used
Seismotectonics of the region in the GRNN model:
Tectonically, the city area is situated in the western part of the N c = NCN (1)
Faridpur Trough of the Bengal Foredeep, which is a part of the
Bengal Basin, one of the largest sedimentary basins of the world where the correction factor for CN was calculated using the equation
(Alam 1990). The trough is filled with Tertiary and Quaternary sand-
and clay-rich sediments with a few coarse-grained sand beds. CN = 0.77 log10 ( 2000 / p′ ) (2)
Bangladesh covers most of the Bengal Basin, which is located in the where p’ is the effective overburden pressure (kN m−2)
calculated by
eastern part of the Indian Plate. The northward collision between the determining the effective unit weight of soil (γ’). In this study the
Indian Plate and Eurasian Plate created the Himalayan Ranges in the effective unit weight of sand was 17.20 kN m−3 and for the other types
north, the Indo-Burman Ranges in the east and the Bengal Basin in of soil a value of 15.75 kN m−3 was used. The negative pore pressure
the south (Curray & Moore 1974; Alam 1989). The Bengal Basin is
that developed during driving of the SPT sampler into saturated sand
surrounded by the Precambrian Shillong Massif in the north, the
and silts may result in higher shearing resistance as well as SPT blow
Tertiary sedimentary folded belts in the east, the Indian Shield in the
count. Therefore, Nc values greater than 15 obtained from equation
west and it is open to the Bay of Bengal in the south. The northern,
(1) were adjusted using the equation of Terzaghi & Peck (1968):
eastern and southeastern parts of the Bengal Basin (Bangladesh and
NE India) are seismically active zones located close the plate bound- N c′ = 15 + 0.5 ( N c − 15 ) . (3)
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
192 G. Sarkar et al.

Fig. 1. Location map of Khulna City (source: Banglapedia, National Encyclopedia of Bangladesh, 2011).
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 193

Fig. 2. Geology map of Khulna City

(modified from Roy et al. 2005).

To assign the coordinate of the borehole locations, a reference example, regression models work well with one response but not
point with coordinates (0,0) was selected from which all the dis- for multiple responses. In addition, most of the models are devel-
tances were measured. Using latitudes and longitudes, borehole oped by considering linear behaviour of the parameters. Kurup &
site locations were identified on the city map (Fig. 3). To identify Griffin (2006) stated that the existing traditional experience-based
the training and testing boreholes, two different symbols are used systems and non-traditional mathematical-based (statistical and
in Figure 3. fuzzy subset theory) systems emphasize only the soil behaviour
rather than the actual soil type and composition. Therefore, many
Overview of GRNN researchers are now interested in using ANN in the field of geo-
technical engineering because this approach is general, flexible,
In recent years, artificial neural networks (ANN) have been widely robust, versatile and does not require a physical model to start the
used in various fields of science and engineering. They also have process (Goh 1999; Penumadu & Zhao 1999; Itani & Najjar 2000;
been used for a great variety of subjects such as selection of equiv- Juang et al. 2001; Kurup & Dudani 2002). Moreover, it has the
alent vehicle axle load, prediction of potential earthquake hazard, ability to approximate many arbitrary functions, tolerate the pres-
project planning and construction modelling (Nouri 2004). Several ence of chaotic components, classify patterns, estimate continuous
empirical, numerical and statistical methods are used in computing variables and forecast time series.
geotechnical problems, but owing to the spatial nature of soil and ANN models are extensively used for predicting unknown param-
its uncertainties these models have numerous limitations. For eters for a given set of inputs and they show better performance for
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
194 G. Sarkar et al.

Fig. 3. Borehole locations of Khulna City

Corporation.

