Peter Doucette 
  
ROAD CENTERLINE VECTORIZATION BY SELF-ORGANIZED MAPPING 
Peter DOUCETTE , Peggy AGOURIS', Mohamad MUSAVI””, Anthony STEFANIDIS” 
University of Maine, USA 
"Department of Spatial Information Science & Engineering 
(doucette, peggy, tony) @spatial.maine.edu 
"Department of Electrical and Computer Engineering 
musavi @eece.maine.edu 
  
Working Group III/3 
KEY WORDS: Semi-Automated Extraction, Self-Organization, Cluster Analysis, Minimum Spanning Tree. 
ABSTRACT 
A novel approach to semi-automated road centerline extraction from remotely sensed imagery is introduced. Providing 
inspiration is Kohonen's self-organizing map (SOM) algorithm. With IFOV « 2m, road features are open to region- 
based analysis. A variation of the basic SOM algorithm is implemented in a region-based approach to road vectorization 
from high spatial (1.0m) and spectral resolution imagery. Using spectrally classified road pixels as input, centerline 
nodes are located via cluster analysis of the local density fluctuations in the input space. Linking the self-organized 
locations of the nodes with a minimum spanning tree algorithm provides global topological structure, which is 
subsequently refined. The idea is use contextual analysis from which to derive optimum topology. The result is a 
vectorized road centerline network suitable for direct GIS database population. Preliminary results demonstrate the 
algorithm's potential for robust vectorization when presented with noisy input. 
1 INTRODUCTION 
Road network extraction ranks among the most fundamental of image analysis operations in support of many GIS 
applications. A complete road extraction task entails a raster-to-vector conversion, which is basically a lossy signal 
compression operation whose objective is to derive compact geometrical representations for raster features. A primary 
goal within the Image Understanding (IU) community is the minimization of human effort required to vectorize roads 
(among other features) from images when populating GIS databases. Despite substantial IU advancements with 
automated feature extraction techniques over the years, the population of geospatial feature databases remains in large 
part a costly manual process. With a growing variety of next-generation imaging platforms offering enhanced spatial 
and spectral resolutions, more robust extraction tools will play a larger role in improving future productivity. With 
IFOV < 2m, most road features are not characterized as single raster edges. Rather, they manifest as elongated regions, 
a.k.a. ‘ribbon’ features, and as such become subject to region-based analysis in addition to (dual) edge analysis. 
However, an inherent dilemma with high spatial resolution is the potential for a decrease in local signal-to-noise ratio, 
and therein lies the trade-off for higher geometric accuracy. The goal of this paper is to investigate the use of self- 
organization methods for robust raster-to-vector conversion of road centerlines from high spatial and spectral resolution 
imagery. 
Road extraction algorithms are often categorized according to their degree of automation. Most prevalent are the semi- 
automated methods, which attempt to strike a synergistic compromise between human and machine. Semi-automated 
methods are further divided into two broad categories. The first includes line-following methods, in which local 
exploratory image filters sequentially trace a minimum cost path from one point to the next (Mckeown and Denlinger, 
1988; Vosselman and Knecht, 1995). The second group includes active contour models, i.e., snakes (Kass et al., 1988), 
which are adapted for 2D and 3D road extraction (Trinder and Li, 1995; Neuenschwander et al., 1995; Gruen and Li 
1997). Snakes are fundamentally edge-based curvilinear feature extractors that are distinct from line-following in that 
they exemplify ‘simultaneous’ curve-fitting as opposed to sequential. Notwithstanding, either method is considered 
semi-automatic owing to their requirement of user provided seed point inputs. On the other hand, fully automated 
methods attempt to completely circumvent user intervention during the extraction process. A considerably more 
rigorous approach, it usually requires a skillful integration of contextual information and a priori knowledge into the 
road extraction task (Mckeown et al., 1985; Baumgartner et al., 1999). Still, a basic concern of fully automated methods 
is limited generalization ability, particularly as a priori knowledge becomes more specialized with scene complexity. 
  
246 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.