A vector machine based approach towards object oriented classification of remotely sensed imagery.
Arun, Pattathal Vijayakumar ; Katiyar, Sunil Kumar
Reference to this paper should be made as follows: Arun, P. V.;
Katiyar, S. K. 2014. A vector machine based approach towards object
oriented classification of remotely sensed imagery, Geodesy and
Cartography 40(1): 1-7.
UDK 528.85
Introduction
Land cover plays a pivotal role in impacting and linking many parts
of the human and physical environments, hence monitoring its changes is
highly significant. Remote sensing techniques are widely used for land
cover classification and urban analyses. The accuracy of pixel based
classification approaches are getting affected due to the increased
resolution of satellite images and object based strategies are being
devised as an alternative (Vapnik 1998). Literature reveals a great deal
of advanced methodologies in this context; however these approaches may
be further improved with the incorporation of more object specific
parameters (Srivastava 2004; Schnitzspan et al. 2008). The spectral and
spatial information can be combined to increase the seperability between
classes to yield higher classification accuracy (Hosseini, Homayouni
2009).
Vector Machines (VM) along with mercer kernels have been widely
used for the classification of multispectral as well as hyper spectral
images (Tan et al. 2011; Schnitzspan et al. 2008). The technique
constitutes of finding an optimal separation between the classes and
also uses kernel method to project linearly inseparable data to a higher
dimension space. Kernel methods have useful properties which facilitate
the separation of even closely related classes, using a low number of
(potentially high dimensional) training samples (Mercier, Lennon 2003).
The existing VM based approaches do not consider the spectral meaning
and behavior of the data, but instead rely on geometric measures for
class separation (Schnitzspan et al. 2008). Mercier, Lennon (2003)
proposed the linear mixing of quadratic with spectral kernels (Spectral
Information Divergence& Spectral angle based) to achieve better
classification results as compared to the statistical based approaches.
Yanfen Gu et al. (2007) suggested soft classification of hyper spectral
imagery by incorporating spatial and spectral information using
composite kernel. However most of the object specific interpretation
parameters have not been considered in these approaches. We propose a
composite kernel strategy to analyze the spatial features along separate
spectral bands, hence combining the spectral and spatial information for
effective interpretation.
We have also investigated to augment vector machine based
approaches with the incorporation of object specific interpretation keys
such as shape and texture. Support vector machine (SVM), the commonly
used vector machine, is an Independent and Identically Distributed (IID)
classifier that does not consider interactions of adjacent data points;
but have appealing generalization properties (Hosseini, Homayouni 2009;
Huang et al. 2002). Support Vector Random Fields (SVRF) are Conditional
Random Field (CRF) based extensions of vector machines that facilitate
to better model the adjacency interactions (Chi-Hoon et al. 2005). SVRF
model is robust to class imbalance, can be efficiently trained,
converges quickly during inference, and can trivially be augmented with
kernel functions to improve results (Lee et al. 2005). SVRFs have
appealing generalization properties of SVMs along with the spatial
dependency modelling capabilities of CRFs. We investigate to incorporate
contextual parameters on SVRF variation of vector machines since it
provides flexibility for effective augmentation (Melgani, Bruzzone
2008).
We have adopted a hierarchical SVRF approach (Gustavo, Luisel 2006)
that incorporates SVMs along with multilayer CRFs in a consistent
framework, in order to automatically model the optimal interplay between
local, semi-local and global feature contributions. Proposed approach
considers feature shape along with other interpretation keys and hence
misclassification of objects under due to illumination variance may be
avoided. We have used evolutionary computing approaches such as Cellular
Automata (CA) as well as Genetic Algorithm (GA) for enhanced feature
modeling.
In this paper we propose a vector machine based frame work for the
incorporation of object specific interpretation keys to facilitate
effective classification of spatial images. We have also adopted
adaptive kernel strategy in which the kernel parameters have been
adjusted with reference to the ensemble distributions. Different
existing approaches along with the proposed approach have been evaluated
with reference to the study area.
1. Theoretical background
1.1. Random modelling techniques
Evolutionary computing approaches such as CA, GA and their variants
such as Cellular Neural Network (CNN) and Multiple Attractor Cellular
Automata (MACA), have been found to be useful for modelling random
features (Mitchell et al. 1996; Mnih, Hinton 2010). CNN is an analogue
parallel computing paradigm defined in space and is characterized by the
locality of connections between processing elements (Orovas, Austin
1998). Cell dynamics of this continuous dynamic system may be denoted
using ordinary differential equations as given in equation (1), where
vector G is the gene which determines the random nature.
