Chen Yuen Teoh 
  
JUNCTION EXTRACTION FROM HIGH RESOLUTION IMAGES BY COMPOSITE LEARNING 
Chen Yuen TEOH, Arcot SOWMYA 
Department of Artificial Intelligence 
School of Computer Science and Engineering 
University of New South Wales 
Sydney, Australia 
{cteoh|sowmya} @cse.unsw.edu.au 
KEY WORDS: Road Extraction, Remote Sensing, Pattern Recognition. 
ABSTRACT 
This paper aims to present the junction recognition level of RECoiL, a system for Road Extraction from High Resolution 
Images by Composite Learning. By modelling junctions as specific junction types and learning different rules for each 
junction type, junction extraction is achieved. A clustering technique also has been applied to the same task using the 
same features identified during supervised learning, and the results are compared. 
1 INTRODUCTION 
Roads are a major man-made surface feature and communication infrastructure among people. Maps are an abstraction 
of the real world, and creation and revision of maps are important processes for governments and businesses. Even 
though modern cartographers utilize Geographical Information Systems (GIS) databases for map storing and manipulation 
purposes, map creation is still dominated by manual techniques. Therefore semi or full automatic road recognition and 
extraction from remote sensing images will certainly aid faster and more accurate map maintenance. Machine learning is 
an Artificial Intelligence (AI) approach that could achieve automatic road recognition. 
In previous research (17), we have described Road Extraction from High Resolution Images by Composite Learning, 
RECoiL, which is a semi-automatic, multi-level, adaptive and trainable road recognition system based on supervised 
learning and K-means clustering (6) technique. RECoiL produces promising output by reducing the data set, while 
also generating comprehensible rules for recognition and avoiding the arduous labour of example selection tasks at lower 
levels. In addition using clustering techniques, we also manage to achieve similar results to supervised learning techniques 
currently maintains for the same task. 
In this paper we discuss the junction recognition level of RECoiL. Our approach to junction recognition is to divide 
junctions into different types and tackling them one by one. By applying a supervised learning technique to different 
junction types, we manage to extract junction based edges from an edge image. The current results of extraction are still 
preliminary and many more examples of a particular junction type are required to train the system. 
The most common approach for extracting roads is by detecting or tracking elongated lines in low resolution images, 
while profile matching or detection of parallel lines is used for high resolution images (2). Others combine these methods 
with prior knowledge to achieve semi or automatic road extraction (11, 19). The knowledge based approach is another 
common technique for road extraction, which can be divided into two categories. In the first category, an expert system 
and expert knowledge are used to formulate rules to guide the extraction system (7, 9). However these systems have 
limited application as they require extensive prior knowledge and are source dependent. The second category is to make 
use of external knowledge such as GIS data and maps (2, 4). However, availability of such prior data is not always 
guaranteed. 
Roads may also be extracted from multi-source images such as multi-sensor, multi-temporal, multi-scale and multi- 
resolution images (3, 10, 18). However this approach suffers from image registration problems and paucity of multiple 
images of the same location. A grouping approach is yet another alternative (1), where the extraction system groups edges 
into road parts and segments, finally linking with junctions to form road networks. A model based extraction system was 
introduced (16), that groups roads using different knowledge of roads. Contextual information of a scene was used (13), 
where different techniques for road extraction are applied to different contexts of images. 
Machine learning techniques for remote sensing applications remain rare. Huang et al. (8) proposed a supervised learning 
based system which generates rules for scene analysis and Dillon et al. (5) proposed an Image interpretation and Scene 
understanding system (CITE) which uses incremental, relational learning to label a scene. 
  
882 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.