analysing incomplete data (Patterson 1998). In the ANN model, a of the inputs and transfers it to a linear or nonlinear active function for
large number of simple processors known as neurons are intercon- generating output (Kurup & Griffin 2006). Generalized regression
nected according to the relative importance of the inputs as well as neural networks (GRNN) and probabilistic neural networks (PNN)
the weight of each input. Among all the available ANN techniques, are the two basic types of radial basis networks. GRNN accomplishes
the back propagation algorithm is widely used in the field of geotech- regression and the target variables are continuous, whereas PNN
nical engineering (Kurup & Griffin 2006; Trivedi et al. 2014). works well for classification and the target variable is categorical.
However, this method takes a significant time to train, which involves GRNN has several advantages, especially when compared with
computing the optimal weight and convergence to desired solutions, the more widely used back propagation neural network algorithm.
owing to a complex iteration process. This problem can be resolved The back propagation is performed in a repeatable manner and
through radial basis networks (Kurup & Griffin 2006; Specht 1991). often gives solutions with a deviation between the actual and pre-
A radial basis network is a modified neural network, which requires dicted results. However, the computations of GRNN are not repet-
more neurons than standard feed-forward back propagation net- itive and show better correlation of results with the actual values
works. The training time mainly depends on the time required for (Specht 1991; Wasserman 1993). The advantages of GRNN can be
loading the training matrix (Demuth et al. 2008). After receiving an summarized as follows: (1) simple and fast training; (2) independ-
input, a neuron processes the input by computing the weighted sum ence from initial condition; (3) nonlinear interpolation based on a
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 195

Fig. 4. Geotechnical cross-sections of the study area.

sound mathematical foundation; (4) training time required only for  − m ( x − x )2 

∑ i ij 
loading training matrix; (5) it predicts very well for a large quan-
∑ j = 1 j  i = 12σ 2
n
y exp
tity of data because the estimate surface converges to the optimal 
y^ =  .
regression surface for increasing sample size.
 − m ( x − x )2 
The GRNN also has some limitations. The main disadvantage ∑ i ij  (4)
∑ j = 1 exp  i =12σ 2
n

of this neural network is that it requires a substantial amount of 

 
computation to evaluate new points, but this problem can be
overcome by clustering the raw data as described by Kurup &
Griffin (2006). The training data need the full range of expected
The value of (xi – xij) represents the distance between the train-
output values for complete analysis. The drawbacks of the
ing sample xij and the point of prediction xi of the ith variable. The
GRNN model are as follows: (1) smoothing; that is, it underesti-
prediction performance mainly depends on the summation of
mates the output values when testing data fall far away from the
(xi – xij). For example, the model predicts well when the summation
peak of the output surface and it overestimates when the data fall
of (xi – xij) is small but the opposite behaviour is found in the case
near the peak of the output surface (Kurup & Griffin 2006); (2)
of larger value. This is because the weight, that is,
results mainly depend on the density of the monitoring data; that
is, it shows better results when the samples are evenly distrib-  −∑ m ( x − x ) 2 
exp  
i =1 i ij
uted; (3) it cannot estimate a point outside the range of observed
samples.  2σ 2
 of the corresponding target value yj of
 
the jth training case is larger for a small value of (xi – xij), whereas
GRNN structure for a larger value the resultant weight is small. The standard devia-
The GRNN and PNN can be designed using the radial functions tion, known as the smoothness parameter (σ), determines the accu-
newgrnn in MATLAB (MathWorks, Natick, MA). GRNN is a racy of the function obtained from training data through the
multilayer feed-forward neural network, which uses a normally GRNN. The value of σ should be selected according to the nature
distributed probability density function of the observed data for of the data; if the data are irregular, a large value of σ helps to
performing regression. Considering m number of output variables, eliminate any wild points and creates a better regression during
and n number of training cases, the predicted value of output (ŷ) training. On the other hand, a smaller value of σ provides an accu-
can be determined from the GRNN model as follows (Specht rate fit regression model for regular set of training data as described
1991): by Kurup & Griffin (2006).
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
196 G. Sarkar et al.

Fig. 6. Structure of a typical GRNN model to predict soil type.

Fig. 5. Map showing geotechnical zones of Khulna City.