[X.sub.k](f) = -X1 + f(G, [Y.sub.k], [U.sub.K]). (1)
CNN is effectively used for modelling object shape to facilitate
the incorporation of feature specific information in to the SVRF
kernels. Random rules governing the shape of a feature can be identified
by evolving the feature from a single state using CNN and GA. Abstract
representations of objects are obtained by evolving features
continuously until they can be separated from the background.
1.2. Vector machines
N-Dimensional classifiers such as VMs are non-probabilistic binary
linear classifiers that constructs a set of hyperplanes to optimally
separate the classes (Melgani, Bruzzone 2004). SVRF is a CRF based
extension for SVM (Chi-Hoon et al. 2005; Lee et al. 2005).
It considers interactions in the labels of adjacent data points
while preserving the same appealing generalization properties as support
vector machine (SVM) (Lennon et al. 2007). SVRF may be mathematically
represented using equation (2) where O([y.sub.i], i(X)) is an SVM-based
Observation-Matching potential and V([y.sub.i], [y.sub.j], X) is a
(modified) DRF pair wise potential:
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]. (2)
The observation-matching function captures relationships between
the observations and class labels, whereas local-consistency function
models relationships between neighboring data points. SVRF is used along
with the kernel functions to implement effective segmentation augmented
with contextual knowledge.
1.3. Kernels
Kernels augment vector machines in measuring the similarity between
two data points that are embedded in a high, possibly infinite,
dimensional feature space (Mercier, Lennon 2003; Gustavo, Luisel 2006).
Adaptive kernel strategy is implemented Mixture Density Kernel (MDK)
that measures the number of times an ensemble agrees that two points
arise from same mode of probability density function (Srivastava 2004).
Mixture density kernels are used to integrate an adaptive kernel
strategy to the SVRF based clustering as they facilitate learning of
kernels directly from image data rather than using a static approach.
The composite kernel concept is used to incorporate spectral and
spatial information, given X = [{[x.sub.1], [x.sub.2], ...,
[x.sub.m]}.sup.T] be the spectral characteristics of an M-band
multispectral imagery and Y = [{[y.sub.1], [y.sub.2], ...,
[y.sub.n]}.sup.M] be the spatial characteristics, then the possible
spectral and spatial kernel [K.sub.x](P,[P.sub.i]) < [PHI](P),
[PHI]([P.sub.i])>, [K.sub.y](P,[P.sub.i]) < [PSI](P),
[PSI]([P.sub.i]) respectively. Preferably a weighted combination of the
kernels are adopted as discussed in (Gustavo, Luisel 2006) such that K
(P,[P.sub.i]) [mu][K.sub.x](P,[P.sub.i]) (1 -
[mu])[K.sub.y](P,[P.sub.i]) and the value of tuning parameter is
adjusted with respect to the objects.
2. Experiments
2.1. Data
Proposed approach has been evaluated over multispectral LISS III
and LISS IV sensors images of Indian Remote Sensing Satellites and
details are given in Table 1. The image has been geo referenced using
ERDAS 9.1 and has been sub set for the Bhopal Area. The study area
constitutes of five land cover classes namely agriculture, urban,
barren, water, and forest. Study area is so selected that the spatial
distributions as well as area fraction of almost all classes are uniform
and hence the effect of these factors over classification accuracy is
made negligible.
[FIGURE 1 OMITTED]
The ground truthing has been done using Differential Global
Positioning System (DGPS) survey data collected over the study area
using Trimble R3 DGPS equipment with centimeter level accuracy. Details
of the data are presented in Table 2.
2.2. Proposed algorithm
The schematic representation of the proposed algorithm is as given
in the (Fig. 1). During training phase, first the edges are detected
using Canny operator, then a CA based region growing strategy is adopted
to approximatelyextract the objects. Each pixel is assigned a state,
namely 'B' for boundary pixel, 'NB' for non boundary
pixel and 'NR' for non region pixels. Initially boundary pixel
states are assigned as 'B' and non boundary pixel states as
'UB. The 'NB' pixel state is changed to 'NR'
iteratively if it is near to a boundary pixel. The whole procedure is
repeated until no further state change is experienced, thereby detecting
different objects in the image. Further CNN along with GA is used to
find rules that iterate from a given initial state to a desired final
state. This inverse mapping or evolution is exploited to model feature
shapes, and CNN rules used to evolve a particular feature are used to
distinguish it. These extracted rules reveal the shape information of
various features and are used for classification along with other
interpretation keys.
