Annual Progress Report , 2013

The University of Kalyani

Clustering and its application to GIS

Parthajit Roy
Department of Computer Science,
The University of Burdwan, Burdwan-713104

Abstract
Pattern recognition is an emerging field in the modern computations. The modern world is gradually moving towards the
generic solutions rather than specific solution to a problem. For this the recognition of the pattern of the problem is very
important. On the other hand, learning is also pattern recognition. Clustering is a field of pattern recognition or pattern
classification where the leveled data are not available. i.e. the unsupervised type pattern classification is clustering. This
is more challenging because the domain knowledge is not available in clustering. This article is a survey of the clustering
techniques that have been explored by the author and the application of them in the field of geographic information systems.
Keywords
Clustering — Pattern Recognition — Geographic Information System

1. Introduction a value representing the degree of membership of ith object to

the jth cluster [2].
Pattern Recognition or Pattern Classification an emerging and
Duda et al [3] discussed a solid foundation of classifica-
promising field in the modern day computing and in the field
tion whereas Everitt et al [4] explored the clustering from
of Pattern Recognition, clustering is one of the interesting
theoretical aspects. An excellent survey on clustering up to
and difficult task. Clustering means unsupervised grouping
1999 is done by Jain et al [1]. An outstanding survey on
of objects where labeled data for training is not available [1].
clustering algorithms is done by Xu and Wunsch II [5].
This restriction has made the clustering more challenging.
Geographical Information System (GIS), on the other
Clustering mainly deals with the topology of the data set or
hand, is another emerging field [6] [7]. Clustering is an in-
proximity of the data set or some kind of density or distribu-
tegrated part of a GIS system. The application of clustering
tion measures. Learning by these tools only is a major focus
in GIS ranges from soil type grouping to crop clustering and
area of data clustering.
many more. Clustering in GIS is even more challenging be-
There are five major steps of data clustering proposed by
cause GIS deals with huge data. So, the clustering as well as
Jain et al [1]. These are
data structure for storing the knowledge about cluster is one
• Feature extraction of the major concerns of literally every GIS system. This re-
port is a survey of clustering and its application to geographic
• Introduction of the idea of proximity information systems.
• Clustering of grouping
2. Proposed Methods
• Data abstraction
The present article covers fuzzy clustering techniques geo-
• Validation of cluster metric and graph based clustering techniques. First of all a
novel variation of traditional fuzzy C-means algorithm has
Given n objects {o1 , o2 , · · · , on } to be clustered, the first been proposed. This is named as selective scale space based
thing is to extract the features that represents an object. If fea- fuzzy C-means technique. The second method is a geometric
ture vectors are m-attribute or m dimensional, then objetcs are clustering based on Delaunay triangulation. This method has
represented by S = p1 , p2 , p3 , · · · , pn ∀pi ∈ Rm . After the fea- applied parzen window technique for fine tuning. The third
tures are extracted, the notion of proximity is applied. There one is a variation of minimum spanning tree based clustering.
are different distance measures like Mahalanobis distance, Following subsections discusses them in brief along with the
Minkowski distance etc. available but the most frequently proosed algorithms.
used distance measure is Euclidean distance. Having the fea-
ture vectors and the distance in hand, the objective of the 2.1 Selective Scale Space based FCM
clustering is to find out a partition matrix P(S) of order C × n The of this technique is to cluster n number of objects whose
representing the clustering into C number of clusters where representative feature vectors are S = {p1 , p2 , p3 , · · · , pn } ∀pi ∈
each cell of the matrix ui j , ∀i = 1, · · · , n and j = 1, · · · ,C has Rm . First of all we are going to present the standard fuzzy C-
 Clustering and its application to GIS — 2/8