The structures of GRNN for predicting soil types and Nc val-

ues are shown in Figures 6 and 7, respectively. The structures of
these models consist of a large number of nodes (neurons),
Fig. 7. Typical GRNN model structure for Nc value.
which are arranged in four layers: an input layer, a pattern layer,
a summation layer and an output layer. The input layer consists
of input variables where each input variable represents one neu- ing Nc values. The first network (Fig. 6) contains three input neu-
ron in the GRNN structure, the pattern layer represents training rons corresponding to the three input variables X, Y and Z, where
of the input parameters using the probability density function X, Y and Z represent the coordinates of boreholes from the refer-
 −∑ m ( x − x ) 2  ence point in the three geometric directions. These three input neu-
exp   . The summation layer is divided into
i =1 i ij
rons were connected with the pattern neurons with a connection
 2σ 2
 weight equal to the value of the three input neurons in each training
  case. However, the number of pattern neurons represents the num-
two parts: an A-summation layer and a B-summation layer. The ber of training cases. The pattern neurons are also connected with
A-summation layer represents the sum of weighted responses of the A-summation neurons with a connection weight equal to the
the pattern layer (numerator of equation (4)) and the B-summation value of the target parameters. Both the A-summation layer and
layer represents unweighted outputs of pattern neurons (dominator output layer contain five neurons corresponding to the five target
of equation (4)). Moreover, the A-summation layer contains one parameters (% sand, % silt, % clay, % organic and % peat).
neuron for each target parameter, whereas the B-summation layer Therefore, both the pattern layer and summation layer were used
contains only one neuron for all target parameters. The output for training the incoming inputs (i.e. borehole coordinates) with
layer comprises simple division of the numerator and dominator of the target parameters (i.e. soil types) using a probability density
equation (4). The output layer is also called the decision layer, function to obtain desired outputs from a given input set of data.
which yields the predicted value to an unknown input variable. The second model is an extension of the first because the output
Similar to the A-summation layer, the output layer also contains parameters (X, Y, Z, % sand, % silt, % clay, % organic and % peat)
one neuron for each target parameter. of the first model were used as input neurons for the second model.
In this study, two interconnected networks were proposed for The input neurons were processed though a pattern layer and a
estimating the soil type and Nc values. The first network was summation layer using a probability density function to obtain the
developed for predicting soil types and the second was for predict- desired output. This model comprises only two output parameters,
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 197

Table 1. Input and output parameters of the testing boreholes nos 16, 58 and 84

X (m) Y (m) Z (m) Soil type SPT value

  Clay (C) Sand (S) Silt (M) Organic (O) Peat  

Borehole no. 16 (Fulbarigate)  

1060 1260 0 1 0 0 0 0 0
1060 1260 1.50 1 0 0 0 0 5
1060 1260 3.00 1 0 0 0 0 7
1060 1260 4.60 1 0 0 0 0 5
1060 1260 6.00 0 0 0 1 0 3
1060 1260 7.60 0 0 0 1 0 5
1060 1260 9.15 0 0 0 1 0 7
1060 1260 10.70 1 0 0 0 0 5
1060 1260 12.20 1 0 0 0 0 4
1060 1260 13.70 0 0 0 1 0 4
1060 1260 15.25 0 0 0 1 0 7
Borehole no. 58 (Goalkhali)  
2978.9 6222.1 0 1 0 0 0 0 0
2978.9 6222.1 1.50 1 0 0 0 0 3
2978.9 6222.1 3.00 1 0 0 0 0 2
2978.9 6222.1 4.60 0 0 0 1 0 5
2978.9 6222.1 6.00 0 0 1 0 0 2
2978.9 6222.1 7.60 1 0 0 0 0 3
2978.9 6222.1 9.15 1 0 0 0 0 4
2978.9 6222.1 10.70 1 0 0 0 0 9
2978.9 6222.1 12.20 0 1 0 0 0 15
2978.9 6222.1 13.70 0 1 0 0 0 17
2978.9 6222.1 15.25 0 1 0 0 0 15
Borehole no. 84 (KDA)  
6510 10849 0 1 0 0 0 0 0
6510 10849 1.50 1 0 0 0 0 2
6510 10849 3.00 1 0 0 0 0 5
6510 10849 4.60 0 0 0 1 0 5
6510 10849 6.00 0 0 0 1 0 4
6510 10849 7.60 1 0 0 0 0 4
6510 10849 9.15 1 0 0 0 0 2
6510 10849 10.70 0 1 0 0 0 6
6510 10849 12.20 0 1 0 0 0 10
6510 10849 13.70 1 0 0 0 0 5
6510 10849 15.25 1 0 0 0 0 7
6510 10849 16.80 1 0 0 0 0 17
6510 10849 18.30 1 0 0 0 0 17

