ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
INDUCTIVE CLUSTERING: 
AUTOMATING LOW-LEVEL SEGMENTATION IN HIGH RESOLUTION IMAGES’ 
Annie Chen“, Gary Donovan*, Arcot Sowmya®, John Trinder® 
: School of Computer Science and Engineering — (anniec, garyd, sowmya)@cse.unsw.edu.au 
School of Surveying and Spatial Information Systems - j.trinder@unsw.edu.au 
University of New South Wales, Sydney 2052, Australia 
KEY WORDS: Classification, Edge, GIS, High Resolution, Identification, Land Use, Segmentation, Spatial Infrastructures. 
ABSTRACT: 
In this paper we present a new classification technique for segmenting remotely sensed images, based on cluster analysis and machine 
learning. Traditional segmentation techniques which use clustering require human interaction to fine-tune the clustering algorithm 
parameters and select good clusters. Our technique applies inductive learning techniques using C4.5 to learn the parameters and pick 
good clusters automatically. The techniques are demonstrated on level 1 of RAIL, a hierarchical road recognition system we have 
developed. 
1. INTRODUCTION 
Road detection and recognition from remotely sensed imagery is 
an important process in the acquisition and update of 
Geographical Information Systems (GIS). Recent research has 
tended towards automatic approaches such as snakes, and has 
combined additional image information like multiple scales 
(Baumgartner et al, 1999; Baumgartner et al, 2000; Laptev et al, 
2000) and image bands (Desachy et al, 1999). See (Laptev et al, 
2000) for an extensive summary of current road extraction work. 
In our previous papers (Teoh et al, 2000a; Teoh et al 2000b; 
Trinder et al, 1999) we described the RAIL/Recoil/KBRoad 
system, which is a semi-automatic, multi-level, adaptive and 
trainable edge-based road recognition system. RAIL is intended 
to demonstrate the use of various Artificial Intelligence 
approaches in road extraction, particularly for choosing 
parameters. These methods are relevant for most applications, 
even though we only demonstrate them on the RAIL system. 
Recently, we have developed a new learning technique designed 
to automatically combine different algorithms in the best way 
for a given image - inductive clustering. This uses inductive 
learning to improve the results obtained from clustering, by 
learning the optimal clustering parameters for any situation. 
The details of this inductive clustering method are presented in 
this paper. Section 2 gives a brief overview of RAIL, inductive 
learning and clustering. In Section 3, we introduce the framework 
that combines inductive learning with clustering, and discuss the 
experimental components of that framework. Finally, in Section 
4 we present some results from evaluating the framework. 
  
2. BACKGROUND INFORMATION 
Our research builds upon the existing RAIL system, which 
includes components for inductive learning and clustering 
algorithms, amongst other features. RAIL has been used for 
many experiments in the past, including preliminary trials of 
Artificial Intelligence techniques in road-detection; a new 
junction recognition algorithm (Teoh et al, 2000a); and as part of 
an expert system written in PROLOG (Trinder et al, 1998). 
RAIL works by classifying images using different techniques, 
such as inductive learning or various clustering methods 
(described below). Classification is the process of sorting image- 
level features (e.g. edges) into different classes based on their 
attributes. Our classification systems are designed to arrange all 
the road edges into one class. 
2.1 RAIL 
RAIL is a multi-level edge-based road-extraction program, where 
straight edges are detected from a single-spectrum image using 
VISTA's implementation of the Canny operator (Pope et al, 
1994). Level 1 joins pairs of opposite edges together, whilst 
Level 2 links the edge pairs together to make road sections. 
Levels 3 and 4 relate to intersection detection and integration, 
respectively. At each level, a different classification technique 
can be applied. 
At level 1, the objective is to join matching edges into edge pairs, 
or road segments. A road segment can be thought of as a pair of 
edges which are part of a road, and oppose each other. The 
attributes, or properties, of such a road segment at RAIL Level 
1 are listed in Table 1. 
" A preliminary version of this paper (Chen, 2002) was presented at ICML2002, Machine Learning in Computer Vision Workshop. 
A - 73