means algorithm. Here it is assumed that there are C number In the proposed model scale space technique has been
of clusters (A priori which is essential for FCM). Given such applied. This technique works good if the clusters are linearly
information, the standard Fuzzy C-Means technique is stated non-separable. The idea of scale space filter is that some
using algorithm 1. transformation will be applied the parameters so that the clus-
ters will become linearly separable. Some previous work on
Algorithm 1 Standard Fuzzy C-Means Algorithm clustering using scale space filtering can be found in [10].
Input: S = {p1 , p2 , · · · , pn } . Points to be clustered. Scale Space has a tendency to make far points even further.
Output: C = {k1 , k2 , · · · , kC } . The final cluster centers. This may destroy the proximity of data sets. For this, a novel
1: declare Fn×C S as Matrix Standard Deviation(S.D.) based selective model has been pro-
2: declare Ob jectiveValue as Real posed. The scale space is applied to those parameters where
3: INITIALIZE C LUSTER C ENTERS (C) the variability is not good. The low standard deviation means
4: while Ob jectiveValue ≤ Benchmark do not good variability. The scale space has been applied to
5: POPULATE F UZZY PARTITION M ATRIX (Fn×C ,C, S) those parameters only. The parameters having high standard
6: UPDATE C LUSTER C ENTERS (C, Fn×C ) deviation remains untouched. The final model after all kind
7: Ob jectiveValue ← OBJECTIVE F UNCTION(C, Fn×C ) of preprocessing and internal reorientation has been proposed
8: end while in the algorithm 2.

The fuzzy membership should be such that Algorithm 2 S.D. based Selective Scale Spaced FCM

n C
Input: S = {p1 , p2 , · · · , pn } , sdT hreshold
Output: C = {k1 , k2 , · · · , kC } . The final cluster centers.
∑ ∑ µik = n 1: declare Fn×C S as Matrix
k=1 i=1
2: declare Ob jValue as Real
with 0 ≤ µik ≤ 1∀i = 1, 2, · · · ,C and ∀ j = 1, 2, · · · n , 0 <
3: NORMALIZE DATA S ET (S)
∑nk=1 µik < 1 and ∑Ci=1 µik = 1 4: for All parameter [i] o f dataset S do
There is a need of validity index to check the quality of
5: sd ← GET S TANDARD D EVIATION(parami )
the clusters. In the present model this is measured by Xie-
6: if sd < sdT hreshold then
Beni index [8]. This index is a function of Variation v and
7: pi ← G AUSSIAN S CALE S PACE(pi )
separation of the clusters centers. Variation v can be stated
8: end if
using equation 1 [9].
9: end for
C n 10: INITIALIZE C LUSTER C ENTERS (C)
v(P,C, X) = ∑ ∑ µi2j E 2 (ci , x j ) (1) 11: while Ob jValue ≤ Benchmark do
i=1 j=1 12: POPULATE F UZZY PARTITION M ATRIX (Fn×C ,C, S)
where ci are cluster centers and x j are feature vectors. The 13: UPDATE C LUSTER C ENTERS (C, Fn×C )
separation measure can be stated using equation 2. 14: Ob jValue ← X IE B ENI O BJ F UNCTION(C, Fn×C )
15: end while
d(C) = mini6= j E 2 (ci − c j ) (2)
where the E(., .) ia the Euclidean norm and finally the Xie-
Beni index (XB) can be given using equation 3. 2.2 Delaunay Triangulation based clustering
This model is based on a variation of computational geometric
model known as Delaunay triangulation [11]. Given a set of
C n 2 2
vX (P,C) ∑i=1 ∑ j=1 µi j E (ci , x j ) representative feature vectors P = {p1 , p2 , p3 , . . . , pn } in a two
XB(P,C)X = = (3)
nd(C) nmini6= j E 2 (ci − c j ) dimensional space, The Delaunay Triangulation is a planar
decomposition of the point set. The notion of proximity in
The Xie-Beni index measures compactness of a single a Delaunay triangle has been adopted for the first phase of
cluster and separability of different clusters. The value of clustering in the present model. The algorithm is given below.
fuzzy membership functions for fuzzy C-menas algorithm is The second phase of this model introduced a fuzzy re-
given by the equation 4 [9]. finement to the clusters. The Fuzzy based final technique
NSFCDT is illustrated in Algorithm 4.
1
µi,k = 2 , 1 ≤ i ≤ C, 1 ≤ k ≤ n (4)

E(ci −xk )

m−1 2.3 Spanning Tree based Clustering
∑Cj=1 E(c j −xk ) This model is a minimum spanning tree based model. The
edges of the graph will have weight and this weight is, in
The Center update rule [9] is done through the equation 5
our case, Euclidian distance of the two nodes in the two di-
∑nk=1 (µik )m xk mensional plane. The edges between two data point gives the
ci = (5)
∑nk=1 (uik )m notion of proximity in our present model.
 Clustering and its application to GIS — 3/8