In the validation phase, SVRF uses training samples to classify the
input image where tone, texture and CA rules are adopted for an
effective approach. Kernel parameters are adjusted from an ensemble of
probabilistic mixture models, where each model is generated from a
Bayesian mixture density estimate. Features are calculated along each
band and composite kernel strategy is used to incorporate these spectral
and spatial information where tu [mu] is adjusted based on the object
nature.
2.3. Implementation
The algorithms have been implemented in MATLAB and were compared
with the commonly used conventional approaches. Relevant statistical
parameters such as Overall Accuracy (Mnih, Hinton 2010) and Kappa
Coefficient of agreement (Melgani, Bruzzone 2008; Nasset 1996) have been
used for comparative evaluation. The procedure of accuracy estimation is
as summarized in (Fig.2).
[FIGURE 2 OMITTED]
3. Results and discussions
The investigations of this research work revealed that augmentation
of vector machine based classification scheme with feature specific
parameters and spectral information reduces false alarms for thematic
classification. For instance, recreational forest area (VanVihar
national park- Bhopal), which has been difficult to classify due to
small trees and shadows, has been correctly classified with the
approach. Efficiency of the approach with reference to traditional
classifying techniques has been evaluated using various statistical
measures and results are as summarised in Table 3.
Investigation results reveal that the classification accuracy of
traditional methods has been affected due to the increase in resolution
of satellite images. This is evident from the lower accuracy of these
methods over LISS 4 image when compared to LISS 3. Accuracy of the
proposed approach has been found to be comparatively stable over the
change in resolution and has also found to perform better. Higher values
of kappa and over all accuracy indicate that the proposed algorithm is
giving better results at every resolution. Performances of the proposed
& SVM based approaches are similar for lower resolution data (say
LISS 3) since the differences in performance is the effect of object
specific parameters which are relatively less achievable at lower
resolution.
The performances of these methodologies have been also evaluated by
comparing areal extents of various features. The features having well
defined geometry like lakes, parks etc have been selected for
comparative analysis vector machine based classification. The original
surface areas of the features are calculated by manual digitization
using ERDAS and comparative the results are presented in Table 4.
Comparative analyses of the areal extents also indicate that the SVRF
approach yields better results compared to the other methods. The
VanVihar national park which is a recreational forest area has been
distinguished using the proposed approach and this indicates superiority
of this method for object based classification.
[FIGURE 3 OMITTED]
The classified results for the LISS 3 imagery using various
methodologies are as given in (Fig. 3) and visual interpretation also
reveals the accuracy of SVRF based methodology.
We have also investigated the effect of spectral considerations
with reference to vector machine approaches and results are summarized
in Table 5. It has been found that the spectral and spatial
considerations separately do not yield good results and spectral
considerations alone have the worst. However when combined together as
proposed it results in a significant improvement in the accuracy.
Investigations have revealed that the support vector based
approaches (SVRF&SVM) outperforms conventional counterparts and that
the proposed method is performing better. The main disadvantage of the
suggested approach is its computational complexity which can be improved
using coreset optimization and similar approximation techniques.
Complexity can be further reduced by storing the detected rule
variations; optimization methods such as GA can be exploited to optimize
the strategy. This research provides a basic framework and further
investigations are needed to enhance it. Integration of a fuzzy approach
to the inverse mapping also seems to be promising, since
fuzzy/neutrosophic cognitive maps can be exploited for effectively
organizing and selecting CA rules.
Conclusions
Vector machine based approaches have found to give better results
when augmented with probabilistic approaches like CRF, since the spatial
dependencies between classes have been taken in to consideration. The
investigation revealed that use of spectral knowledge along with object
specific parameters into SVRF classification reduces false alarms for
thematic classification. The proposed use of CA for the incorporation of
feature specific rules has found to yield better results. SVRF based
approach is found to outperform the contemporary methods and can be made
semi supervised by enhancing with Learning Automata. We have presented
the basic framework which needs further improvement for effective use.
doi: 10.3846/20296991.2014.890271
Caption: Fig. 1. Proposed algorithm
Caption: Fig. 2. Accuracy Analysis
Caption: Fig. 3. Visual comparison of different Classification
method results for LISS3 sensor imagery
References
Gustavo, C. V.; Luisel, G. C. 2006. Composite Kernels for
hyperspectral image classification, IEEE Geo science Remote Sensing
Letters, vol. 3, no. 1: 93-97.
http://dx.doi.org/10.1109/LGRS.2005.857031
Hosseini, R. S.; Homayouni, S. 2009. A SVMS-based hyperspectral
data classification algorithm in a similarity space, in Work Shop
onHyperspectral Image and Signal Processing: Evolution in Remote
Sensing, 2009: WHISPERS '09, Grenobe, France, vol. 1: 1-4.
