ethod for
mensional
nal hyper-
asses were
» sufficient
asses were
acy of par-
percents,
extracted
o the case
classes in
Lhe perfor-
‚tion accu-
n that the
ys optimal
designing
bands for
tive infor-
mensional
ixtraction,
M ultispec-
1123-1131.
istics, Wi-
> Process-
mic Press,
> Sensing:
58-361.
ature Ex-
reoscience
yo, Japan,
1atic Clas-
tely Sensed
Alabama,
eighted Fea-
Proc. SPIE,
isible/ In-
>IE, vol.834,
COMPUTER UNDERSTANDING OF SUB-PIXEL LAND COVERS
Jayanta Kumar Ghosh.
Research Scholar. Civil Engineering Department.
University of Roorkee. Roorkee. UP. 247 667. INDIA.
Commission Ill, Working Group 3
KEY WORDS : Knowledge Base, Sub-pixel, Classification, Computer Understanding System,
Automation, Fuzzy Labels, Linguistic Variables.
ABSTRACT
Sub-pixel analysis of satellite data is important for accurate information and estimation of the different land
cover classes. This paper defines a new method for direct classification and estimation of the sub-pixel land
covers. Itis a fuzzy knowledge based approach wherein features (LINGUISTIC VARIABLES) are addressed
by fuzzy labels. This approach can be utilised in developing an image understanding system.
1. INTRODUCTION
In satellite images, land cover classes are classified
with a certain degree of uncertainity, especially
when mixed pixels occur. This is due to their con-
tinuous spatial coverage rather than abrupt and
inter-grade gradually. Thus, mixed pixels occur in a
satellite image either at the boundaries of the cover
types or due to the sub-pixel phenomena (Fisher and
Pathirana, 1990). Previous studies show that differ-
ent classification methods for classifying or assign-
ing labels to pixel may achieve classification accu-
racy greater than 8596 for "pure" pixels butless than
7596 correct classification for regions having mixed
pixels (Metzler and Cicone, 1983). This is due to the
fact that the mixed pixel displays a composite spec-
tral response which may be dissimilar to each of its
component classes (Campbell, 1987). Thus, due to
the inherent presence of mixed pixel, classification
schemes are prone to errors.
To get through the crux, a number of approaches,
basically statistical in nature, have been developed
and tested to unmix the pixels into their constituent
classes (Settle and Drake 1993, Jasinski and
Eagleson 1990, Fisher and Pathirana 1990). Typi-
cally, pixel unmixing is achieved through the appli-
cation of a spectral mixture model. An alternative to
this is to define relationships between a measure of
the strength of class membership, which may be
derived from some image classification routines
and the pixel composition (Foody and Cox 1994).
This paper defines a new method for direct classifi-
cation and estimation of the sub-pixel land cover
classes in addition to pure pixels.
2. PHILOSOPHY OF THE METHOD
The approach outlined in this paper is based on the
premise that classes of objects in which the transi-
tion from membership to non-membership is gradual
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
rather than abrupt, which intuitively correlates with
the spatial distribution of the land cover classes. As
land covers are imprecise in spatial distribution with
reference to IFOV (Instaneous Field Of View) of the
satellite sensors, fuzzy labels (Zadeh, 1973) pro-
vides a better framework of representation of class
information. Generally, the more a pixel contains a
cover class, greater is the proportion of spectral
characteristics of that class in that pixel (Wang,
1990). As the mixture proportion changes from pixel
to pixel, the spectral characteristics will also change.
In fuzzy representation, for remote sensing image
analysis, land cover classes are defined by fuzzy
sets (Zadeh, 1965) where FEATURES are linguistic
variables, image pixels are set elements and the
membership grades attached to a pixel indicate the
extent to which the pixel belong to a certain class/
classes.
As better resolution of images enhances the intrin-
sic heterogeneity (scene noise) of images, conven-
tional classification techniques do not lead to better
results (Townshend, 1980). So an "image under-
standing system" using subtle differences of multi -
spectral responses in synergistic consideration has
been adopted. It is characterised by a priori knowl-
edge of the real world. In determining the a priori
knowledge, the domain knowledge of land covers in
multi-spectral perspective, expert's heuristics and
training area informations are used.
3. METHODOLOGY
The land cover classes are first categorised in a
hierarchical order ( See Appendix A). Then the
working of the understanding system proceed from
top to bottom i.e., from general to particular type of
cover.
The steps involved in the operation of the proposed
understanding system are enumerated as follows:
239
