Gary Priestnall 
  
SEMI-AUTOMATED LINEAR FEATURE EXTRACTION USING A KNOWLEDGE RICH 
OBJECT DATA MODEL 
Gary Priestnall!, and Steve Wallace? 
! University of Nottingham, UK. 
School of Geography. 
* Defence Evaluation and Research Agency, UK 
Space Department. 
gary.priestnall ? nottingham.ac.uk 
sjwallace @dera.gov.uk 
KEY WORDS: Linear Feature Extraction, context, object oriented data model 
ABSTRACT 
This paper describes an approach to the extraction and classification of linear features from imagery using an object 
oriented data model as a framework to store contextual knowledge. The work reported is the foundation of a three year 
project, funded by the UK Ministry of Defence, named Automatic Linear Feature Identification and Extraction 
(ALFIE). The project team comprises the UK Defence Evaluation and Research Agency (DERA), the University of 
Nottingham, and Laser-Scan Limited; DERA is leading the work. The research is driven from a requirement for rapid 
database generation for Synthetic Natural Environment (SNE) applications, although the potential for the research is m 
limited to these applications. The linear objects under investigation are described in detail in terms of the properties 
used to differentiate feature classes. The interrelationships between feature classes as observed in imagery are 
characterised in the form of a matrix. This will form the basis for an object schema that will allow contextual 
knowledge such as regional containment and the relationships between local neighbouring features to be modelled. Th 
overall role of context in a strategy for geographical linear feature extraction is discussed and certain parallels made to 
approaches investigated in non-geographical domains. 
1 INTRODUCTION 
The work reported is the foundation of a three year project, funded by the UK Ministry of Defence, named Automatic 
Linear Feature Identification and Extraction (ALFIE). The project team comprises the UK Defence Evaluation and 
Research Agency (DERA), the University of Nottingham and Laser-Scan Limited; DERA is leading the work. The 
requirement for this research project was identified as a result of research into rapid Synthetic Terrain Database 
Generation. A prerequisite for this is the ready availability of suitable raw Synthetic Natural Environment (SNE) data 
(terrain elevation, feature and attribute data) for any area in the world that may need to modelled. Up to date and 
accurate source information at the required resolution is likely only to be available from remotely sensed imagery. 
This offers a range of geographical information from regional contextual knowledge provided by low resolution multi 
spectral imagery, to the large scale detail offered by digital ortho-photography and very high resolution satellite senso 
However, the abundance and detailed content of this imagery will remain inaccessible unless the information content 
can be readily extracted through automation. Tools to automate stages of the process of linear feature extraction from 
imagery are required. An approach to feature extraction is presented which uses an object-oriented data model as the 
framework to store contextual knowledge at a number of levels in an attempt to discriminate types of linear features and 
to reconstruct networks. 
This paper considers the nature of the problem of extracting linear objects from imagery focussing on the use of 
contextual knowledge. The generic properties of the linear objects under investigation are presented which form the 
basis for defining the structure of an object schema. The interrelationships between objects are considered, to offer à 
framework for incorporating knowledge of the place within a scene of an object, in terms of both the context of its 
surrounding region and the local juxtaposition to other objects. Finally the role of context in an overall strategy for 
object recognition from imagery is discussed. 
  
740 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
The 
clue 
in a 
ima 
inte 
ope! 
proj 
fron 
Lin 
repr 
repr 
rast 
bec 
The 
prol 
in re 
eng 
of d 
stan 
insp 
prec 
assu 
com 
idea 
dom 
cont 
the | 
this 
cert: 
obje 
the | 
31 
The 
left 
mor 
mor 
  
The 
kno 
inte; 
ima,