IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
MULTI-RESOLUTION SEGMENTATION AND CLASSIFICATION OF
REMOTE SENSING DATA
P.V.Narasimha Rao', K. Dhanunjaya Reddy*, T.Sanjeev Kumar* and B.R.M.Rao'
* A&S Group, National Remote Sensing Agency, Balanagar, Hyderabad, Andhra Pradesh, India — 500 037
email: rao_pvn@nrsa.gov.in, rao_brm@nrsa.gov.in
*ND Soft Spatial Systems Pvt. Ltd., Panjagutta Colony, Hyderabad-16, Andhra Pradesh, India, email:dreddy @ndsss.com
KEY WORDS: Remote sensing, Multi-resolution segmentation, Classification, Image objects, Image texture
ABSTRACT:
Classification of moderate to coarse resolution remote sensing data into land use and land cover classes by per pixel classifiers is
known for decades. Availability of high resolution data from space platforms in recent years necessitated alternative approaches in
digital classification to satisfactorily account for inter and intra class variability of such data. Among several possible alternatives,
object oriented multi-resolution segmentation and classification approach, commercially known as eCognition, has shown greater
potential. It has flexibility to use relations across different segmentation levels and features within a segmentation level to realize
better classification accuracy. In the present study, object oriented multi-resolution segmentation and classification approach has
been applied to the 24m resolution IRS-1D LISS-III data of two test sites with a wide range of heterogeneity in land use and land
covers. Performance of the adopted approach is superior to that of per pixel classifier. The approach provided the options of (i)
interactive as well as automatic classifications, (ii) using class hierarchy and contextual information, and (iii) provides information at
different resolutions and scales. However, multi-resolution segmentation is based on certain heuristics and data dependent. Sufficient
care needs to be exercised during image segmentation to realize acceptable accuracy of classification.
1. INTRODUCTION
Segmentation is the subdivision of an image into separated
regions. Grouping of pixels in a remote sensing image (or
classification) into meaningful land use and land cover classes
using per pixel classifiers is known for decades. While per pixel
classifiers by supervised and unsupervised techniques yield
satisfactory results with medium to coarse resolution datasets,
their performance in general, is inadequate with highly textured
radar data and high resolution optical data. Increased
availability of high resolution remote sensing satellite data from
such sensors as IRS-LISS-III and IKONOS in recent years has
brought out the inadequacies of the per pixel classifiers to
account for the variability associated with the natural features.
The assumption of the per pixel classifiers that pixels are
statistically independent is unrealistic as neighbouring pixels
are more likely to be similar. The per pixel classifiers do not
take into account the relationship between the scale of objects
being classified and the size of the image pixel. They do not use
two of the key features used by the human visual recognition
system, image texture and context. Patterns are built up from
below on a pixel by pixel basis on the assumption that each
pixel is separate entity. Texture which is a measure of the local
variation in pixel values and context, the logical relationship
between an object and its neighbours, are ignored. (Mather,
P.M., 1990). In addition, need for extraction of information
residing at different spatial resolutions and scales is appreciated
more in recent times. As a result, alternative methods, including
texture and contextual information, information fusion,
hierarchical classification and object oriented multi-resolution
segmentation and classification of remote sensing data have
come into prominence.
Some of the simplest approaches to segmentation are global
thresholding, region growing by clustering, texture
segmentation algorithms by co-occurrence matrices (Haralick et
al. 1963), MRF models (Mao and Jain, 1992), wavelet
coefficients (Salari and Zing, 1995), fractal indices (Chaudhuri
and Sarkar, 1995) and knowledge based approaches (Gorte,
150
1998). More recently, Definiens Imaging, Germany developed
eCognition™, an object oriented multi-resolution segmentation
and classification scheme that extracts homogeneous image
objects and their relations.
In the present study the eCognition software has been used with
the main objectives of investigating the suitability of object
based segmentation and classification of remote sensing data
for land use and land cover.
1.1 Object oriented multi-resolution segmentation and
classification:
The adopted technique, which is commercially available by
name eCognition'“ is based on the concept that important
semantic information necessary to interpret an image is not
represented in single pixels but in meaningful image objects
and their mutual relations. The basic difference of the adopted
procedure with the per pixel based classifiers is that it does not
classify single pixel but rather image objects formed by a group
of pixels during the segmentation step. In the first step, the
software extracts homogeneous image objects in any chosen
resolution that are subsequently classified by means of fuzzy
logic. The basic strategy is to build up a hierarchical network of
image objects, which allows the representation of the image
information content at different resolutions (scales)
simultaneously. By operating on the relations between
networked objects, it is possible to classify local context
information. Beyond the pure spectral information this "context
information" (which often is essential) can be used together
with form and texture features of image objects to improve
classification. Multi-resolution segmentation separates adjacent
regions in an image as long as they are significantly contrasting
to each other and generates homogeneous regions. In addition,
a hierarchical network of image objects allows the
representation of image information in different resolutions
simultaneously, allowing classification of relationships at
different scales. As the image objects represent meaningful
areas the networking of adjacent and hierarchically grouped
21
In
ren
Sy!
bee
Pra
lan
chi
pac
anc
lan
ser
20(
ext
Jar
251
Jan
the
Ins
ch:
bla
wit
fie
acc
lan
ren
gre
rec
2.2
Pr
pre
ser
an
cla
sul