the depth increment (ΔZ) and Nc value. Consequently, the input variables (X, Y and Z). To estimate the soil type and Nc val-
A-summation layer and output layer contain only two neurons, ues, two algorithms, one for training and the other for testing, were
whereas the B-summation layer contains only one neuron. compiled.
The advantage of this algorithm is that one can estimate the
GRNN algorithms type of soil and Nc values of any point within the trained region by
loading the coordinates of boreholes only as an input. The coordi-
The newgrnn function was used to construct the GRNN network nates of the boreholes define the location of the boreholes and their
model using the neural network toolbox in MATLAB. For the first depth increments. The presence and absence of a soil type was
network model, coordinates of the boreholes and soil types were expressed in binary form (unity and zero). For example, when clay
the input and output parameters of the function and the spread con-
is dominant at a particular depth, it is expressed as unity and other
stant was taken as unity, so that neurons respond strongly to over-
soil types are expressed as zero. The input and output parameters for
lapping regions of the input space. The spread constant is a radial
the test boreholes 16, 58 and 84 are shown in Table 1. This table
basis function constant that is used to fit the model appropriately.
shows that, for a particular borehole (no. 16, Fulbarigate), the coor-
A higher value of spread constant makes the model over-fit and a
dinates of the borehole are expressed in terms of X, Y and Z from the
lower value makes it under-fit (Zhao 2008). The output parameters
reference point and the soil type is expressed in terms of numerical
of the first network such as borehole coordinates (X, Y and Z) and
value (zero and unity). The soil in this region mainly consists of clay
soil types (% sand, % silt, % clay, % organic and peat) were con-
and organic matter to a depth of 15.25 m having lower Nc values.
sidered as input variables in the second network. These input vari-
ables and the output parameters (ΔZ and Nc values) were compiled
Results and discussion
in another newgrnn function to train them using a probability den-
sity function. Following this, the network was able to predict the Regression plots were used to display the network outputs with
soil type and Nc values for the testing boreholes from a given set of respect to targets for both training and testing data. It was found
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
198 G. Sarkar et al.

Fig. 8. Comparison of the estimated output and targeted field values from a GRNN model, a result of all training boreholes. (a) Soil profile; (b) SPT
values.

that GRNN performed better than a typical ANN model (Koekkoek value from the GRNN model. In the GRNN model the soil is clas-
& Booltink 1999). According to the results of the network training, sified qualitatively rather than quantitatively because the soil clas-
the network successfully captured the relationship between the sification is qualitative in nature (Kurup & Griffin 2006). Based on
input parameters and outputs. the value of the coefficient of correlation (R) (R > 0.99) it can be
Figure 8a and b shows the comparison between the targeted concluded that the prediction of soil composition was accurate for
(field data) and the GRNN predicted values resulting from all all types of soil.
training boreholes. These results are derived from the network The model was then extended for the prediction of Nc values at a
structure where inputs are coordinates (depth or distance) and out- fixed depth increment of ΔZ = 1.52 m. The networks were trained to
puts are obtained as different matrices (soil type and Nc value) in associate the Nc recorded in the field with the predicted values from
an Excel spreadsheet. Figure 8a shows the graph of the GRNN the GRNN model. The performance plot of the model predicting Nc
model for testing soil composition at different locations and vary- values for the same locations and depths as described for soil com-
ing depths. The network was trained to associate the soil composi- position is illustrated in Figure 8b. The prediction of Nc values also
tion obtained from the particle size distribution with the predicted showed very good correlation with the measured Nc values.
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 199

Fig. 9. Comparison between actual and predicted results for soil composition.

To verify the results, the soil types obtained from the GRNN- captured a good representation of the field soil composition.
predicted model were compared with the tested boreholes as Test borehole BH-107 represents the sounding location at
shown in Figure 9. The estimated soil types from the GRNN Farazipara; the data comprise two different silt layers at the top
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
200 G. Sarkar et al.

D
%+ %+ %+
1FYDOXH 1FYDOXH 1FYDOXH
                 

)LHOG )LHOG )LHOG
*511 *511 *511



'HSWKP













E
%+ %+ %+
1FYDOXH 1FYDOXH 1FYDOXH
                   


)LHOG )LHOG )LHOG

 *511 *511 *511


'HSWKP








Fig. 10. Comparison of Nc value with
depth from actual Nc data and GRNN

predicted data: (a) for boreholes BH-6,
BH-28 and BH-34; (b) for boreholes BH-
 45, BH-72 and BH-78;

