KNOWLEDGE BASED CLASSIFICATION OF LANDSCAPE OBJECTS COMBINING
SATELLITE AND ANCILLARY DATA
Jerzy Chmiel, doctoral student, Institute of Photogrammetry and Cartography, Warsaw University of Technology, Poland, and
Thomas Gumbricht, doctoral student, Division of Land and Water Resources, Royal Institute of Technology, Sweden.
KEY WORDS: Expert System, Knowledge Base, GIS, Integrated image classification, Fuzzy Logic.
ABSTRACT
The land surface pattern is related to both natural and anthropogenic processes, for which domain experts have developed
corresponding semantic. Traditional image classification is based on statistical relations, disregarding qualitative relations between
processes and patterns. The article presents an image classification system integrating remote sensing and georeferenced data
knowledge rules inferred via a simple and transparent expert system. Performance was tested against traditional maximum
likelihood classification. The expert system classification gave the best results. It is concluded that_simple and transparent expert
modelling can enhance understanding of spatial relations between processes and patterns, but that accuracy in georeference is
crucial for inference of expert rules.
1 INTRODUCTION
The land surface is a non-random structure. The textural and
structural pattern of both the natural and the cultural landscape
have process derived logic (Ripl and Gumbricht, 1996).
Regolith, wetness, vegetation and e.g. infrastructure are
strongly interconnected and site related. Domain experts have
developed corresponding object oriented semantic. However
traditional image classification disregards these relations, and
rely heavily on stochastic probability density functions (pdf)
(cf. Argialis and Harlow, 1990). Categorisation is mostly based
on procedural rules related to pixel-wise multidimensional
vectors (Fig. 1). Classification accuracies have been improved
by advanced statistical data modelling (e.g. Franklin and
Peddle, 1989; Lauver and Whistler, 1993), by integration of
multitemporal or ancillary data (e.g. Middelkoop and Jansen,
1991), and by multisource field data (e.g. Wu et al., 1988;
Congalton et al, 1993; Fiorella and Ripple, 1993; Zeff and
Merry, 1993). Digital elevation models (DEM) have been most
widely employed. Inter alia used for correction of reflectance
because of inclination, stratification before classification, for
assigning a priori probabilities during classification, and for
post processing of problematic classes.
Present developments in image classification include expert
System integration, and the application of neural networks.
Neural networks combine relation knowledge in hierarchical
nodes, often by an inductive, backward driven iterating
process. Reported successful applications include land cover
classification (Hepner et al., 1990; Civco, 1993; Dreyer, 1993)
and shoreline extraction (Ryan et al., 1991). However so far
traditional classification methods are reported to perform
equally well. Expert systems can be either inductive and
backward driven, or deductive and forward (or data) driven. A
recent trend has been to develop simple and transparent expert
Systems for integrated image classification, and domain expert
languages for imprecise (fuzzy) knowledge inference (e.g.
Leung and Leung, 1993; Wang, 1994). By combination with
statistical discrimination classification rules can be
transparently compacted (Srinisavan and Richards, 1990;
183
Dymond and Luckman, 1994). The strong relation between
landscape processes and patterns suggest that declarative
knowledge rules should be powerful for integrated image
classification (e.g. Skidmore et al., 1991; Gumbricht et al,
1995).
Data structure and integration methodology Indicators
simple advanced (examples)
Level of integration
Boolean classes | | Geomorphometry Texture
Signatur RS signals PCA, CCA, ratios| (color, slope, NDVI)
Signature
interpreation
Boolean Thematic| | Thematic with ^
Feature membership |
| Size, shape, pattern
(fragmentation,
flow length)
Image regions
Bool. boundaries
Objects
Fuzzy boundaries
Figure 1. Simplified scheme for image classification
methodologies (modified after Gumbricht et al., 1995).
This article presents a compact expert system for knowledge
based image classification combining procedural and
declarative knowledge representation. Its performance was
compared with traditional maximum likelihood classification.
The aim of the study is to define and evaluate object oriented
classifications of the landscape pattern with relevance for
functional management (cf. Worboys, 1994). The study is part
of a larger Swedish-Polish research program aiming at defining
and modelling sustainable landscape management and
restitution (cf. Gumbricht, 1995).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
