assifiers is
roaches in
ternatives,
vn greater
to realize
roach has
= and land
ions of (i)
rmation at
Sufficient
developed
‘mentation
us image
used with
of object
nsing data
tion and
ailable by
important
age is not
ge objects
ie adopted
it does not
Dy a group
| step, the
ny chosen
s of fuzzy
1etwork of
the image
(scales)
between
al context
is "context
d together
) improve
s adjacent
'ontrasting
1 addition,
lows the
esolutions
nships at
neaningful
y grouped
image objects adds a topological structure, which enables
contextual analysis. The image objects act as the building
blocks for the subsequent image analysis. In comparison to
pixels, image objects carry much more useful information.
Thus, they can be characterised by far more properties than
pure spectral or spectral-derivative information, such as their
form, texture, neighbourhood or context. After basic object
segmentation is completed, it's possible to apply supervised
classification in which image objects of known classes are
selected and others of similar properties are automatically
detected. Such grouping takes advantage of the object's
physical and spectral properties. The approach is similar to a
traditional nearest-neighbor method, but it uses "fuzzy"
methods. In addition, exporting image objects in vector
shapefile form and/or their visual interpretation at each of the
segmentation levels are some of the options available to the
analyst. For more information on eCognition software and it's
methodical ^ principles, reader may please visit
http://www.definiens-maging.com/index.htm.
2. MATERIALS AND METHODS
2.1 Study Area and Satellite Data
In the present investigation, two sites with the availability of
remote sensing data acquired by an optical sensor and C-band
Synthetic Aperture Radar (SAR) in the microwave region have
been selected. Site-1 covers part of Guntur District in Andhra
Pradesh and is characterized by a wide range of land use and
land cover variations, which included crops viz, cotton,
chillies, paddy, maize, tobacco, pulses, banana plantation,
paddy fallows, uncultivated bare fields, water bodies and built
up areas. For site-1, two sets of remote sensing data viz., optical
and microwave data have been used independently for land use
land cover information. The optical data from IRS-1D LISS-III
sensor in three spectral channels were acquired on January 3,
2002 and microwave data from RADARSAT SAR in C-band in
extended low (16? look angle, 35m spatial resolution) mode on
January 3, 2002 and standard beam — 7 (45? look angle and
25m spatial resolution) mode on Decemberll, 2001 and
January 4, 2002. Site-2, also in Andhra Pradesh, is located in
the Medak district and covers the International Crop Research
Institute for Semi Arid Tropics (ICRISAT) farm, which is
characterized by well laid out agricultural plots with red and
black soils, cultivated fields, lakes etc. and the surrounding area
with settlements, water bodies, mixed plantations and bare
fields. The IRS-LIIS-III data covering the second site was
acquired on 15 March 2000. Ground truth collection for major
land use and land covers has been carried out on the day of
remote sensing data acquired by the microwave sensor. The
ground truth locations have been marked with hand held GPS
receiver.
2.2 Procedure
Procedure adopted in the current investigation involved (i) data
preparation for analysis, training area extraction and
separability analysis, classification of satellite data of two sites
by per pixel classifier using Maximum Likelihood algorithm
and by object oriented multi-resolution segmentation and
classification, analysis of classification results in terms of
subjective evaluation and classification accuracy measures.
2.2.1 Data preparation and analysis: The IRS-LISS-III data
of the two sites were geocoded by identifying the common
GCPs on topomap and the satellite data and resampling.
Similarly, the SAR data pertaining to the Site-1 were geocoded
IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India,2002
using the geocoded IRS-LISS-III data of the study area as the
master image. The SAR data sets were processed for speckle
suppression using a 3x3 convolving window of Lee
Multiplicative filter iteratively. Based on the preliminary field
survey, training and test area sites for major land use and land
cover classes were identified and their statistics generated. Pair
wise separability was computed using the Transformed
Divergence measure for all the classes considering (i) only the
IRS-LISS-III multispectral data set, and (ii) the dual angle
RADARSAT SAR data set for the Site-1 and (iii) IRS-LISS-III
multispectral data set for the Site-2. In site-1, while separability
was good for several classes with the Transformed Divergence
being maximum at 2.0, pair-wise separability was poor for
banana and cotton with TD=1.23, maize and chillies
(TD=1.27), early pulses and tobacco (TD=1.39), shallow water
bodies and bare fields (TD=1.49) and harvested paddy fields
and settlement (TD=1.42). In site-2, good separability was seen
with most of the classes. However, pairwise separability for
settlements and bare fields (TD=1.78), plantations and other
vegetation (TD=1.8) was marginally low. The LISS-III datasets
pertaining to the two sites have been classified by identifying
the training areas and using maximum likelihood algorithm and
by object oriented multi-resolution segmentation and
classification approach.
2.2.2 Multi-resolution Segmentation and Classification of
satellite data: The segmentation and classification technique
used in the study has two steps: Multi-resolution segmentation
of an image and its classification by Nearest Neighbour. Multi-
resolution segmentation has been carried out to minimizes the
average local heterogeneity of image objects for a given
resolution. In the segmentation step, initially several parameters
like image layers, scale parameter, homogeneity criterion in
terms of colour and shape have been defined. While the image
layers refer to the number of input features, scale parameter is
an abstract term, which determines the maximum allowed
heterogeneity for the resulting image objects. In heterogeneous
data the resulting objects for a given scale parameter are
smaller than in more homogeneous data. The color criterion
defines to which percentage the homogeneity is defined by the
spectral values of the image layers as opposed to the shape.
Similarly, the shape criterion is defined by two parameters:
smoothness and compactness. The smoothness criterion is used
to optimize image objects with regard to smooth borders and
the compactness is used to optimize image objects with regard
to compactness. Table 1 shows site-wise details of these
parameters.
Nearest Neighbour rule, which was used in the study, offers a
quick and simple classification of image objects by clicking
typical image objects as representative samples of a class. In
addition, class related feature options such as mutual relations
with the other objects (at the same segmentation level) or with
the sub objects ( in lower level) or with the super objects (in the
higher level) have been exercised to optimize the classification
results. In addition, classification options such as manual,
semi-automatic and automatic are possible at all levels of
segmentation, which have been judiciously used. Figure 1
shows results of segmentation and classification of IRS-LISS-
III data of Site-1. It is clear from the figure that larger the scale
parameter larger the resulting image objects by grouping
smaller image objects formed in the lower level of
segmentation. To avoid merging of image objects that belong to
different land use classes, segmentation has been carried out at
each level by interactively selecting the scale parameter. For
example, in site-1, fallow fields and settlements were merged at
segmentation level-6