As the volume of GIS data is very large, the computation

of MST should be very efficient. Delaunay triangulation is a
Algorithm 3 Procedure Delaunay Classification planar decomposition and for planar graph, the upper bound
Input: PointSet = {p1 , p2 , . . . , pn } of number of edges is
Declare DT As DelaunayTriangleSet  
DT ← D ELAUNAY T RIANGLES(PointSet) g
E(G) ≤ max (n − 2) , n − 1 (6)
Declare ClusterSet As Set g−2
Declare NewSet As Set
where g is the grith of the graph [12]. As the grith of DT is
Declare TmpSet As Set
g = 3. Putting the value of g in equation 6 we get, E(V ) =
for all dt In DT do
3n − 6. i.e. the edges are of O(n). instead of O(n2 ) in case
if V ISITED(dt) = False then
of complete graph Kn . Also, if the distance is Euclidean, the
I NITIALIZE(NewSet)
MST of a complete graph is the MST of a Delaunay graph.
P USH(dt, Stack)
This suggests that a Delaunay based MST is efficient than bare
while E MPTY(Stack) = False do
MST. Also there is efficient algorithm due to Bowyer [13] and
Triangle T ← P OP(Stack)
Watson
√ [14] for finding DT of a point set with time complexity
if S ATISFIABLE(T )&N OT V ISITED(T ) then
O(n√n) . The final time complexity
√ of the algorithm will be
T.Visited ← True
O(n n) + O(n lg n) = O(n n) instead of O(n2 lg n) in case
NewSet ← NewSet ∪ {T }
of complete graph.
T mpSet ← N EIGHBORS(T )
for all Element In T mpSet do The algorithm can be stated as follows.
if N OT V ISITED(Element) then
1 Scan The Image and identify the pixels qualified for
P USH(Element, Stack)
being a part of the forest and let that set is V and let
end if
|V | = n.
end for
end if 2 Compute the Delaunay Triangulation of the given point
end while set V .
ClusterSet ← ClusterSet ∪ NewSet
end if 3 Compute the minimum spanning tree of the complete
end for graph having V as the vertex set. (We have to calculate
the MST of the DT because the MST of the Complete
Graph is the MST of the DT.)

4 Consider a benchmark distance.

Algorithm 4 Procedure Fuzzy Based NSFCDT 5 Start from a vertex and keep including vertices in a
Input: PointSet = {p1 , p2 , . . . , pn } single cluster as long as the edge length is less than
1: B UILD ( ClusterSet) . By Algorithm 3 benchmark value.
2: for all Ci In ClusterSet do 6 If an edge crosses the benchmark distance, then create
3: Area ← G ETA REA(Ci ) a separate cluster and start putting the new vertices into
4: Degree ← G ET F UZZY D EGREE(Area) the new cluster.
5: S ET S ATISFIABILITY C ONDITION(Degree)
6: for all Triangle In Cluster do 7 continue this process until all the points of V is not
7: N ← Neighbour(Triangle) processed.
8: for all E In N do
9: if F UZZY S ATISFIABLE(E) then The model further proposed fuzzy post processing known
10: if E ∈ C j , ∀ j = 1 . . . m & j 6= i then as fuzzy encroachment. The idea is, if a cluster is very large
11: Ci ← Ci ∪C j then the nearby cluster is not a separate cluster. instead it is a
12: else part of the larger one and the larger cluster will encroach the
13: Ci ← Ci ∪ {E} smaller one. The final fizzy encroachment algorithm can be
14: end if stated as follows.
15: end if
16: end for • Take the cluster sets C, produced by the SMT-based
17: end for algorithm.
18: end for
• For all c ∈ C do the following:

