Soil data clustering by using K-means and fuzzy K-means algorithm

Abstract

A problem of soil clustering based on the chemical characteristics of soil, and proper visual representation of the obtained results, is analysed in the paper. To that aim, K-means and fuzzy K-means algorithms are adapted for soil data clustering. A database of soil characteristics sampled in Montenegro is used for a comparative analysis of implemented algorithms. The procedure of setting proper values for control parameters of fuzzy K-means is illustrated on the used database. In addition, validation of clustering is made through visualisation. Classified soil data are presented on the static Google map and dynamic Open Street Map.

Full text

56 Telfor Journal, Vol. 8, No. 1, 2016.
Abstract — A problem of soil clustering based on the
chemical characteristics of soil, and proper visual
representation of the obtained results, is analysed in the paper.
To that aim, K-means and fuzzy K-means algorithms are
adapted for soil data clustering. A database of soil
characteristics sampled in Montenegro is used for a
comparative analysis of implemented algorithms. The
procedure of setting proper values for control parameters of
fuzzy K-means is illustrated on the used database. In addition,
validation of clustering is made through visualisation.
Classified soil data are presented on the static Google map and
dynamic Open Street Map.
Keywords — clustering, data mining, K-means, fuzzy K-
means, pedologic map.
I. INTRODUCTION
N Montenegro, in the period 1958-1988, a detailed soil
map with scale of 1:50 000 was made. Unfortunately, as
in other Former Republics of Yugoslavia, the enormous
effort and the work was not properly presented to the wide
professional community and land users, since the data and
data map were available only in a hard copy version. The
goal of this paper is to investigate the possibility of using
data mining tools for soil classification based on the
available data and to illustrate adequate visualisation, in
order to make this data understandable to wider society.
Data mining (DM) is a set of techniques that aims to
discover implicit useful information from big data.
Information discovery is usually performed by identifying
patterns and establishing relationships. Data Mining allows
focusing on the most important information in the data.
DM includes: clustering, anomaly detection, association
rule learning, classification, regression and summarization
and sequence or path analysis and forecasting. DM is the
computer-assisted process of "digging" through enormous
databases in order to analyse and extract the meaning of the
data.
In this paper focus is on clustering. Clustering has found
applications in many research areas such as mathematics,
engineering, economics, marketing, machine learning,
pattern recognition, genetics, bioinformatics, psychology,
biology, data compression and information retrieval.
Clustering is a process of grouping similar sets of data. This
grouping is unsupervised; it is done without using known
structures in the data. Clustering aims to make clusters with
data samples which are more similar to each other than to
data samples that belong to the other clusters. Each cluster
is defined by a central point, a centroid. Similarity of data
in one cluster is measured by using different criteria. Thus,
there are lots of different methods which can solve the
general task of clustering [1].
Two types of K-means algorithm are analysed in this
paper and the obtained results are discussed. The standard
K-means algorithm divides a data sample into exclusive
clusters. The initial values of clusters’ centroids are
randomly selected from the available data. Updating
centroids and clustering of data is then repeated until
convergence is reached or for a defined number of
iterations. A new centroid for a cluster is calculated based
on each data sample that belongs to that cluster. The first
issue of application of K-means-type algorithms is that the
number of clusters should be known in advance. Thus,
before discovering knowledge from big data, we have to
know how many cluster we expect in a database. The
second issue is that this kind of algorithms is very sensitive
to the initial clusters’ centroids. Usually, initial centroids
are chosen randomly.
In fuzzy K-means clustering data samples belong to
every estimated cluster, with a certain belonging degree.
Hence, a result of this algorithm are not exclusive clusters,
but clusters with fuzzy borders. A fuzzifier is a parameter
which defines fuzziness of fuzzy K-means clusters. This
method is used for clustering data in cases where clusters
are not clearly defined and one cannot estimate clear
borders among data samples.
In this paper, standard and fuzzy K-means clustering of
soil data is implemented for a database and results are
graphically presented. The results of fuzzy K-means
clustering for different values of fuzzifier are tested, and a
procedure for the selection of fuzzifier for a database is
proposed. In addition to simple graphical representation,
maps are a proper way of presenting the results of clustering
of soil. Thus, the results of K-means clustering of soil are
present on the static Google map and dynamic Open Street
Map.
Soil data clustering by using K-means and fuzzy
K-means algorithm
Elma Hot and Vesna Popović-Bugarin, member, IEEE
I
Paper received May 24, 2016; accepted June 12, 2016. Date o
f
publication July 20, 2016. The associate editor coordinating the review o
f
this manuscript and approving it for publication was Prof. Miroslav
Lutovac.
This paper is a revised and expanded version of the paper presented a
t
the 23rd Telecommunications Forum TELFOR 2015 [21].
This research is supported in part by FP7 Project Fore Mont and Pro ject
for establishment of pilot Montenegrin Centre of Excellence in Bio-
informatics - BIO-ICT (Contract No. 01-1001).
Elma Hot, Faculty of Electrical Engineering Podgorica, University o
f
Montenegro, Džordža Vašingtona bb, 81000 Podgorica (Phone: +382
20 245 839, Fax: +382 20 245 873, e-mail: elma_hot@live.com).
Vesna Popović-Bugarin, Faculty of Electrical Engineering Podgorica,
University of Montenegro, Džordža Vašingtona bb, 81000 Podgoric
a
(Phone:+382 20 245 839, Fax: +382 20 245 873, e-mail: pvesna@ac.me).