and bottom with a middle organic deposit and, interestingly, the file. Similarly, the clay layer was not found in the predicted
GRNN model predicted a similar soil profile. A similar success model for both BH-100 and BH-120. However, the only varia-
rate was observed for the sounding boreholes BH-78 (KMC), tion in the estimated data for BH-118 was at a depth of 4.5–
BH-84 (KDA), BH-100 (PRF Police Line), BH-118 (Sir Iqbal 7.5 m, where an unexpected organic layer was found. Moreover,
Road) and BH-120 (Labanchara). The only variation in the esti- boreholes BH-34 (Tarer Pukur), BH-45 (Boyra) and BH-114
mated profiles for these boreholes is due to a small missing (Ahashan Ahmed Road) showed more inconsistency in predic-
layer. For example, BH-78 and BH-84 contain mostly clay with tion, where the original soil layers obtained from the field test
a thin sand and organic clay layer. The GRNN model also are thin and numerous. For example, in BH-34, the field data
showed that the soils found in these boreholes are mainly clay showed that there are alternating layers of clay and sand with a
with sand but the organic layer was not in the predicted soil pro- silt and organic layer, but the silt and sand layer was missing in
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 201

d
%+ %+ %+
1FYDOXH 1FYDOXH 1FYDOXH
                  


)LHOG )LHOG )LHOG

 *511 *511 *511


'HSWKP










Fig. 10. (Continued) (c) for boreholes
BH-84, BH-90 and BH-100; (d) for
 boreholes BH-107, BH-114 and BH-118;

the GRNN model. Therefore the developed GRNN cannot pre- these three boreholes are located near the periphery of the train-
dict very well when the data showed more variations in soil lay- ing zone.
ers over a small depth. A comparison of both field and predicted Nc values with depth
The main inconsistency was observed in the testing bore- is presented in Figure 10. Although the prediction of soil type in
holes BH-16 (Fulbarigate), BH-28 (Daulatpur), BH-90 the borehole BH-16 was inconsistent, the prediction of Nc value is
(Sabujbag) and BH-143 (Shipyard). For example, borehole nearly consistent over depth. The only exception was noted at a
BH-16 comprises mainly clayey soil with organic deposits, depth of 12.5 m, where the model predicts an abruptly higher Nc
whereas the model p­ redicted silt with organic deposits. A simi- value than the maximum value of the entire depth. This type of
lar inconsistency in prediction of soil layers was observed for inconsistency is termed smoothing. On the other hand, the varia-
the other boreholes. The inconsistencies in prediction of soil tion of the field Nc with depth was found to zigzag in both BH-28
profiles in these boreholes are higher than for the others because and BH-34, whereas the model predicted a smoother profile with
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
202 G. Sarkar et al.

e
%+ %+
1FYDOXH 1FYDOXH
            

)LHOG )LHOG
*511 *511



'HSWKP










Fig. 10. (Continued) (e) for boreholes
BH-120 and BH-143.


depth. In this study, some inconsistency was often observed for (1) irregular variation of actual Nc values with depth, (2) location
prediction of Nc values when the measured data have a zigzag pat- of boreholes (boreholes near to the periphery of the trained zones
tern with depth. provided limited prediction) and (3i) smoothing.
Figure 10b presents Nc values for three test boreholes (BH-45,
BH-72 and BH-78), in which BH45 and BH-78 showed zigzag Conclusions
variations of Nc value from a depth of about 9 to 15 m, and BH-72
showed little variation in Nc. Interestingly, the prediction of Nc A GRNN model for predicting soil types and Nc values from exist-
value was found to be consistent for both smooth and zigzag pro- ing SPT data is presented in this study. The geological and geo-
files. The inconsistencies from smoothing continued, unchanged, technical investigation of the study site revealed that soil in this
even when the boreholes were from within the spatially distributed zone is formed mainly from alluvial deposits including four types:
training data. sand, silt, clay and organic with peat. Two interconnected training
Figure 10c shows that the model prediction of Nc for BH-90 is models were developed to predict soil profiles and Nc values. The
better than for BH-84 and BH-100, although the estimated Nc key advantage of this model is that one can estimate the unknown
­values for these two boreholes did not show considerable deviation soil types and Nc values using the coordinates of the location,
from the measured profile. The main inconsistency was found for which can help engineers, geologists and urban planners to obtain
these boreholes below a depth of 15 m, where the GRNN models a preliminary understanding of soil conditions during planning and
overestimate the minimum and underestimate the maximum Nc construction of new infrastructure. The results indicate that the
values. A similar behaviour was found for the testing boreholes developed GRNN model can predict well when the boreholes are
BH-114 and BH-118 as shown in Figure 10d, where the inconsist- within the range of training data. However, it failed to predict well
ency was found after a depth of 9 m, for similar reasons. Figure 10e when the soil profiles were variable. Based on the results of this
shows the estimated Nc values of BH-120 and BH-143, where pre- study, it is safe to state that GRNN can perform similarly to or
dicted data captured well the field Nc values for BH-120. The suc- sometimes better than conventional predicting methods in many
cess rate of prediction of these boreholes was relatively poor situations. It is possible to apply the algorithm to any location; the
compared with other locations. rate of success in model predictions may also increase in the case
Generally, the testing borehole results showed that the proposed of less variable geological profiles. In conclusion, within the scope
GRNN model predicts the Nc values better than it predicts soil pro- of the investigation, GRNN proved to be a powerful tool in com-
file. The main inconsistency in prediction was found to be due to puting reliable solutions for infrastructure construction problems.
 Downloaded from http://qjegh.lyellcollection.org/ at The City University on March 21, 2016
Soil type prediction and SPT value, Khulna, Bangladesh 203