Huang, C.; Davis, L. S.; Townshend, J. R. G. 2002. An assessment of
support vector machines for land cover classification, International
Journal of Remote Sensing 23(4) 725-749.
http://dx.doi.org/10.1080/01431160110040323
Lee, C.-H.; Greiner, R.; Schmidt, M. 2005. Support Vector random
fields for spatial classification, in 9th European Conference on
Principles and Practice of Knowledge Discovery in Databases, PKDD 2005,
3-7 October, 2005, Porto, Portugal, 121-132.
Lee, C. H.; Schmidt, M.; Murtha, A.; Bistritz, A.; Sander, J.;
Greiner, R. 2005. Segmenting brain tumors with conditional random fields
and support vector machines, in Proc. Workshop Computer Vision
Biomedical Image Applications 3765: 469-478. 2005.
http://dx.doi.org/10.1007/11569541_47
Lennon, M.; Mercier, G.; Hubert-Moy, L. 2007. Classification of
hyperspectral images with nonlinear filtering and support vector
machines, IGARSS 22(2): 141-147.
Melgani, F.; Bruzzone, L. 2004. Classification of hyperspectral
remote sensing images with support vector machines, IEEE Transactions on
Geosciences and Remote Sensing 42(8): 1778-1790.
Melgani, F.; Bruzzone L. 2008. Support vector machines for
classification of hyperspectral remote sensing images, IGARSS 23:
234-240. http://dx.doi.org/10.1109/TGRS.2004.831865
Mercier, G.; Lennon, M. 2003. Support Vector machines for
hyperspectral image classification with spectral-based kernels,
IEEE-Transactions of Geo Science and Remote Sensing 45(3): 123-130.
Mitchell, M.; Crutchfield, J. P.; Das, R. 1996. Evolving Cellular
Automata with genetic algorithms: a review of recent work, in First
International Conference on Evolutionary Computation and Its
Applications (EvCAX96), May, 1996, Moscow, Russia, vol. 1, no. 1:
120-130.
Mnih, V.; Hinton, G. 2010. Learning to detect roads in
high-resolution aerial images, in 11th European Conference on Computer
Vision (ECCV), January, 2010, Marseille, vol 10, no 1: 120-130.
Nasset, E. 1996. Use of the weighted Kappa coefficient in
classification error assessment of thematic maps, International Journal
of Geographical Information Systems 10(5): 591-603.
Orovas, C.; Austin, J. 1998. A cellular system for pattern
recognition using associative neural networks, in IEEE International
Workshop on Cellular Neural Networks and their Application, February,
1998, Rome, Italy, vol. 2, no. 5. 11 p.
Schnitzspan, P.; Fritz, M.; Schiele, B. 2008. Hierarchical Support
Vector random fields: joint training to combine local and global
features, in COMPUTER VISION--ECCV 2008 Lecture Notes in Computer
Science, vol. 5303. Heidelberg: Springer, 527-540.
http://dx.doi.org/10.1007/978-3-540-88688-4_39
Srivastava, A. N. 2004. Mixture density Mercer kernels: a method to
learn kernels directly from data, in Proceedings of 4th SIAM
International Conference on Data Mining, 22-24 April 2004, Lake Buena
Vista, Florida, USA.
Tan, K. C.; Lim, H. S.; Mat Jafri, M. Z. 2011. Comparison of Neural
Network and Maximum Likelihood classifiers for land cover classification
using Landsat multispectral data, in IEEE Conference on Open Systems
(ICOS2011), 25-28 September, 2011, Malaysia, vol. 3: 121-127.
Vapnik, V. 1998. Statistical learning theory. New York: Wiley
Publishers Inc., 230-240.
Yanfeng Gu; Ying Liu; Ye Zhang. 2007. A soft classification
algorithm based on spectral-spatial kernels in hyperspectral images, in
International Conference on Innovative Computing, Information and
Control, ICICIC '07, 2007, Kumamotto, Japan.
Pattathal Vijayakumar Arun (1), Sunil Kumar Katiyar (2)
(1) Ames center, NASA, California, USA
(2) Maulana Azad National Institute of Technology, Bhopal, India
E-mail: (1) arun.p.1988@ieee.org (corresponding author)
Received 21 July 2013; accepted 24 February 2014
Pattathal Vijayakumar ARUN. He has completed his Masters from
NIT-Bhopal, and is currently pursuing PhD. His main area of interest are
artificial Intelligence, spatial mining, and image processing.
Sunil Kumar KATIYAR is a Professor as well as Chairman of Remote
Sensing Center at Maulana Azad National Institute of Technology, Bhopal.
