region, using knowledge base. These categories are the
sub-classes of topographical map factors to be extracted.
For example, settlement places are divided into asphalt
and concrete roof settlement places. The third step is to
merge the sub-classes of same map factors into one. The
reseach scheme is shown in Fig. 1.
The following sections are devoted to the production
of texture energy image, the establishment of knowledge
base and the decision making of classification. Principal
component transformation and unsupervised classification
algorithms are presented in relevant documents in detail.
They will not be discussed here.
PRODUCTION OF TEXTURAL IMAGE
The Production of texture image is to enhance the
image data of a band selected from multispectral images
using rule-based localized enhance technique(Zhang et al,
1991), and based on this to generate textural image
using a texture energy technique.
The texture energy technique was presented by Laws
(Wechsler,1980) . The idea of this technique is similar to
that of Fourier transformation, but it is implemented
in space domain. Two steps are required for the
implementation of this technique.
The first step is to compute the convolution of the
given image f(i,j) with mini-mask h(s, p) (e.g., Lapla-
cian and gradiant operator, generally 3 x 3 or 5x 5) to
detect the pixels on edges:
N-1 N-1
gi, = E, E f(i-s,j-p) hG. p) (1)
where N is the size of the mini-mask.
The second step is to abstract texture energy in a
larger size window, which is defined as the square sum
of grey levels within the window. The sum of absolute
value is used to substitute for the square sum for the
convenience of computation, then
GG, - EE |8G-k,i-D) Q)
where M is the size of the window. Generally M —9, 15
or 21.
In essence, the texture energy algorithm is to find out
the pixels on edges and to regard the statistical signal
energy as the apparent feature of textural signal.
ESTABLISHMENT OF KNOWLEDGE BASE
In order to use knowledge for automated extraction
of map factors, a knowledge base must be established.
Two things must be done: one is to research the spectral
and textural properties of objects, the other is to repre-
sent the inherent and interrelative features of objects.
Each category of objects has its inherent spectral and
textural properties. The properties can be acquired with
visual interpretation experiences and the spectral
reflectance curves of objects that have been known. In
the following paragraph, the spectral properties of 4
88
ground objects(ie.,vegetation, soil, water and buildings)
are discussed simply.
Green vegetation. has a low reflectance in the
0.35-to-0.7 um region with a slight rise at 0.554 m, and
a higher reflectance in the near-infrared(0.75 — 1.354 m),
but low reflectances at 14, 19 and 2.7um. The
reflectance curve of soil is less complex than that of veg-
etation. The reflectance of soil generally increases with
the increasing of wavelength, especially in the
0.35-to-1.4#m region. Because soil is a complex mix-
ture, and has various types, the spectral reflectance fea-
tures of which are different. Water has a low reflectance
in the visible band and a tendency of decreasing in the
near-infrared. Asphalt and concrete roofs of buildings
have the reflectance profile typical of man-made materi-
als, ie., a generally increasing reflectance with increasing
wavelength. The reflectance varies with the age of the
building. Roofs of natural materials have the same
reflectance as the natural materials. All above can be
found in relevant documents.
Besides spectral properties for object interpretation in
images, the textural properties are very important. For
example, soil and buildings have the similar spectral
properties, i.e.,a generally increasing reflectance with
increasing wavelength, but their textural properties are
quite different.
In the knowledge base we have established, there are two
kinds of knowledge, i.¢., band-to-band (BB) and category-
to-texture (CT) ,which are expressed in terms of rules and
weights. A rule is a comparision between band and band or
texture and threshold. A weight is a supporting or
opposing evidence weight when the rule is right or false.
A knowledge sub-base for a certain category is com-
posed of a set of rules and weights. For example, the
knowledge sub-base for the forest interpretation in TM
image is composed of
rule 101: band(4)>sum (1,2,3) [50, 0]
rule 103: band(2)> band (1) [10, 10]
rule 106: min(4,5,6)<sum (1,2,3) [0, 50]
rule 114: band(0) (texture image) « 20 [20, 10]
where band(i) stands for the value of band i, sum( )
denotes the sum of the bands in parentheses, min( ) de-
notes the minimum of the bands in parentheses, the first
and second number in brackets| ] indicate the sup-
porting and opposing evidence weight, respectively. The
minimum weight is 0, and maximum weight is 100. A
rule is a constraint in essence. According to the roles
they play in category interpretation, constraints are di-
vided into mandatory constraints, optional constraints
and contradictory constraints. Mandatory constraints are
represented by having the supporting and opposing
weights greater than zero so that evidence is added for
fh no 73, 6 t" mm Ww 3 0 al
eum ph DÀ) 0M
ld