Hot and Popović-Bugarin: Soil data clustering by using K-means and fuzzy K-means algorithm 57
The paper is organized as follows. In Section II K-means
clustering is presented; in Section III fuzzy K-means
algorithm is reviewed. The performances of analysed
algorithms in the case of clustering data samples with
chemical characteristic of soil are analysed in Section IV
through simulation results. A conclusion is drawn in
Section V.
II. K-MEANS CLUSTERING
K-means (KM) clustering is a widely used partitioning
method. This method aims to make K mutually exclusive
clusters of n data samples characterised by d parameters.
Each cluster K is defined with one central point (centroid)
determined by a certain combination of parameters
contained in each data sample.
KM is known as a method of vector quantization, since it is
based on the location of points and their mutual distances
[2]. Namely, data samples described by d parameters can be
presented as points in a d-dimensional space, where their
coordinates are determined by the values of d parameters.
A data sample belongs to a cluster defined with a centroid
which is the closest one to the considered sample (point).
The closest centroid is chosen after calculating the distance
of each data from each centroid. Each data sample can
belong to exactly one cluster. Hence, KM clustering is also
called hard clustering.
A. K-means clustering algorithm
The KM algorithm aims to distribute a set X of n data
samples into K clusters. Each data sample is defined by d
parameters. We consider data samples as points in a d-
dimensional space for better visualization. Input to the
algorithm is the number of clusters K. The initial values of
centroids 11 1
12
, ,...,
K
cc c, d
i
cR, are chosen randomly from
the available data samples. After the calculation of the
distance of each data sample from set X to each clusters’
centroid, each data sample is declared to be a member of its
closest cluster. A set of data samples that belong to a cluster
defined by centroid i
c is denoted as i
c, 1iK . In each
iteration, a centroid is estimated as a mean value of d
corresponding parameters of all data samples which are a
member of a corresponding cluster. Calculating K new
centroids in each iteration is equivalent to changing a
clusters' position in a d-dimensional space, till optimal
clusters positions are reached. The processes of clustering
and updating centroids are repeated until convergence has
been reached or for a specified number of iterations.
One drawback of KM clustering appears when a point is
equally close to more than one centroid. In this case, the
algorithm will not converge, since this point belonging will
oscillate among a few different clusters, resulting in
different clusterings. However, this rarely happens in
practice [3], [4].
III. FUZZY K-MEANS CLUSTERING
Fuzzy K-Means (FKM) clustering method is a modification
of the standard KM clustering. As in KM clustering, initial
centroids are chosen randomly. Each iteration in FKM
clustering also starts with calculating the distance of each
data sample to each centroid. However, in FKM there is a
belonging degree, which is inversely proportional to that
distance. A data sample belongs to every cluster with a
certain degree [5]. Hence, borders among clusters are fuzzy.
FKM clustering is also referred to as soft clustering.
In FKM clustering, all data samples affect the calculation
of new centroids. The impact of a data sample on the
calculation of clusters’ centroids is proportional to the
degree of its belonging to that cluster. The other part of the
FKM algorithm is the same as in the KM algorithm.
Fig. 1. K-means clustering algorithm.
A. Fuzzy K-means clustering algorithm
Input to the FKM algorithm is the number of clusters K.
For n data samples the algorithm gives as a result an n