– Calculate the size of the cluster

 Clustering and its application to GIS — 4/8

– If the size is large, then Table 4. The Comparative Study of Models on Iris Data
∗ Encroach the nearby clusters if any, depend-
ing upon the fuzzy strength. Model Correct wrong
∗ Else make no change to the cluster. Name Prediction Prediction
• After the completion of the above process, discard all Pure FCM 89.3% 10.7%
negligible clusters and produce only fair clusters. Scale Space FCM 97.3% 2.7
SD Scale Space FCM 98% 2%
3. Results and Analysis
3.2 Result of Delaunay Triangulation based Cluster-
The clustering algorithms are applied to benchmark data, ran- ing
dom data and GIS data. The results are discussed below. The result of proposed NSFCDT is given below. We have
compared NSFCDT with the NSCABDT [16]. The Scheme
3.1 Result of Selective Scale Space based Model
has been tested with a series of data sets consisting of 5000,
The Result of pure FCM for iris data set [15] is given below.
6000, 7000, 8000, 9000 and 10000 random points. In each and
Table 1. The Result of Pure FCM every case, the proposed NSFCDT shows reasonably good
and comparable results. NSFCDT shows better performance
when the number of clusters is more (more then 5). The result
Pure FCM of a typical data set is shown for NSFCDT in grafical form in
Class Actual Correct Wrong %of Figure 1(a),1(b), 3(a) and 3(b). In Figure 1(a), a data set of
Name No. Pred. Pred. Error 4088 points has been taken. Figure 1(b) shows the Delaunay
Setosa 50 50 0 0% Triangulation of the data set. In Figure 3(a), the Graphical
Versicolor 50 47 13 19% result of crisp clustering is shown and in the Figure 3(b) the
Verginica 50 37 3 29% final fuzzy output of the second pass of the model is shown.
Figure 3(b) shows, that though crisply there are four clusters,
the fuzzy model produces three instead of four clusters as it
The Result of scale space applied to all parameters for iris
seems from the Figure 1(a).
data set [15] are shown below.

Table 2. The Result of Scale Space FCM

Pure FCM
Class Actual Correct Wrong %of
Name No. Pred. Pred. Error
Setosa 50 50 4 4%
Versicolor 50 49 0 2%
Verginica 50 47 0 6%

The Result of scale space applied to selective parameters

for iris data set [15] are shown below. Figure 1. A Set of 4088 data points in a spatial domain.

Table 3. The Result of Standard Deviation based Scale

Spaced FCM

Pure FCM
Class Actual Correct Wrong %of
Name No. Pred. Pred. Error
Setosa 50 50 0 0%
Versicolor 50 50 3 3%
Verginica 50 47 0 6%

The following table is a comparative study of the varia-

tions. This clearly shows the superiority of the selective scale
space based model than the other models. Figure 2. Crisp Clustering produces four different clusters
 Clustering and its application to GIS — 5/8

Figure 4. Comparison of the proposed model with

NSCABDT [16]

Figure 3. Fuzzy Clustering produces three different clusters.

The time-space complexity of the proposed model and
the time-space complexity of the model (based on Complete
graph) proposed by Foggia et al. [17] is given in the table 6.

The comparative study between NSFCDT and NSCABDT

Table 6. Complexity measures of Complete Graph based
for various data sets are presented in the table 5. The table
model and the Present model
clearly shows that the NSFCDT is superior in crisp clustering.
As we have proposed a second pass for fuzzy refinement, there
is slight overhead and the time complexity is slightly high.
The interesting point is that, the little time overhead of the Attribute Complete-Graph Model Present Model
√
present model offering a fuzzy rationality which models the Time Complexity O(n2 lg n) O(n n)
real life clustering better. Space Complexity O(n2 ) O(n)

Table 5. A Comparative Study of NSCABDT and NSFCDT The proposed NSFCDT model has also been applied on
GIS data taken from Remote Sensing image of LANDSAT-7
(Courtesy of Bidhan Chandra Krishi Viswavidyalaya, West
Run Time in Seconds Bengal, India). A very satisfactory results are found on appli-
No of NSCABDT NSFCDT cation of the techniques. A part of Sundarban area of West
Points Crisp Fuzzy Bengal having number of rivers and a very dense forest, the
Clustering Clustering model shows a very good result on a gray scale Remote Im-
Run Time Run Time Run Time age. The Crisp classification shows several tree set area (Six
5000 48.2 40.04 55.18 large clusters). Even a very low aperture canal separates two
6000 71.4 54.72 81.99 clusters distinctly (The Green color and Red color clusters in
7000 93.8 63.19 87.45 fig 7(a)). For Blue and Cyan cluster of fig 7(a), a small area
8000 117.9 97.37 135.82 having shallow distribution separates the clusters. Both the
9000 151.3 129.48 183.88 cases are clearly not realistic (clear from fig 5). The fuzzy
10000 189.4 167.73 228.31 inclusion gives a rational classification of the tree set areas
and it is established from the algorithm that there are broadly
four clusters instead of six, suggested by the crisp clustering.
Figure 5, 7(a) and 7(b) shows the results.
3.3 Result of Spanning tree based Clustering