Acknowledgements and Funding Kurup, P.U. & Dudani, N.K. 2002. Neural networks for profiling stress history
of clays from PCPT data. Journal of Geotechnical and Geoenvironmental
The authors would like to express their thanks to laboratory technicians and
Engineering, 128, 569–579, http://dx.doi.org/10.1061/(ASCE)1090-
CRTS of the Department of Civil Engineering, KUET, for their co-operation and
0241(2002)128:7(569).
tolerance in the course to this project. In addition, the authors wish to express
Kurup, P.U. & Griffin, E.P. 2006. Prediction of soil composition from CPT
their gratitude to the anonymous reviewers, whose suggestions and remarks
data using general regression neural network. Journal of Computing in
have greatly helped us to improve the quality of the paper.
Civil Engineering, 20, 281–289, http://dx.doi.org/10.1061/(ASCE)0887-
3801(2006)20:4(281).
Scientific editing by Tom Dijkstra; Joel Smethurst Neaupane, K.M. & Achet, S.H. 2004. Use of backpropagation neural network
for landslide monitoring: A case study in the higher Himalaya. Engineering
Geology, 74, 213–226, http://dx.doi.org/10.1016/j.enggeo.2004.03.010.
References Ni, S.H., Lu, P.C. & Juang, C.H. 1996. A fuzzy neural network approach to eval-
Agrawal, G., Chameau, J.L.A. & Bourdeau, P.L. 1997. Assessing the liquefac- uation of slope failure potential. Computer-Aided Civil and Infrastructure
tion susceptibility at a site based on information from penetration testing. Engineering, 11, 59–66, http://dx.doi.org/10.1016/0148-9062(96)81903-6.
In: Kartam, N., Flood, I. & Garret, J.H. (eds) Artificial Neural Networks for Nouri, Y., Choobbasti, A.J. & Kaveh, A. 2004. Soil Profile Prediction Using
Civil engineers: Fundamentals and Applications, American Society of Civil Artificial Neural Networks. Electronic Journal of Geotechnical Engineering,
Engineers, New York, 185–214. 9. http://www.ejge.com/2004/Ppr0462/Abs0462.htm
Alam, M. 1989. Geology and depositional history of Cenozoic sediments of Ozer, M., Isik, N.S. & Orhan, M. 2008. Statistical and neural network
the Bengal Basin of Bangladesh. Palaeogeography, Palaeoclimatology, assessment of the compression index of clay-bearing soils. Bulletin of
Palaeoecology, 69, 125–139, http://dx.doi.org/10.1016/0031-0182(89)90159-4. Engineering Geology and the Environment, 67, 537–545, http://dx.doi.
Alam, M. 1990. Bangladesh in World Regional Geology. Columbia University org/10.1007/s10064-008-0168-8.
Press, New York. Park, H.I. & Kim, Y.T. 2011. Prediction of strength of reinforced lightweight
Banglapedia. 2011. National Encyclopedia of Bangladesh. http://geospatialworld. soil using an Artificial Neural Network. Engineering Computations, 28,
net/Paper/Application/ArticleView.aspx?aid=1429#sthash.katdxv8S.dpuf. 600–615.
HBRI-BSTI. 1993. Bangladesh National Building Code. HBRI-BSTI, Dhaka. Patterson, D.W. 1998. Artificial Neural Networks: Theory and Applications.
Curray, J.R. & Moore, D.G. 1974. Sedimentary and tectonic processes in the Prentice Hall, Upper Saddle River, NJ.
Bengal deep-sea fan and geosyncline. In: Burk, C.A. & Drake, C.L. (eds) The Peck, R.B. Hanson, W.E. & Thorburn, T.H. 1974. Foundation engineering,
Geology of Continental Margins. Springer, Berlin, 617–627. John Wiley & Sons, New York.
Demuth, H., Beale, M. & Hagan, M. 2008. Neural Network Toolbox™ 6 User’s Penumadu, D. & Zhao, R. 1999. Triaxial compression behavior of sand and