K
matrix W, with elements:
2
1
1
1
(, ) .
k
m
m
n
ii
kic
wij
xc
x









(1)
,ij
wis the degree to which element i
x
belongs to cluster
j
c,,
01
ij
w

. mR

is the fuzzifier, it defines the level
of cluster fuzziness, and 1m. In the absence of a priori
knowledge of the data fuzziness, it is recommended to set
the fuzzifier according to the database and expected results
of clustering. By adjusting the fuzzifier m, border between
clusters can be more fuzzy or more clear. The procedure
for its selection is illustrated in the next section through the
simulation results of FKM algorithm for different values of
m.
New centroids for K cluster are calculated on the basis of
all the data samples:
()
()
m
k
x
km
k
x
wx x
cwx


. (2)
The procedure repeats until reaching convergence or for a
specified maximum number of iterations. Both algorithms

58 Telfor Journal, Vol. 8, No. 1, 2016.
converge when
1tt t
kk k
cc c


, where t is the number of
iteration and

is a sensitivity threshold. In all our
experiments, this threshold is 0.01.
FKM clustering is computationally more complex than
KM clustering. KM calculates a distance and chooses the
smallest o ne in order to find to which cluster a data belongs,
while FKM performs additional Kxd multiplications for
each data sample (1), (2). However, the FKM algorithm has
a better result for a data set with overlapped clusters than
the KM algorithm. Moreover, in a case where data samples
are equally close to more than one centroid, the FKM
algorithm will not oscillate, unlike KM algorithm. FKM
algorithm will give an equal belonging degree of these data
samples to more than one cluster.
IV. S
IMULATION
R
ESULTS
Algorithms for KM and FKM are implemented in Java
and adapted for a soil database. A database of soil samples
sampled in Montenegro is used for clustering; 2526 data
samples are used. The goal is to estimate to which soil type
a sample belongs, using KM and FKM methods of
clustering. Based on the knowledge of the types of soil in
Montenegro, the number of clusters is chosen. The number
of clusters is essential to proper clustering. Thus, it is
important to have a correct value of this parameter. Since
this parameter is not always known in advance, the cases of
clustering with a smaller or higher number of clusters is also
investigated. The conclusion is that, the final centroids are
quite similar to initial centroids in case of using a number
of clusters smaller than a correct one. On the other hand, in
the case of using a larger number of clusters than a correct
one, clusters with a small number, zero or nearly zero, of
elements appears. This was an additional confirmation of
defined number of soil types in Montenegro. Moreover,
these conclusions can be used as a guide for determination
of a correct number of clusters in the case when this
information is not known in advance.
In Montenegro there are four to six types of soil. In our
simulations, data samples with only six parameters are used
for clustering. The used parameters represent the numerical
values of chemical characteristics of soil samples.
Consequently, this clustering of soil can be considered as
basic one. Hence, the optimal number of clusters for this
basic clustering is four.
Fig. 2. Visualisation of soil database, each point is a data
sample with three parameters.
For the verification and visualization of the performance
of the algorithm, clustering of soil samples based on two or
three parameters in three or four clusters was done and is
presented in this section. MATLAB is used for a graphical
representation of results.
Data samples with three parameters from soil database
are presented in a 3-D space before clustering (Fig. 2.).
Clustering soil samples is done by KM and FKM
algorithms. The maximum number of iterations is 100 in
both algorithms.
The mean value of the number of iterations needed for
achieving the convergence of KM algorithm in the case of
clustering in four clusters based on six parameters is 7 in
500 runs. Fig. 3 shows how the parameters of centroid
converge, in the case of clustering in four clusters based on
six parameters. Convergence is reached after 8 iterations.
Fig. 3. Convergence of centroids of K-means clusters in
the case of clustering in four clusters based on six
parameters.
The mean value of the number of iterations for the FKM
algorithm is 21 in 500 runs for the case of clustering in four
clusters based on six parameters. In the case of clustering in
four clusters based on six parameters, presented in Fig. 4,
the convergence is reached after 14 iterations.
Fig. 4. Convergence of centroids of fuzzy K-means
clusters in the case of clustering in four clusters based on
six parameters.

