Research Article
Interactive and visual fuzzy classification of remotely sensed imagery
for exploration of uncertainty
1
ARKO LUCIEER and MENNO-JAN KRAAK
International Institute for Geo-Information Science and Earth Observation
(ITC), Department of Geo-Information Processing (GIP), PO Box 6, 7500 AA
Enschede, The Netherlands; e-mail: lucieer@itc.nl, kraak@itc.nl
(Received 21 October 2002; accepted 18 September 2003 )
Abstract. In this study, we propose, describe, and demonstrate a new geovisua-
lization tool to demonstrate the use of exploratory and interactive visualization
techniques for a visual fuzzy classification of remotely sensed imagery. The
proposed tool uses dynamically linked views, consisting of an image display, a
parallel coordinate plot, a 3D feature space plot, and a classified map with an
uncertainty map. It allows a geoscientist to interact with the parameters of a
fuzzy classification algorithm by visually adjusting fuzzy membership functions
and fuzzy transition zones of land-cover classes. The purpose of this tool is to
improve insight into fuzzy classification of remotely sensed imagery and related
uncertainty. We tested our tool with a visual fuzzy land-cover classification of a
Landsat 7 ETMz image of an area in southern France characterized by objects
with indeterminate boundaries. Good results were obtained with the visual
classifier. Additionally, a focus-group user test of the tool showed that insight
into a fuzzy classification algorithm and classification uncertainty improved
considerably.
1. Introduction
Geospatial data quality is a topic frequently covered in recent scientific
literature on GIS and remote sensing (Foody and Atkinson 2002). With increasing
use of remotely sensed data as input to a GIS, uncertainty in remotely sensed image
classification is also receiving more attention. In recent years, much research has
focused on modelling uncertainty in remotely sensed image classification (van der
Wel 2000, Zhang 2001, Foody and Atkinson 2002). Uncertainty can arise from
poor class definition, mixed pixels, and transition zones (fuzzy boundaries). Both
fuzzy and probabilistic classification techniques can help to model and quantify
uncertainty. Most proprietary GIS and remote-sensing software provide numerous
conventional and advanced classification algorithms. However, most if not all of
these packages do not offer tools to model, visualize, and manage uncertainty in
classifications. Many of these packages can produce good thematic maps, but
International Journal of Geographical Information Science
ISSN 1365-8816 print/ISSN 1362-3087 online # 2004 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/13658810410001658094
1
A web supplement with colour figures and a demo version of the software can be found at
http://parbat.lucieer.net
INT. J. GEOGRAPHICAL INFORMATION SCIENCE
VOL. 18, NO.5,JULY–AUGUST 2004, 491–512

nowadays users want to be presented with information about the quality of these
maps. Uncertainty in classification is an important aspect of the data quality and
usability or ‘fitness for use’ of a classified map. Therefore, modelling and
communication of uncertainty is becoming more important. An appropriate way to
present spatial and thematic behaviour of uncertainty is via visualization. In recent
years, visualization prototypes have been proposed, which focus on presentation
and exploration of uncertainty in a remotely sensed image classification (van der
Wel et al. 1997, Blenkinsop et al. 2000, Bastin et al. 2002). However, none of these
tools provides a way of interacting with the classification algorithm itself. Visual
interaction with the classification algorithm could greatly improve insight into
classification and related uncertainty.
In this paper, we propose, describe, and demonstrate a novel visualization tool,
which allows interaction with a supervised fuzzy classification algorithm. The
objective of this study is to develop and implement a geovisualization tool by which
a geoscientist can interact with the parameters of a fuzzy classification algorithm in
order to gain insight into the working of a fuzzy classification and related
uncertainty, and to possibly refine the classification result. For this study, we
selected a Landsat 7 ETMz image of an area characterized by semi-natural
vegetation types. Transition zones between these vegetation types are known to be
problematic in image classification. We evaluate the prototype with a focus group
user test.
2. Methods
2.1. Fuzzy classification
Supervised image classification is a commonly performed analysis of remotely
sensed data. The result of such a classification is a thematic map with a label for
each pixel of the class with which it has the highest strength of membership. This
hard or crisp classification is based on conventional crisp set theory. A conventional
classification of remotely sensed imagery models the study area as a number of
unique, internally homogeneous classes that are mutually exclusive. However, these
assumptions are often invalid, especially in areas where transition zones and mixed
pixels occur. Land-cover types are rarely internally homogeneous and mutually
exclusive, so classes can seldom be separated by sharp or crisp boundaries, in
feature space as well as geographic space. Furthermore, complex relationships exist
between spectral responses recorded by the sensor and the situation on the ground,
where similar classes, pixels, or objects show varied spectral responses, and similar
spectral responses may relate to dissimilar classes, pixels, or objects. Moreover,
remotely sensed images contain many pixels where boundaries or sub-pixel objects
cause pixel mixing, with several land covers occurring within a single pixel. Finally,
classes are often hard to define, resulting in vagueness and ambiguity in a
classification scheme. Most, if not all, geographical phenomena are poorly defined
to some extent and, therefore, fuzzy set theory as an expression of concepts of
vagueness is an appropriate model for working with remotely sensed imagery
(Fisher 1999, Zhang and Foody 2001). To adapt to the fuzziness characteristic of
many natural phenomena, fuzzy classification approaches have been proposed
(Wang 1990, Foody 1996, Zhang and Foody 2001).
Fuzzy classification is based on the concept of fuzzy sets (Zadeh 1965). In the
fuzzy set model, the class assignment function attributes to each element a grade of
492 A. Lucieer and M.-J. Kraak