Guide. MathWorks, Natick, MA. gravel using Artificial Neural Networks (ANN). Computers and Geotechnics,
Goh, A.T. 2002. Probabilistic neural network for evaluating seismic liquefac- 24, 207–230, http://dx.doi.org/dx.10.1016/S0266-352X(99)00002-6.
tion potential. Canadian Geotechnical Journal, 39, 219–232, http://dx.doi. Roy, M.K., Saha, S., Ahmed, S.S. & Mazumder, Q.H. 2005. Tide, morphol-
org/10.1139/t01-073. ogy, litho-facies, zonation and evolution of a middle Holocene to present
Goh, A.T.C. 1995. Closure to ‘Modeling Soil Correlations Using Neural estuary—Meghna, in South Central Bangladesh. Journal of the Geological
Networks’ by Anthony T. C. Goh. Journal of Computing in Civil Engineering, Society of India, 60, 354.
9, 275–278, http://dx.doi.org/10.1061/(ASCE)0887-3801(1997)11:1(79.2). Samui, P. & Sitharam, T.G. 2010. Site characterization model using Artificial
Goh, A.T.C. 1999. Soil laboratory data interpretation using generalized regres- Neural Network and Kriging. International Journal of Geomechanics, 10,
sion neural network. Civil Engineering and Environmental Systems, 16, 171–180, http://dx.doi.org/10.1061/(ASCE)1532-3641(2010)10:5(171).
175–195, http://dx.doi.org/10.1080/02630259908970261. Specht, D.F. 1991. A general regression neural network. IEEE Transactions on
Hubick, K.T. & Hubick, K.T. 1992. Artificial Neural Networks in Australia. Neural Networks, 2, 568–576, http://dx.doi.org/10.1109/72.97934.
Department of Industry, Technology and Commerce, Canberra, ACT. Terzaghi, K. 1996. Soil Mechanics in Engineering Practice. Wiley, New
Itani, O.M. & Najjar, Y.M. 2000. Three-dimensional modeling of spatial soil York.
properties via Artificial Neural Networks. Transportation Research Record: Terzaghi, K. & Peck, R.B. 1968. Soil Mechanics in Engineering Practice. 2nd
Journal of the Transportation Research Board, 1709, 50–59. edn. Wiley, New York.
Jaksa, M.B. 1995. The influence of spatial variability on the geotechnical Trivedi, R., Singh, T.N., Mudgal, K. & Gupta, N. 2014. Application of artifi-
design properties of a stiff, overconsolidated clay. PhD thesis, University of cial neural network for performance evaluation of vertical axis wind turbine
Adelaide, Adelaide, SA. rotor. International Journal of Ambient Energy, 3, 564–574, http://dx.doi.org
Juang, C.H., Jiang, T. & Christopher, R.A. 2001. Three-dimensional site charac- /10.1080/01430750.2014.915889.
terisation: Neural network approach. Géotechnique, 51, 799–809. Ural, D.N. & Saka, H. 1998. Liquefaction assessment by Artificial Neural
Kim, Y.-S. & Kim, B.-T. 2006. Use of artificial neural networks in the pre- Networks. Electronic Journal of Geotechnical Engineering, 3. http://www.
diction of liquefaction resistance of sands. Journal of Geotechnical and ejge.com/1998/Ppr9803/Abs9803.htm
Geoenvironmental Engineering, 132, 1502–1504, http://dx.doi.org/10.1061/ Wasserman, P.D. 1993. Advanced Methods in Neural Computing. Wiley, New
(ASCE)1090-0241(2006)132:11(1502). York.
Koekkoek, E.J.W. & Booltink, H. 1999. Neural network models to predict Zhao, H. 2008. Slope reliability analysis using a support vector machine.
soil water retention. European Journal of Soil Science, 50, 489–495, http:// Computers and Geotechnics, 35, 459–467, http://dx.doi.org/10.1016/j.​
dx.doi.org/10.1046/j.1365-2389.1999.00247.x. compgeo.2007.08.002.

View publication stats