He is a countable expert in the field of remote sensing with years of
experience in the field. His area of interest include image processing,
GPS, and Cartography. He had served as the PI in several national and
international projects.
Table 1. Details of experimental data
S. Imaging Spatial Satellite Area Date of
No. sensor resolution acquisition
(m)
1 LISS-III 23.5 IRS-P6 Bhopal 5th April 2009
(India)
2 LISS-IV 5.6 IRS-P6 Bhopal 16th March 2010
(India)
Table 2. Ground truthing information
S.No Area Date of No. No. of Accuracy
procurement Classes points/
class
1 Bhopal November, 2012 5 40 cm level
([10.sup.-2])
Table 3. Results of accuracy analysis
S. Sensor Methodology Kappa Overall
No statistics accuracy
(%)
1 LISS 3 Mahalanobis 0.88 86.13
2 LISS 3 Minimum Distance 0.91 87.58
3 LISS 3 Maximum Likelihood 0.91 89.83
4 LISS 3 Parrellelepiped 0.93 92.81
5 LISS 3 Feature Space 0.94 92.85
6 LISS 3 SVM (Spectral & 0.97 96.92
spatial factor
considered)
7 LISS 3 SVRF (Spectral & 0.98 97.81
spatial factor
considered)
8 LISS 4 Mahalanobis 0.85 84.40
9 LISS 4 Minimum Distance 0.89 85.00
10 LISS 4 Maximum Likelihood 0.88 86.80
11 LISS 4 Parrellelepiped 0.90 88.62
12 LISS 4 Feature Space 0.92 90.56
13 LISS 4 SVM (Spectral & 0.98 97.84
spatial factor
considered)
14 LISS 4 SVRF (Spectral & 0.99 98.89
spatial factor
considered)
Table 4. Comparison of the geographical extent of various features
S.No Sensor Feature Reference Methodology Areal
area extent
([km.sup.2]) (W)
1 LISS3 Lake 32.5 Mahalanobis 25.42
Minimum 24.31
Distance
Maximum 27.37
Likelihood
Parallelepiped 28.58
Feature Space 26.82
SVM(Spectral 28.71
& spatial
factor
considered)
SVRF(Spectral 30.72
& spatial
factor
considered)
2 LISS3 Parks 2.13 Mahalanobis 0.82
Minimum 0.89
Distance
Maximum 1.45
Likelihood
Parallelepiped 1.37
Feature Space 1.51
SVM(Spectral 1.58
& spatial
factor
considered)
SVRF(Spectral 1.65
& spatial
factor
considered)
3 LISS3 Artificial 4.41 Mahalanobis --
Forest Minimum --
area Distance
(Vanvihar) Maximum --
Likelihood
Parallelepiped --
Feature Space --
SVM(Spectral 2.61
& spatial
factor
considered)
SVRF(Spectral 3.52
& spatial
factor
considered)
4 LISS4 Lake 32.81 Mahalanobis 24.31
Minimum 23.40
Distance
Maximum 25.12
Likelihood
Parallelepiped 26.24
Feature Space 27.17
SVM(Spectral 29.43
& spatial
factor
considered)
SVRF(Spectral 31.08
& spatial
factor
considered)
5 LISS4 Parks 2.37 Mahalanobis 0.51
Minimum 0.72
Distance
Maximum 1.53
Likelihood
Parallelepiped 1.14
Feature Space 1.46
SVM(Spectral 1.63
& spatial
factor
considered)
SVRF(Spectral 1.71
& spatial
factor
considered)
6 LISS4 Artificial 3.95 Mahalanobis --
Forest Minimum --
area Distance
(Vanvihar) Maximum --
Likelihood
Parallelepiped 1.81
Feature Space --
SVM(Spectral 3.42
& spatial
factor
considered)
SVRF(Spectral 3.62
& spatial
factor
considered)
Table 5. Analysis of spatial & spectral considerations
over classifier
S. No Sensor Methodology Kappa Overall
statistics accuracy (%)
1 LISS 3 SVM (spatial) 0.94 89.13
2 LISS 3 SVRF (spectral) 0.72 70.21
3 LISS 3 SVM (spectral+spatial) 0.97 96.92
4 LISS 3 SVRF (spectral+spatial) 0.98 97.81
5 LISS 4 SVM (spatial) 0.95 91.08
6 LISS 4 SVRF (spectral) 0.75 74.07
7 LISS 4 SVM (spectral+spatial) 0.98 97.84
8 LISS 4 SVRF (spectral+spatial) 0.99 98.89
