1086
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
3. KNOWLEDGE BASED EXPERT SYSTEMS
STRATEGIES
Two fundamental extreme types of inference mechanisms are
utilised in Artificial Intelligence :
- Domain Dependent Inference (DPI): is suitable for
relatively explicit body of evidence in which the
conclusion is entailed and no, or low level of inference
is required. DDI inference mainly controls the
combination order of the different evidences and is
represented in the procedural knowledge. The
sophistication in expert systems based on DDI is
embedded within them during their rules’ formation
rather than in their functioning way. The construction of
procedural knowledge and evidential basis requires
heavy information analysis, learning procedures and
feature assessment. Rule based expert systems
implementing binary decision trees are extreme
examples of DDI and are widely used in remote sensing
(e.g., Goodenough et al.,1994; Kartikeyan et. Al.,1995;
Chan et al.,1999). GeoAIDA (Tonjes et al.,1999) is a
good example of such strategy, where expert system
based on semantic networks was developed
implementing specific (explicit) evidence, including
contextual information. The inference mechanism
applied was then a procedural sequential decision tree
type. Other extremely different example of this strategy
concerns the use of Classification Tree Analysis (CTA)
techniques and their recent improvements with Boosting
and Bagging methods (e.g., Lawrence et. al.,2004).
Again in this method there is maximal exploitation of
the information content in the data and the
(computerized) construction of a decision tree
specifically (explicitly) applicable to the data at hand.
- Domain Independent Inference (DII): is suitable for
relatively implicit body of evidence. The DII inference
relate to their associated characteristics: the relative
belief, support, certainty and weight, rather the
information sources themselves and thus is independent
of any specific recognition or decision making problem.
The combination of evidence would then be based on
general deductive, inductive or abductive procedures
(Durkin, 1994). A significant element of the intelligence
represented by DII is embedded within it during the
development of the generalized inference algorithm
which is independent of any specific recognition or
decision making problem, and usually adopted without
investing in such development. The sophistication of the
system which is embedded in the inference
methodology is then dependent on the inference
capabilities to overcome non-uniqueness and conflicts
(e.g., Cohen and Shoshany,2005).
In reality there is rather a mixture of strategies tailored
according to information sources and expert systems tools
available. However, there is relatively little comparison made
between DDI and DII based strategies in terms of their
performance and the types of evidences used. Performance
assessment of these strategies in areas which differ from the
training areas is of special interest. Would it be better in terms
of 'cost' and performance to deepen the search for domain
explicit evidence or to broaden the implicit evidential basis?
These same questions are addressed in "Information Foraging".
4. THE INFORMATION FORAGING PERSPECTIVE
Chamov,(1976) and Stephens and Kerbs,(1995) proposed the
foraging theory to explain foraging behavior and strategies in
nature. This theory is based on the observation that animals
evaluate the availability of food sources and at the same time
the efforts required in order to consume them. Prey searching
efforts, competition with other predators and energy required to
catch the prey are among the main ' costs’ of obtaining the
food. The amount of food (prey size) gained and its quality
represents the gain from these efforts. 'Within Patch' searching
strategies concern exploiting the food available at a certain area
before migrating to another patch. While such strategy
minimizes the search and migration energy, there is decrease in
food availability and maybe increase competition. In the remote
sensing context such strategy would be represented by limiting
the information search for one data source: visible, IR or
microwave spectral bands. Domain Dependent Inference (DDI)
strategies are also a type of 'Within Patch' foraging behavior.
'Between Patch’ strategies represent large area search for
patches of high food availability or low competition. Its
implementation requires gathering of larger amounts of data
and better skills in optimizing their use. Animals adopting
these strategies will have higher resilience to changing
conditions and would reduce the potential degradation from
overgrazing. In the remote sensing context 'Between Patch'
strategies would be characterized by implementing multiple
sources (e.g., multi-spectral; multi-temporal; multi-resolution).
The uses of Domain Independent Inference (DII) strategies are
analogous to this later animal behavior.
'Translating' the foraging theory into the information and
knowledge extraction arena required quantitative treatment of
the information gained from the efforts made for searching it.
Such treatment is crucial for assessing the relative performance
obtained from the implementation of the different strategies.
Pirolli and Card,(1999) provided the following simplified
expression for assessing relationships between the time invested
in producing the information items and the information gain.
G = gT b /t b (1)
Where, T b is the total amount of processing time, t b is the
average processing time per information item and g is the
average information gain per item.
In the next section the treatment of information gain from
developing and implementing expert systems in remote sensing
is presented.
5. ESTIMATING RECOGNITION ENERGY AND GAIN
FROM INFERENCE
The complexity of the recognition task is determined by the
level of confusion between the different surface phenomena as
their appear in the multi-temporal and /or multi-spectral and or
multi-resolution feature space. In other words: highly separable
classes would not require much work in recognizing them and
vice versa. There are numerous unsupervised classification
algorithms which may facilitate 'automatic' determination of
classes in the feature space. The inherent separability which
exists inbetween these classes is inversely related to the "effort"
needed to be invested in constructing the knowledge base.