The model has also been applied on a set of 5000, 6000,

7000, 8000, 9000 and 10000 points and the results have been
compared with Yang et al. [16]. Instead of the post processing
proposed in NSCABDT, the present paper proposed a novel
fuzzy encroachment technique. The results are neck to neck
and as the present model proposes fuzzy refinement, a better
rationality can be achieved and also the quality measures like
validity, stability of clusters can also be achieved by tuning
the fuzzy parameters. The comparison of the proposed model Figure 5. A portion of Sundarban area of West Bengal,
with NSCABDT is given in the figure 4. India.(Courtesy of BCKV, West Bengal, India.)
 Clustering and its application to GIS — 6/8

Figure 9. The Delaunay Triangulation of the attribute

points.
Figure 6. Crisp Clustering produces six main clusters. But
clusters in Blue and Cyan are essentially same. The same is
true for Green and Red Clusters also.

Figure 10. The minimum spanning tree generated from the

Delaunay Triangulation of the attribute points.

Figure 7. Fuzzy Clustering produces four different clusters.

The Cyan and Blue are merged to form a single cluster and so
also Red and Green and they have formd a single cluster
shown in Pink.

3.4 Application Spanning Tree Model to GIS

The spanning tree based proposed model has been applied on
Figure 11. The initial minimum spanning tree based
a 640 × 400 remote gray scale image [Fig 8] of Sundarvan,
clustering. The model creates clusters correctly but takes
West Bengal, India. The results are quite satisfactory. In the
benchmark distance for all of the clusters same. That is why
crisp clustering the result is showing several small clusters
it creates some extra clusters which are actually the part of
[Fig 11], which are not rational because from the figure 8, it is
the larger one. The red color circles shows that.
clear that they should form a single cluster. After running the
fuzzy encroachment process, the clustering has become even
more rational and ends up with five clusters[Fig 12]. The inter-
mediate triangulation and minimum spanning tree generation
are also shown in figure 9 and figure 10 respectively.

Figure 12. Final output of the proposed model. The model

produces only five clusters by the method of fuzzy
encroachment, which looks realistic. It also discards
negligible clusters shown in green circles in figure 11.
Figure 8. A portion of the Sundarvan Area of West Bengal,
India (Gray Scale Image).
 Clustering and its application to GIS — 7/8