Hot and Popović-Bugarin: Soil data clustering by using K-means and fuzzy K-means algorithm 59
A. Results of K-means Clustering
Results of K-means clustering based on two parameters
for three and four clusters are presented graphically in Fig.
5.a) and b), respectively. Each data sample belongs to only
one cluster. Hence, each colour of data sample presents one
cluster.
Clustering based on three parameters in four clusters of
the same data is presented in Fig. 6.
Fig. 5. K-means clustering based on two parameters. Each
colour presents one cluster a) K=3; b) K=4.
Fig. 6. Clustering in four clusters (types) based on three
parameters. Each colour presents one cluster. a), b), c)
show different projections in three-dimensional space;
B. Results of Fuzzy K-means Clustering
Validation of FKM clustering results is also done by its
visualisation. In FKM clustering each data sample belongs
to each cluster, with a different belonging degree. Hence,
each FKM clust er was drawn on a different graph, while the
colours of samples depend on the belonging degree of each
sample to that cluster. This way of visualisation allows a
proper way of presenting the propagation of clusters. Dark
red presents a belonging degree near to 100%. Yellow and
green present data samples witch partly belong to a plotted
cluster. Light blue to dark blue colours present belonging
degrees whose values are less than 30%.
Parameter m is the fuzzifier and it defines the level of
cluster fuzziness [6]. FKM clustering of soil database was
done with different values of m. Results of FKM clustering
with different fuzzifiers are presented in Fig. 8-9.
Fig. 8.a) presents four FKM clusters of soil database,
where fuzzifier m is 1.1. Dark red is a dominant colour
around centroids and the borders between clusters are clear.
Hence, fuzziness of clusters is missing, so results are similar
to KM clusters.
The second example is FKM clustering where fuzzifier is
2 (Fig. 8.b). The appearance of all colours, from dark red to
dark blue, means that all belonging degrees from 0 to 100%
appear. Different clusters are visible but borders are still
fuzzy.
Fig. 7. Fuzzy K-means clustering based on three
parameters, each graph presents one cluster, colour
depends on belonging degree, K=4 a) m=1.1; b) m=2.
FKM clusters with a fuzzifier 3 are shown in Fig. 9.a)
Clusters are visible, but dark red colours are missing. This
means that belonging degrees near to 100% do not exist. All
belonging degrees in this case are smaller, while a blue
colour is dominant. Hence, fuzziness of clusters is
significant. A border between clusters is less visible than in
the case where the value of m is smaller.
The results of FKM clustering of the same database with
a fuzzifier 5 are shown in Fig. 9.b). Clusters are not visible;
all belonging degrees are equal. Hence, this clustering is
unsuccessful and pointless. The belonging degree of each
sample tends to be equal for every cluster when a fuzzifier
takes values higher than 3.
A conclusion from these examples is that a fuzzifier
value 2 is optimal for this database. Clusters are visible and
borders of clusters are fuzzy enough. These examples show
the choosing procedure of fuzzifier for a specific database.
Having in mind the previous conclusion, FKM clustering
of soil database with a fuzzifier 2 is performed.

