IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
HIFRARCHICAL CLASSIFICATION OF MULTIDATE WIFS IMAGES USING
ARTIFICIAL NEURAL NETWORKS
J.Kannan, Jayesh V. Acharekar and B. Krishna Mohan
CSRE, Indian Institute of Technology Bombay, Powai, Mumbai — 400 076
KEY WORDS: Neural Networks, Hierarchical Classification, Multitemporal data, Image classification
ABSTRACT:
The use of neural networks offers useful properties and capabilities like nonlinearity, input-output mapping, adaptivity, evidential
response, contextual information, fault tolerance etc. Voluminous amount of data are sent by various sensors from different satellites to
the respective ground stations, from earth resources to meteorology. To extensively work with this data, it is very essential to integrate
these available data sets in a prominent platform with the aid of modern technology and techniques. The recent development in the neural
network and its advantage over the multitemporal data sets has been encouraging and the present paper represents the effectiveness of the
same taking Varanasi, India, as the study area. The Principal Component Analysis (PCA) was carried out for the available multi-
temporal data sets, and the output was classified with conventional MLC and also with Neural Networks.
INTRODUCTION
A neural network is a massively parallel-distributed processor
made up of simple processing units, which has a natural
propensity for storing experiential knowledge and making it
available for use (Haykin, 1999). Neural network classification
was found to be superior to conventional classification like
Maximum Likelihood Classification in cases wherein the data sets
are of varying dynamic ranges, and statistical distributions
(Peddle et al., 1994). The use of neural networks offers useful
properties and capabilities like nonlinearity, input-output
mapping, adaptivity, evidential response, contextual information,
fault tolerance etc. Some of the earliest experiments on
classification of remotely sensed images using artificial neural
networks are reported by Hermann and Khazenie., (1992) and
Liu and Xiao, (1991); for geological mapping by An and Chung
(1994). Image segmentation techniques and hierarchical
approaches for analysis of remotely sensed images can be found
in (Benie and Thomson, 1992; Qiu and Goldberg, 1985; Gonzalez
and Lopez Soria, 1991; Gerylo et al., 1988);
In the present scenario, remote sensing has proved a powerful
technology for monitoring the earth’s surface and atmosphere at a
global, regional, and even local scale. This is made possible by
the large amount of data acquired by different type of sensors,
which provide repeated coverage’s of the planet on a regular
basis. As a consequence, an increasing quantity of multi-source
and multi-temporal remote-sensing data acquired in many
geographical areas is available. For proper explanation of these
data, it is mandatory to develop effective data fusion techniques in
order to take advantage of such multi-temporal data
characteristics. In particular, in the context of classification
problems, data fusion may provide an improvement in accuracy
(as compared with standard techniques applied to single-
sensor/single-date images), which may be primary importance in
real application.
Multi-temporal Data Sets are useful in quantifying the changes
occurring within the test site over a period of time. These data sets
are also essential to distinguish temporally varying vegetation
cover such as crops from other vegetation types such as forest and
plantation. A number of studies in this direction are published in
literature (Rencz, 1985; Guindon,1988; Kohei, 1991; Ban and
Howarth, 1999; Byeungwoo and Landgrebe., 1999; Bruzzone et
31.1999) The data acquired from IRS Wide Field Sensor has
proven much effective for such multi-temporal data analysis
because, of larger swath (770 Km), high repetitivity (5 days) and
operation in two vegetation specific bands, the sensor providing
vegetation index at regional level, thus helping in assessment of
crop condition and drought monitoring.
WiFs Specification
Values
3 - 0.62-0.68 (red)
B4 - 0.77-0.86 (near IR
Resolution (m 188
CCD devices 2048 elements
Swath 810 (5 days
ivalent focal len 56.4
o. of levels 128 (7 bits
SNR 2128
Spectral bands (microns)
Geology of Study Area:
The study area comprises extensively of Neogene-Quaternary
sediments. The Ganga Valley is a vast, yet less studied, Neogene-
Quaternary basin of India in between the Peninsular shield and
the Himalayan tectogene. The pre-Quaternary rocks of this basin
are entirely covered by a fairly thick sequence of unconsolidated
and semi-consolidated Quaternary deposits. The age of the