4. Conclusion and Future Scope on Computing and Systems-2013, (Published in: Proce-
dia Technology of Elsevier, DOI:10.1016/j.protcy.2013.
The clustering is a very important and effective thing in the 12.400), vol. 10 pp. 596-603, September, 2013.
field of GIS. The main aim of clustering is to group data
sets. The grouping of data is a fundamental part of every
GIS systems. Different types of clustering techniques have
been proposed. The result of clustering is very promising in References
the generic data set as well as the standard benchmark data [1] A. K. Jain, M. N. Murty, and P. J. Flynn. Data cluster-
sets. The GIS results are also promising. Nevertheless, there ing: A review. ACM Computing Surveys, 31(3):264–323,
are scope of improvements. The newer area of GIS where 1999.
the clustering can be applied as well as the newer clustering [2]
techniques to GIS can be proposed. The use of more efficient Ujjwal Maulik and Sanghamitra Bandyopadhyay. Per-
data structures can improve the present models also. formance evaluation of some clustering algorithms and
validity indices. IEEE Transaction on Pattern Analysis
and Machine Intelligence, 24(12):1650–1654, December
List of Publications 2002.
1. Parthajit Roy and J. K. Mandal, “A Novel Fuzzy-GIS [3] Richard O. Duda, Peter E. Hart, and David G. Stork.
Model Based on Delaunay Triangulation to Forecast Pattern Classification. John Wiley, 2 edition, 2008.
Facility Locations (FGISFFL)” International Sympo- [4]
sium on Electronic System Design (ISED), 2011, Index Brian S. Everitt, Daniel Stahl, Morven Leese, and Sabine
at IEEE Xplore, DOI:10.1109/ISED.2011.48, pp. 341– Landau. Cluster Analysis. John Wiley & Sons, 5 edition,
346, 2011. 2011.
[5] Rui Xu and Donald Wunsch II. Survey of clustering algo-
2. Parthajit Roy and J. K. Mandal, “A Relaxed Parzen Win- rithms. IEEE Transactions on Neural Network, 16(3):645–
dow Based Multifeatured Fuzzy-GIS Model to Forecast 678, May 2005.
Facility Locations (RPWMFGISFFL)” Proceedings of [6]
the International Conference on Information Systems C. P. Lo and Albert K.W. Yeung. Concepts and Tech-
Design and Intelligent Applications 2012 (INDIA 2012), niques of Geographic Information Systems. Prentice Hall
(Published in Advances in Intelligent and Soft Comput- of Iindia, 2 edition, 2007.
ing, 2012 of Springer, DOI:10.1007/978-3-642-27443- [7] T.M. Lillesand, R.W. Kiefer, and J.W. Chipman. Remote
5), vol. 132/2012, pp. 623-630, 2012. Sensing and Image Interpretation. Wiley India, 5 edition,
2004.
3. Parthajit Roy and J. K. Mandal, “A Novel Spatial Fuzzy
[8] X. L. Xie and G. Beni. A validity measure for fuzzy
Clustering using Delaunay Triangulation for Large Scale
GIS Data (NSFCDT)” PProceedings of the 2nd Inter- clustering. IEEE Transactions on Pattern Analisys and
national Conference on Communication, Computing & Machine Intelligence, 13(8):841–847, August 1991.
Security [ICCCS-2012], (Published in: Procedia Tech- [9] Ujjwal Maulik and Sanghamitra Bandyopadhyay. Fuzzy
nology of Elsevier, DOI:10.1016/j.protcy.2012.10.054, partitioning using a real-coded variable-length genetic
vol. 6, pp. 452-459, 2012. algorithm for pixel classification. IEEE Transactions on
Geoscience and Remote Sensing, 41(5):1075–1081, May
4. Parthajit Roy and J. K. Mandal, “A Novel Spatial Fuzzy
2003.
Clustering using Delaunay Triangulation for Large Scale
[10] Yee Leung, Jiang-She Zhang, and Zong-Ben Xu. Cluster-
GIS Data (NSFCDT)” Proceedings of the 2nd Interna-
tional Conference on Communication, Computing & ing by scale-space filtering. IEEE Transactions on Pattern
Security [ICCCS-2012], (Published in: Procedia Tech- Analisys and Machine Intelligence, 22(12):1396–1410,
nology of Elsevier, DOI:10.1016/j.protcy.2012.10.054, December 2000.
vol. 6, pp. 452-459, 2012. [11] Mark de Berg, Otfried Cheong, Marc van Kreveld, and
Mark Overmars. Computational Geometry-Algorithms
5. Swati Adhikari, J. K. Mandal and Parthajit Roy, “On
and Applications. Springer, 3 edition, 2008.
The Strength of Self Organizing Feature Map for Clus-
[12] Bélla Bollobás. Modern Graph Theory. Springer-Verlag,
tering Non-Categorical Data” Proceedings of First In-
ternational Conference on Computational Intelligence: 2005.
Modeling Techniques and Applications (CIMTA) 2013, [13] A. Bowyer. Computing dirichlet tessalations. The Com-
pp. 89-93, September, 2013. puter Journal, 24(2):162–166, 1981.
6. Parthajit Roy And J. K. Mandal, “A Novel Selective [14] D.F. Watson. Computing the n-dimensional Delaunay
Scale Space based Fuzzy C-means Model for Spatial Tessalation with application to Voronoi polytops. The
Clustering” Proceedings of Second Internal Conference Computer Journal, 24(2):167–172, 1981.
 Clustering and its application to GIS — 8/8

[15] R. A. Fisher. UCI machine learning repository. 1936.

[16] Xiankun Yang and Weihong Cui. A novel spatial cluster-
ing algorithm based on delaunay triangulation. Journal
of Software Engineering and Applications, 3:141–149,
2010.
[17] Pasquale Foggia, Gennaro Percannella, Carlo Sansone,
and Mario Vento. A graph-based clustering method and
its applications. Proceedings of Advances in Brain, Vision,
and Artificial Intelligence, 4729:277–287, 2007.