K figures and
tized as line
ized as closed
Lfferent soft-
onversion. We
Graphics, and
the attribute
from two dif-
luinet, 1983).
;ext-node num-
ied by a text
tion is cata-
lg to its text
ir labels are
ilygons on the
0 contain the
then its text
placed outside
the polygon to
:o raster con-
>r a text node
scognize those
ted with them;
lygon encloses
ch belongs to
e ambiguities,
digital map by
outside their
1 removing the
•aster conver-
5 is a compact
: maps and the
paper forest
en digitizing
.1 exist today
Ltized. These
updated using
ing images.
all rivers are
on the cover
its have sizes
Small roads or
not appear in
nentation will
S.
; forest cover
ain time; for
images may not
given time.
to have the
selected poly
clusters, and
mputer program
artise. This
a particular
the problems
ty at solving
:ing knowledge
into computer
re this human
ae more diffi-
rce management
olutions. By
dge, one hopes
licability to
ystems can be
used to plan the acquisition and processing of
remotely sensed data. It is not uncommon for
tracking stations to be acquiring data from several
satellites, some of which have complex configur
ations and control. Expert systems can contribute
also to the analysis of remote sensing images, as
discussed in section 5.2, and to the integration of
geographic information. For specific application
areas, such as crop or forest inventories, expert
systems can be effective in guiding inexperienced
users through complex resource management
procedures.
CCRS has been conducting research and development
in artificial intelligence in order to create a
novice advisor, the Analyst Advisor, described in
section 5.2 and an expert system for comparing two
maps and maps and images (section 5.3). This devel
opment has taken place on DEC AI VAXstations. With
a large investment in FORTRAN code on three other
computers, we chose to retain FORTRAN for numeric
processing and to use Prolog for symbolic reasoning.
Figure 9 shows schematically the general architec
ture of a GIS/lAS combination with an expert system.
Since often the GIS is a commercial system for which
the user has little access to the software inter
nals, we show the expert system more intimately
connected to the image analysis system. The user's
knowledge, in the form of facts, rules, and heuris
tics, is used by the expert system to produce
advice, and control, of the IAS.
5.2 The Analyst Advisor
The Analyst Advisor is a hierarchical expert system
(Figure 10) . Such a structure was selected because
there was a large variety of expertise and image
processing techniques to be applied to a particular
application. This approach is highly structured and
modular. There are a number of levels of authority
organized in a pyramidal structure. The highest
level sets broad goals for the next level of com
mand. The lowest level corresponds to the image
processing algorithms, the FORTRAN programs. With
this hierarchy, there are well-defined lines of
communication and a limited span of command.
The top level expert (Figure 10) is also referred
to as the Analyst Advisor. It guides the user
through an analysis session. To simplify our
development, we have focused on the application of
updating forest inventory geographic information
systems. In several boxes of Figure 10, we identify
the expert system by name and include reference to
the LDIAS Task Interface (LTl). The structure of
our FORTRAN software with its unique man-machine
interface enables our expert systems to control the
image processing software.
the earth, altitude above some geoid model surface,
and so forth. These values for a slot can be
obtained by a computational procedure (demon).
Rules are in the form of production rules ("if
CONDITION then CONCLUSION"). There are two control
strategies that can be followed: backward-chaining,
and forward-chaining.
The Analyst Advisor runs on an AI VAXstation and
controls a user workstation. Since the components
of this workstation include image and map displays
attached to different computers, the Analyst Advisor
must control processes across a network of compu
ters. For the grouping of training areas and clas
ses we have made use of evidential reasoning with
uncertainty values derived from the confusion
matrices. In an environment such as ours where one
is analyzing complete thematic mapper scenes, one
must be able to pause the hierarchical expert system
and later resume an analysis session. This problem
has been solved. One can obtain explanations from
several expert systems. The second prototype of the
Analyst Advisor is scheduled to be released to our
users in March, 1987. Experimentation with the
Analyst Advisor will lead to the inclusion of more
knowledge for forestry applications. In the next
year, the MICE expert system described in the next
section will be integrated with the Analyst
Advisor.
5.3 The MICE Knowledge-Based System
As has been shown, there must be a correct geometric
correspondence or registration between the digital
maps (or geographic information system levels) and
the images. This correspondence criterion also
applies to the maps used as a base for geometric
correction of the remote sensing imagery. The pres
ence of non-systematic errors, as shown in section
4, necessitates that one use symbolic reasoning to
identify and reason about feature mis-matches. The
Map/image Congruency Evaluation (MICE) knowledge-
based system is designed to determine the spatial
differences or similarities of maps or maps and
images. The MICE expert system is described in
detail in Plunkett et al (1986).
The architecture of the MICE expert system is
shown in Figure 11. It makes use of the same shell
(inference engine) as the Analyst Advisor. Certain
classes are better for congruency evaluation than
others. Table 6 lists the classes in order of rank
(1 is highest) for matching. The best class is the
hydrography class, which includes rivers, lakes,
shorelines, and creeks.
Table 6. Best classes for congruency evaluation
For each expert we use the same Prolog shell, our
remote sensing shell, RESHELL. A more complete
description of the Analyst Advisor and RESHELL
appear in Goodenough et al (1987) and Goldberg et al
(1985). RESHELL uses a blackboard structure for
storing known and deduced information. On this
blackboard are stored the agenda created by the
meta-rule interpreter, goals from a higher level
expert, and object values. The shell includes an
explanation facility, a scheduler (inference en
gine), a knowledge editor, and a frames database.
The frames database contains more static and global
information and stores these in a manner that pre
serves descriptive, semantic relationships between
objects. For example, a frame called LANDSAT-5
would have attributes (slots) for altitude, acqui
sition frequency, sensors, etc. The value that one
would retrieve from a slot would depend upon the
perspective being used. Thus, one might retrieve
for altitude, for example, altitude directly above
Class Code
Class Description MICE Rank
G
D
E
I
H
C
B
A
F
J
Hydrography
1
Highest
Road and Rail
2
Utility
3
Land Cover
4
Hypsography
5
Structures
6
Buildings
7
Designated Areas
8
Delimiters
9
Text
10
Lowest
The MICE expert system follows the same strategy
as a human; namely, 1) preprocesses the map and the
image to the same spatial datum and symbolic repre
sentation; 2) locates corresponding segments; 3)
reports on the spatial congruency of the corres
ponding segments or structures. Presently,