1085
'7. Beijing 2008
KNOWLEDGE BASED EXPERT SYSTEMS IN REMOTE SENSING TASKS:
QUANTIFYING GAIN FROM INTELLIGENT INFERENCE
M. Shoshany
Geoinformation Engineering, Faculty of Civil & Environmental Engineering, Technion, Israel Institute of Technology,
Haifa 32000, Israel - maximsh@tx.technion.ac.il
Commission VII, WG VII/6
KEY WORDS: Knowledge Base, Reasoning, Theory, Fusion, Classification, Data Mining
ABSTRACT:
How to measure gain from the use of intelligent inference ? How can the complexity of the recognition/ classification task be
estimated? What is the type of evidence which best suits an .inference mechanism? These questions are addressed here in their
theoretical and methodological aspects. Their practical implications are demonstrated with 'real' crop mapping task. For this
purpose, simple rule-based system is compared with expert system based on Dempster - Shafer evidential reasoning algorithm.
The advantage of using 'soft' / implicit evidence with the Dempster-Shafer algorithm over the use of 'hard' / explicit evidence with
decision - tree type procedure is discussed.
1. INTRODUCTION
Spatial, temporal and spectral complexity of remote sensing
recognition tasks necessitates the use of Knowledge Based
Expert Systems (KBES). Such complexity concern the fact that
the same surface phenomenon may emerge in different ways in
imagery sources (see discussion for example in Lu and
Weng,(2007).). These systems facilitate algorithmic
adjustments of the use of classification or recognition rules
according to the local context. KBES combine available
evidence, procedural knowledge regarding priorities in
implementing evidences in parallel or sequentially; and an
inference mechanism. The procedural knowledge may be
constructed based on human expert knowledge or through
extensive learning of the recognition problem at hand. Data
Mining and Knowledge Discovery are technologies suitable for
conducting such extensive learning. In parallel, there are
numerous inference mechanisms available which shift the
emphasis into designing efficient expert systems considering
different strategies for information extraction and processing.
Such strategies also concern the efforts made in the search for
evidence on the one hand and the information gain from using
this evidence on the other hand. Recently, there is growing
interest in the field of information seeking and utilization with
the adoption of Evolutionary - Ecological Models of Foraging
(Pirolli and Card, 1999). Comparison between strategies based
on assessment of information gain versus information cost is a
central element in the Information Foraging Theory. A primary
problem in adopting this theory in remote sensing concern
quantitative evaluation of gain from evidence versus gain from
inference (Shoshany and Cohen,2007). The presentation in the
conference would be divided into three elements: review of
fundamental terms, concepts and strategies in the
implementation of KBES, assessment of the similarity between
information foraging and remote sensing tasks; and finally,
discussion of the ways to estimate gain from evidence and
inference in mapping missions.
2. FUNDAMENTAL TERMINOLOGY
First, it is important to recognize that at the initial stage of the
recognition task we are dealing with propositions, or better say
multiple propositions for each recognition task. Evidence is
then defined (Wikipedia) as “any objectively observable or
demonstrable circumstance which tends to indicate or disprove
a proposition". Two main types of evidence can be discerned:
Explicit Evidence: refers to the "notion of plain distinct
expression that leaves no need to infer" (Merriam-Webster).
In the remote-sensing narrow band signatures of specific
materials and very unique combinations of spectral
reflectance thresholds are examples of explicit evidence.
The production of such evidence would typically require
exhaustive search (see methods proposed by Peddle and
Ferguson,2002).
Implicit Evidence: type of proposition represents a
conclusion "capable of being understood from something
else though unexpressed: capable of being inferred"
(Merriam-Webster).
Implicit evidence would consist of generalizations,
associations and contextual information representing some
level of non-uniqueness and conflicts with reference to the
object subject to recognition (e.g., Cohen and
Shoshany,2005). While the production of such evidence
might be relatively easy, its use would require inference
capable of overcoming such non-uniqueness and conflicts.
Considering that remote sensing evidence would be usually
located on a continuous scale inbetween the extreme explicit to
extreme implicitness, it is clear that the role of inference is
elementary. Inference is defined (Merriam-Webster) as "the
act of passing from one or more propositions, statements, or
judgments considered as true to another the truth of which is
believed to follow from that of the former". Deductive,
inductive, abductive, analogical or common-sense reasoning
would facilitate such inference (Durkin, 1994). The selection of
type of evidence and inference is a crucial step in developing
KBES strategies.