DATA FUSION OF MULTI-SOURCE REMOTE SENSING BASED 
ON LEVEL SET METHOD AND APPLICATION TO URBAN ROAD EXTRACTION 
Cao Guangzhen 3, *, Hou Peng b , Jin Ya-Qiu c 
d Key Laboratory of Radiometric Calibration and Validation for Environmental Satellites, China Meteorological 
Administration (LRCVES/CMA), Beijing 100081, China-gzhencao@hotmail.com; 
b Resources school, Beijing Normal University, Beijing 100875, China - houpcy@163.com; 
c The Key Laboratory for Wave Scattering and Remote Sensing Information (Ministry of Education) Fudan University, 
Shanghai 200433, China - yqjin@fudan.ac.cn 
KEY WORDS: Urban; Multisensor; Fusion; Algorithms; Semi-automation; Extraction 
ABSTRACT: 
Using data fusion of multi-spectral and microwave radar images, a semiautomatic method is developed based on the level set 
method for application to urban road extraction. The fast marching method of level set makes data fusion. Radar remote sensing 
image can make up road breaks due to shadowing of high building or tree canopy in multi-spectral image, while multi-spectral 
image can be of helpful to decrease speckles and identify linear road from flat water surface. As example, the data fusion and road 
extraction from the ERS-2 SAR image and Landsat ETM+, as well as RASARSAT-1 SAR image and Landsat ETM+, in Shanghai 
area are applied. 
1. INTRODUCTION 
Semi-automatic or automatic road detection from remotely 
sensed images in dense urban area is greatly helpful of urban 
transportation mapping, planning and management and in city 
GIS database etc. 
During recent two decades, some approaches for automatic or 
semiautomatic detection of road from the optic or microwave 
radar images have been developed. These approaches usually 
take two steps: extracting possible road candidates, and linking 
road candidates. The first step is based on classification of 
remote sensing image (Dell et al., 2001; Shackelford et al.,2003) 
and conventional edge or line detection (Touzi et al., 1988; 
Tupin et al., 1998). It is most likely suitable to visible or 
infrared images with high spatial resolution and spectral 
resolution. There have been many approaches to linking the 
road candidates (Baumgartnerl et al., 1999; Bums et al., 1986; 
Jeon et al.,2002). 
However, it is well known that remote sensing of multiple 
sensors, e.g. at visible, infrared, and microwave frequencies, 
presents varying views on the geometries and spectrum 
reflectance of roads because of their different measuring 
physics. An infrared image may show rich spectral information, 
but is lack of sensitivity to mistakes of different spectrum for 
the same object and the same spectral for different objects. On 
the other hand, the multiplicative speckles in microwave radar 
image and, especially, complicated distribution of various 
objects in the urban area make the automatic road detection 
more difficult. Therefore, road information retrieval from a 
single sensor’s data is always restrictive. And data fusion of 
multiple sensors may be a new tool to retrieve more accurate 
and richer road information in remote sensing applications. 
In this paper, a semiautomatic method is developed based on 
the level set method for application to urban road extraction 
from multi-spectral and microwave radar remote sensing 
images. At first, the fast marching method of level set (LS-FM) 
is introduced briefly, which will play as a tool to fuse different 
image features for road extraction. Then different methods are 
applied to the extraction of road features from multi-spectral 
and microwave radar images respectively. A new iteration 
difference algorithm is proposed to calculate the spectral 
difference of objects in multi-spectral remote sensing image. 
After that, the speed function for LS-FM algorithm is defined 
with the features extracted from both multi-spectral image and 
microwave radar image. In this way, radar remote sensing 
image can make up road breaks due to shadowing of high 
building or tree canopy in multi-spectral image, while multi- 
spectral image can be of helpful to decrease speckles and 
identify linear road from flat water surface. 
As example, the data fusion and road extraction from the ERS-2 
SAR image and Landsat ETM+, as well as RASARSAT-1 SAR 
image and Landsat ETM+, in Shanghai area are applied. 
2. LEVEL SET AND FAST MARCHING ALGORITHM 
Level Set (LS) method was proposed by Osher and Sethian in 
1989. And the basic idea of it is to present the closed plane 
curve implicitly as the level set of a two-dimensional surface 
function. And with the surface evolution of LS function, the 
curve evolution can be traced implicitly (Sethian, 1999). 
Fast Marching (FM) is the fastest algorithm in LS method 
Sethian, 1996). Supposed that the speed of the curve is always 
positive and T i j is the time the curve passing pixel (i,j), the 
Corresponding author.