WC 
  
A SEMI AUTOMATIC ROAD EXTRACTION METHOD 
FOR ALOS SATELLITE IMAGERY 
H. Hasegawa 
Geographical Survey Institute, 1 Kitasato, Tsukuba, Ibaraki, 3050811, Japan — hase(2gsi.go.jp 
KEY WORDS: Semi-automation, High resolution Satellite, Edge Recognition, Multi sensor Interpretation, GIS, Cartography 
ABSTRACT: 
A semi automatic road extraction method has been constructed and evaluated. Our purpose is to obtain 1/25,000 level road data from 
PRISM image. PRISM is panchromatic, three views, and the 2.5m resolution sensor carried on ALOS satellite that will be launched 
on late 2004. A centre line-detecting algorithm is employed for feature extraction. It picks up pixels where second derivative of 
brightness becomes maximal. Then those pixels are linked if both probability of line and angle difference between adjacent pixels 
satisfy given conditions. Acquired line candidates are classified by its photometric property. We test both automatic grey scale 
threshold method and traditional unsupervised multi band classification method for this stage. After eliminating false line segments, 
a line linking method is applied. If all of angle difference, lateral offset, and net gap are less than threshold, the pair is considered as 
one long line. This process is applied iteratively while a connectable pair remains. The proposed method was tested by a simulated 
ALOS image data set created from three-line airborne sensor images and an IKONOS data set. 
The result shows that 80% of road was extracted before false line elimination while 80% of extracted line was false data. There still 
remains 70% of false road segments even after the classification method. Both correctness and completeness are unexpectedly poor. 
As our method is working on a single image and does not use full feature of PRISM yet, a method using three-dimensional property 
is needed. 
1. INTRODUCTION 
This research focuses on extracting 1/25,000 level road network 
from ALOS optical sensor data. Geographical Survey Institute 
of Japan (GSI) is the national mapping agency in Japan and GSI 
has responsibility to keep 1/25,000 level map data the latest. 
The data acquisition method, however, has mainly depended on 
manual aerial photo interpretation. As a result, map revision 
interval is too long (3-5yr in urban and suburban area, 5-10yr in 
rural area) to follow rapid changes especially in urban area. 
Advanced Land Observation Satellite (ALOS) will be lunched 
at the end of 2004. Tts primary task is making map in Japan. 
ALOS is expected to another source for map making. However, 
as data amount is enormous and data generation from optical 
images is time consuming and expensive task, a semi automatic 
road network extraction system is needed. 
A great amount of research has been targeted on developing 
methods to extract features like roads and buildings. As a full 
automatic feature extraction method is obviously out of reach, 
many investigations have been focusing on semi automatic 
feature extraction method. 
Baltsavias (2002) briefly reviewed the recent trend of image 
analysis, feature extraction strategy, and existing commercial 
systems regarding automated object extraction. The paper 
pointed out that proper definition of the target objects, choice of 
adequate input data, use of existing knowledge, and different 
data analysis method from low level to high level was the main 
considerations. It argued that object modelling had become 
popular as level of modelling was highly depends on desired 
resolution and that combined use of sensor data was more 
common. It also said use of contextual information such as 
relationship between neighbours and usage of as much cues as 
possible (i.e. use multi spectral data for land cover 
classification and use DTM for shadow analysis) had received 
much attention. 
Baumgartner et al (1999) proposed an automated road 
extracting system that used hierarchical road model applied to 
* 
aerial photo (0.2 — 0.5m resolution). Based on their previous 
work, they clearly distinguished global contexts (rural, forest 
and urban) and corresponding local contexts were selected and 
used for road extraction process. Their approach worked quite 
well in rural area with less than 0.5m resolution images whereas 
more improved extraction and grouping methods were needed 
in urban and forest area. 
Price (1999) proposed a road grid extraction method standing 
on a specific assumption. They assumed that streets crossed 
regularly at right angles. The result was quite fine and 
processing speed was very rapid although their assumption may 
not be applicable to complex urban scene. 
Zhao et al (2003) proposed another semi-automatic road 
extraction method using multi spectral high-resolution satellite 
images. Firstly “road mask” was created by multi spectral data 
classification. Chains of edge pixels were tracked based on 
local edge direction and straight lines were obtained. Then this 
method use template matching to determine the best direction 
of line and obtain next road node. A result in urban area was 
good for major road whereas small road were missed, as road 
boundaries were unclear due to the objects surrounding roads. 
In rural area, both major and minor roads were properly 
extracted by indicating adequate control points. 
Treash and Amaratunga (2000) proposed an automatic road 
detection system using high-resolution greyscale images. This 
system consists of three parts, edge detection, edge thinning, 
and edge pairing. The result is good in rather simple image 
whereas in highly occluded road and dense residence area it 
shows poor performance. The authors suggest some additional 
information based on texture of grey value is needed. 
Park and Kim (2001) proposed a template based road extraction 
system. À user needs to input an initial seed point to extract à 
road. Then orientation of the road seed point is calculated 
automatically. They conclude that this template matching 
method have more flexibility than the methods using snake. 
They pointed out the method may not work on the road cast by 
shadow. 
Interne 
Hu an 
road € 
two di 
appliec 
proces 
connec 
extract 
extract 
matchi 
Main 1 
result i 
compli 
Accor 
can b 
topolo 
grounc 
road a 
this st 
metho 
are er 
photor 
group! 
In this 
metho 
segme 
À cent 
emplo: 
seconc 
are lin 
line ar 
condit 
candid 
both : 
unsupx 
After 
applie 
differe 
pair is 
iterat 
In the 
Then 
discus 
2.1 DB 
A sim 
study. 
Data 
simule 
resolu 
(2.5m 
such a 
were | 
study. 
As the 
are sel 
| (a) i 
some 
roads | 
its adj 
(b) is . 
A maj 
be see 
roads