60 Telfor Journal, Vol. 8, No. 1, 2016.
Fig. 8. Fuzzy K-means clustering based on three
parameters; each graph presents one cluster; colour
depends on belonging degree, K=4 a) m=3; b) m=5.
The result of the FKM clustering of soil database based
on two parameters and three clusters is presented in Fig. 9.
Clustering is done based on two parameters, so coordinates
present the position of soil samples in 2D space. Belonging
degrees are inversely proportional to the distance between
samples and centroids in 2D space. Increasing the distance
between a sample and centroid reduces the belonging
degree and colours change proportionally from red to blue.
Based on these graphs, a conclusion is made that the FKM
clustering algorithm made a successful clusterization of soil
data.
Fig. 9. Fuzzy K-means clustering in three clusters, based
on two parameters. Each graph presents one cluster, while
colour depends on belonging degree.
Fig. 10 presents the result of FKM clustering in four
clusters based on two parameters of the same database. A
conclusion that clusterization of soil data is successful can
be made, as in the previous case.
Fig. 10. Fuzzy K-means clustering in four clusters, based
on two parameters. Each graph presents one cluster, while
colour depends on belonging degree, K=4.
Soil samples in used database besides physical and
chemical characteristics, have the coordinates of location.
The coordinates of soil samples from database are given in
meters in the coordinate system MGI 1901 / Balkans zone
6. First, coordinates are converted to longitude and latitude.
Graphical environment for presenting the results of
clustering on the static Google map of Montenegro is
implemented in Java (Fig. 11.). Soil samples are labelled
with markers and the colour of marker depends on a marked
sample's soil type. It allows searching the map by
municipalities and soil types, and adjusting the zoom.
Different types of maps are available: roadmap, satellite,
hybrid and terrain.
Fig. 11. K-means clustering of soil samples presented on
Google static map.

