(3) The application domain knowledge. 
It provides a model, linking man defined attributes 
to the physical properties measured by remote sensing. 
One important aspect of this sort of models is that only a 
measure of performance needs to be defined such that the 
system can modify its strategy during processing 
(optimisation algorithms, model inversion, parameter 
estimation ) in order to reach a mimimum cost solution. 
For systems with a well defined goal, performance 
functions can be defined that measure the distance from 
the existing state to the goal state at any point in time. 
Systematic search leads then to the finding of an optimal 
the path towards that goal. 
4. REASONING 
From the previous section, we have defined the 
domain of knowledge and representation method for the 
system. In this section, we will be concerned with the order 
in which rules are selected. 
Usually, the reasoning in a knowledge based system 
is done at two levels. 
1. Analysis plan generation first reasons out an 
appropriate plan to guide the analysis of a given image. 
The reasoning engine uses characteristics of the image 
and knowledge about standard image analysis processes 
to generate the plan. 
2. procedure selection and parameter adjustment. The 
reasoning at this level instantiates the analysis plan into a 
special subgoal. Procedures are selected and optimal 
parameter values are determined. If the derivatives of cost 
functions to parameters or procedure selection are not 
available then selections are done through trial and error, 
i.e. the system performs image analysis by applying 
promising procedures, and evaluates the analysis results 
for the discovery of a minimum cost solution. The following 
example describes this situation. 
  
Out 
— Smoothing = Labelling 
in Ti2 113 T4 
EPRT Evaluation 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 3. the search path for optimal merging 
Firstly, the system produces a process sequence for 
merging region as following steps: 
(1) smoothing the image using an edge 
preserving filter 
(2) assigning labels to connected components 
(3) measuring the length of perimeter of a region 
(the number of boundary elements). 
(4) producing evaluate the cost of alternative 
512 
merging of regions and select the minnimum cost 
strategy. 
After plan generation, some procedures and 
parameters must be selected. In step (1), the operator of 
edge-preserving-smoothing is selected. The following 
example depicts the situation: 
RULE 111: 
IF noise must be reduced AND edge must be 
preserved 
THEN run EPRT 
Where EPRT is the name of the edge preserving 
operators in SPIDER. 
In step (2), although the system does not need to 
choose the labelling method, a threshold has to be 
defined. We choose the adaptive threshold option[8] for 
region labelling. The thresholding proceeds as follows: 
Given a grid size N, the input image is divided into NxN 
windows. For each NxN window, the statistics (average 
and standard deviation) within the windows are calculated. 
If the standard deviation within the window is smaller than 
the standard deviation of some background patch, then 
there is no object within that windows; if the deviation is 
greater, then we label the object. 
RULE 120: 
IF St.DeV.rindows >= St.DeV.pack. 
In (4) the system measures the length of perimeter of 
each region P, then measures the common length W 
between two regions (R1,R2) on which the difference of 
value is less than a thresholding value 61, and the 
common length B between regions. Rule 130 131 will 
merge such two adjacent regions iteratively. 
RULE 130 
IF: 1 the difference over a border[9] is LOW 
2 W/B > @1 
THEN: merge R1,R2 
RULE 131 
IF: W/min (P1,P2) »61 
THEN: merge R1,R2 
When more than 2 rules can be used , we use 
specificity ordering. It means that the more conditions a 
rule has, the higher matching priority it has. 
In the system, we apply the depth-first search for 
forming a processing strategy. We arrange the most 
promising potential solution for each sort of process as a 
default path. 
It is assumed that the definitions of initial states, 
procedures and goals are all fixed, thus determining a 
search space; the question then is how to search the given 
space efficiently. The techniques for doing so usually 
require additional information about the properties of the 
specific problem domain beyond that which is built into the 
state and procedure definition. Information of this sort is 
heuristic information. The measure by which the promise 
of a node is estimated is called an evaluation function. The 
nex 
min 
par: 
pro 
pro 
ove 
affe 
to € 
prot 
the 
stru 
the 
into 
app 
prol 
rest 
use 
sys! 
laye 
ther 
stat 
Trac 
exp 
type 
to s 
hum 
in in 
prio: 
orge 
proc 
syst 
to d 
for i 
ima( 
for 
dev: 
time 
repr 
kno 
proc 
to bi 
next 
the | 
syst 
func 
func 
bein 
stral 
Scer