IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
CLASSIFICATION METHODS
Dr. R Krishnan
Director, ADRIN , Dept. of Space, 203, Akbar Road, Secundeabad.
krishnan @adrin.res.in
Commission VII, WGVII/1.2
KEY WORDS: Supervised, Unsupervised, MLC, MLP, Sample, Adaptive, Classification NN, fuzzy.
ABSTRACT:
The process of classification can be considered as mapping of continuous data into categorical classes of information about the
landscape. The tools of mapping can be either statistical, heuristic or a combination of both. This paper attempts to provide an
overview of the current classification techniques in the context of increasing spatial and spectral resolution.
INTRODUCTION
Objective of Classification is to transform continuous data into
categorical information classes describing the landscape, which
can be used for decision making for effective management of
natural resources. In the following sections an overview of the
classification methods is provided.
. 1. WHY CLASSIFICATION IS NEEDED?
Classification is needed to convert a multitude of data into
certain meaningful number of labels so that we can make sense
of the environment from which the data has come. The methods
used for classification depend on the types of classes that are
sought to be identified, the resolution of the data (spectral &
spatial), the need for crisp or fuzzy classes, separability
between the classes and the knowledge about the distribution of
the classes and the tolerable degree of penalty or loss associated
with misclassification etc.
2. WHATIS CLASSIFICATION?
Classification in the context of remotely sensed data is to
"link" each pixel in the image to one or more user defined
labels, so that the radiometric information contained in the
image is converted to thematic information, like vegetation,
water, built up etc. "Link" is a mapping function which
constructs a linkage between the raw data and user defined
label set, Fig 1.
Remote Sensin = : J fined
Hu g Classifier User De fined
magery Label Set
0 Lg i
Pasture
E J T3 Water
eer x ii NS sucer T] Bare Soil
: = its [7] Forest
Figure 1. Process of Classification [1]
If the mapping function is a classification technique/algorithm
through which each pixel is mapped to a single label, it is *one-
to-one" mapping. Classifiers, which perform “one-to-one”
mapping are called hard or crisp classifiers.
It is also possible to perform "one-to-many" mapping; in this
case, each pixel is associated with more than one label, with
differing degrees of association between the pixel and each
label and the degree of association is expressed as probabilities
of membership. Classifiers that perform “one-to-many”
mapping are called as soft classifiers.
Before moving to classification techniques it is pertinent to
mention about evolution of space borne sensor systems over the
years.
3. EVOLUTION OF SENSOR SYSTEMS
Over the years there has been steady improvement in terms of
spatial, spectral, temporal resolutions and data volumes, Fig 2,
3 &A4.
A 1 2
um mm i}
fs SN EEE
US IKONOS 2 p c (O.8- 1)
De
KVR- 1000 Ru
IRS-1€: c - LLLI Lx (24
DN
SE
OH eese er Ts (Om):
ges E
SPCY
SEIN
Figure 2. Spatial resolution evolution [10]
Devxaar resp
Data Volume (MB)
The pr
combir
analyst
the dec
are sig
sample
charact
The cl
followi
ASSI
dist
Ki