Hot and Popović-Bugarin: Soil data clustering by using K-means and fuzzy K-means algorithm 61
The second method of presenting the results of soil
clustering is on dynamic maps. Dynamic maps are
implemented using R programing language. In addition, a
Leaflet is used. It is a set of open-source JavaScript libraries
for interactive maps. An Open Street dynamic map is shown
in Fig. 12. Markers present all soil samples from the
database with coordinates. The colour of marker presents
the results of KM clustering. Each colour is related to a
different cluster. Four clusters are depicted.
Fig. 12. Dynamic Open Street map with results of KM
clustering, markers of different colours presents different
KM clusters.
This map can be considered as a basic pedologic map,
because clustering is made based on only six chemical
parameters and only four clusters were made. Mapping of
KM results more precisely is possible by using more soil
parameters and clustering in more clusters. On this map in
Fig. 12. it is visible that two types of soil are dominant, blue
and green colours. Comparing this to pedologic maps of soil
of Montenegro made by experts, it can be seen that two
dominant soil types are the same (at the same parts of
Montenegro) on both maps.
Using maps for presenting soil data and clustering results
allows marking data with markers, polygons, raster images,
images of soil profiles etc.
V. C
ONCLUSION
The problem of soil data clustering and visualization is
analysed in the paper. Data mining techniques, KM and
FKM, are adapted for this purpose. The visualisation of KM
and FKM results is used for the validation of results.
Results obtained by using KM are presented on the Static
Google map and dynamic Open Street Map of Montenegro.
Presented soil data and data mining result on maps are a
proper way of presenting data to scientists, land users and
people who want to get information about soil in
Montenegro. Our future work will be dedicated to
improving data mining techniques and publishing all results
through a WEB application.
R
EFERENCES
[1] J. C. Bezdek, R. Ehrilich, W Fill, “FCM: The Fuzzy C-means
Clustering Algorithm”, Computers & Geosciences, vol. 10, no. 2-3,
pp. 191-203, 1984.
[2] “Vector Quantization and Clustering”, Courses
of Electrical
Engineering and Computer Science, Massachusetts Institute of
Technology
.
[3] J. Macqueen, “Some Methods for Classification and Analysis of
Multivariate Observations” Proc. of the fifth Berkeley Symposium on
Mathematical Statistics and Probability, 1, page 281-297. University
of California Press, (1967)
[4] Andrew Ng,” CS229 Lecture notes”, Machine Learning Course
Materials
[5] S. Ghosh, S. K. Dubey, “Comparative Analysis of K-Means and
Fuzzy C-Means Algorithms”, ((IJACSA) International Journal of
Advanced Computer Science and Applications, vol. 4, no.4, 2013.
[6] E. Hot, V. Popović-Bugarin, Ana Topalović, Mirko Knežević,
“Generating thematic pedologic maps by using data mining and
interpolations,” submitted for 3nd International Conference on
Electrical, Electronic and Computing Engineering IcETRAN 2016,
Zlatibor, Serbia, June 2016
[7] Md. K. I. Rahmani, N. Pal, K. Arora, “Clustering of Image Data
Using K-Means and Fuzzy K-Means”, (IJACSA) International
Journal of Advanced Computer Science and Applications, Vol. 5, No.
7, 2014
[8] R. Suganya, R. Shanthi, “Fuzzy C- Means Algorithm- A Review”,
International Journal of Scientific and Research Publications, vol. 2,
Issue 11, November 2012 1 ISSN 2250-3153
[9] E. Hot, V. Popović-Bugarin, “Analysis Of Fuzzy K-Means
Clustering Method Using Database Of Soil Samples Sampled In
Montenegro,” Information Technology IT 2016, Žabljak,
Montenegro, February 2016
[10] J. Balkovič, Z. Rampasekova, V. Hutar, J. Sobocka and R. Skalsky,
“Digital Soil Mapping from Conventional Field Soil Observations,”
Soil & Water Res., 8, 2013 (1): 13–25
[11] S. Har-Peled, B. Sadri, “How Fast is the k-means Method?*”,
January 2, 2010
[12] Singaravelu.S, A.Sherin and S.Savitha “Agglomerative Fuzzy K-
Means Clustering Algorithm”, Nehru E-Journal A Journal of Nehru
Arts and Science College (NASC) Research Article
[13] S. Chattopadhyay, D. Kumar Pratihar, S. C. De Sarkar, “A
Comparative Study of Fuzzy C-Means Algorithm and Entropy-
Based Fuzzy Clustering Algorithms”, Computing and Informatics,
Vol. 30, 2011, 701–720
[14] L. G. Vendrusculo, A. L. Kaleita, “Terrain Analysis and Data Mining
Techniques Applied to Location of Classic Gully in a Watershed”,
2013 ASABE Annual International Meeting
[15] L. Rokach, O. Maimon, “Clustering Methods”, In The Data Mining
and Knowledge Discovery Handbook, pages 321–352. 2005.
[16] J. C. Dunn, "A Fuzzy Relative of the ISODATA Process and Its Use
in Detecting Compact Well-Separated Clusters", Journal of
Cybernetics 3: pp. 32-57, 1973
[17] D. Rajesh, “Application of Spatial Data Mining for Agriculture,”
International Journal of Computer Applications, (0975 – 8887)
Volume 15– No.2, February 2011
[18] U. Fayyad, G. Piatetsky-Shapiro, P. Smyth, “From Data Mining to
Knowledge Discovery in Databases,” AI Magazine, Vol 17, No 3,
1996
[19] B. Fustic, G. Djuretic, “The Soils of Montenegro,” University of
Montenegro and Biotechnical Institute, Podgorica, Montenegro,
2000.
[20] S. Ghosh, S. K. Dubey, “Comparative Analysis of K-Means and
Fuzzy C-Means Algorithms,” (IJACSA) International Journal of
Advanced Computer Science and Applications, vol. 4, no.4, 2013.
[21] E. Hot and V. Popović-Bugarin, "Soil data clustering by using K-
means and fuzzy K-means algorithm," Telecommunications Forum
Telfor (TELFOR), 2015 23rd, Belgrade, 2015, pp. 890-893.