Raster Image Data: 
satellite data 
scanned aerial photographs 
Expert Knowledge: 
concepts end methods 
from remote sensing, geo-science 
and Image processing 
Known Geographic Information: 
spetiai data 
factual data 
Figure 1: Remote sensing as a knowledge-based analysis task 
X DEVELOPING AN EXPERT SYSTEM TO 
ASSIST REMOTE SENSING TASKS 
Instead of developing new data processing methods in 
remote sensing, the RESEDA project is exploring exi 
sting operational methods in state-of-the-art remote sen 
sing projects. Our objectives are to embed these methods 
into a consistent theoretical framework and to represent 
them in a knowledge base on a computer system. To this 
end, we are developing an intelligent advisory system, 
called the RESEDA Assistant, which makes use of this 
software knowledge base in order to assist the user in 
efficiently performing remote sensing tasks. 
The main purpose of the RESEDA Assistant is to plan the 
sequence of computations necessary for a given analysis. 
For that purpose, the user specifies the available data and 
the desired data. The available data may include sensor 
data and some additional geographic information repre 
sented in a resident database or to be entered by the user. 
From these specifications, the RESEDA Assistant com 
putes a set of processing plans, that is, sequences of 
computations. These computations may consist of ima 
ge-processing operators, statistical evaluations, or mani 
pulations of spatial data. In the first stage of the RESEDA 
project, the computed processing plans will be printed out 
by a prototype system called the RESEDA Advisor and 
will have to be executed explicitly by the user. The final 
system, the RESEDA Assistant, will display a set of 
alternative processing plans as a menu. After selecting 
one of the menu items, the respective processing plan will 
be executed automatically by a plan interpreter. In this 
context, automatic execution means that the appropriate 
data processing programs are called with correct argu 
ments in correct sequence; the programs being called 
may nonetheless run in interactive mode and require 
intervention by the user, such as digitizing training areas 
for a supervised classification. Serving as a knowledge- 
based interface to traditional analysis software, the RESEDA 
Assistant resembles to the Analyst Advisor developed at 
the CCRS in Ottawa (Goodenough, 1987) that interfaces 
the LDIAS data processing system. However, the con 
cepts of these systems differ, since RESEDA focuses 
rather on the methodological knowledge of a remote 
sensing expert than on the facilities offered by a given 
data analysis system. 
The global control strategy of the RESEDA expert sy 
stem is backward chaining. That is, the system first asks 
for the desired information. The desired information may 
be described by its format (e.g., image, map, or factual 
data), its accuracy (i.e., qualitative or quantitative results 
required), and the subject of interest (i.e., which target 
classes or target attributes are to be recognized). The 
system tries to activate processing models that are able to 
compute the desired information from other data. If this 
data is not present, the system recursively tries to activate 
processing models for computing the data. The process 
stops when the system tries to determine data items that 
are explicitly labeled as so-called primary data. In this 
case, the user is asked whether the primary data in 
question is available. If the user says no, the system tracks 
back and tries to activate different processing models. 
The advantage of the backward-chaining strategy is that 
the user is asked for available data only if this is effecti 
vely necessary. An alternate design idea, which is not to 
ask for the available data at all, was rejected, since it 
would have required the implementation of a complete 
data management system. This was found not to be 
feasible because of the great number of foreign data 
sources that can be made available on demand, such as 
remote sensing data, geographic databases, raster-scan 
ned maps, or even printed maps to be digitized manually 
by the user.