hul 2004 
  
vod, J.J, 
idex for 
Remote 
ood, J.J. 
data for 
Meteo.. 
., 1995. 
spective. 
rison of 
or urban 
76. 
1dscapes 
te Sens. 
>, 1988 
'eflected 
Remote 
on from 
idiation: 
acc Res. 
tion: Îts 
ordi.) 
of the St. 
using 
72. 
onment, 
ition of 
perature 
e Sens. 
E., and 
hematic 
Sens. 
qtchard, 
Vionthly 
tion of 
| NDVI 
-284. 
ls” from 
derived 
lization 
> Sens. 
HYPERSPECTRAL IMAGE ANALYSIS FOR MATERIAL MAPPING 
USING SPECTRAL MATCHING 
Saeid Homayouni, Michel Roux 
GET — Télécom Paris — UMR 5141 LTCI - Département TSI 
{saeid.homayouni, michel.roux } @enst.fr 
46 rue Barrault, 75013 Paris, France 
7O9900D0O00© oF HDT 
KEY WORDS: Hyper Spectral Imaging, Image Analysis, Material Mapping, Urban Scene Description, Data Fusion 
ABSTRACT: 
Recently, hyperspectral image analysis has obtained successful results in information extraction for earth remote sensing system. 
The data produced with this type of analysis is an important component of geographic databases. The domain of interest of such data 
covers a very large area of applications like target detection, pattern classification, material mapping and identification, etc. 
Material mapping techniques may be considered like multi-step target detection. Among the strategies for target detection, one of the 
most applied is the use of some similarity measures. In case of hyperspectral data, there are two general types of similarity measures: 
first are deterministic measures and second are stochastic measures. 
In this paper the deterministic measures for spectral matching are tested. These methods use some similarity measures like the 
euclidian distance (Ed), the spectral angle (SA), the Pearson spectral correlation (SC) and the spectral similarity value (SSV). In 
parallel, we have implemented a constrained energy minimizing (CEM) technique, for finding the most similar pixels on our 
materials of interest. These techniques are applied to two data sets which were taken with the Compact Airborne Spectrographic 
Imager (CASI), over the city of Toulouse in the South of France. 
Whereas each method has advantages and limits, a fusion technique is used to benefit from all the strong points and ignores the weak 
points of the methods. Results show that fusion may enhance the final target map; however, the primary algorithms are important 
and are useful for pure pixel targets. 
1. INTRODUCTION minimum of information about the classes of interest. This 
property is really important when we work with a large “data 
Land cover information obtained from remotely sensed data are cube” volume. 
needed for a lot of applications. Historically, there was a great There are two similarity measure categories for the analysis of 
interest in the spectral signature presented in these kinds of data hyperspectral data: stochastic measures and deterministic 
for the identification of objects and materials and the measures. In the first category, one uses the property of sample 
production of Land cover information. Recently, hyperspectral data as self information and defines some spectral information 
imagery systems have proposed a remarkable answer to this criteria such as divergence, probability, entropy, etc (Chang 
need and interest. They may provide hundreds of contiguous 2002). On the contrary, in the second category deterministic 
spectral bands which makes possible the reconstruction of the criteria are defined to measure the similarity. These criteria are 
spectral reflectance of materials. such as the spectral angle, the distance and the correlation 
Pattern recognition and remotely sensed data analysis have been between an unknown pixel and a reference spectrum. 
the subject of intensive research. The emerging techniques, and The definition of the spectral measures needs some a priori 
particularly statistical techniques, are well adapted for both knowledge on the nature of the data, objects, materials of 
multi and hyper spectral images (Kruse 1993, Richards 1999, interest and problems which must be solved. For example, in 
Landgreb 1999 and Landgrebe 2002). However, some of urban area we are facing some problems such as complexity of 
hyperspectral data properties, such as large volume of data, topography and change in materials (Alimohammadi 1998). 
need for prior knowledge about the scene and In this paper we present some similarity measures for material 
hardware/software needs are new challenges in image analysis mapping. These measures belong to the second category of 
and processing. These limitations cause some inabilities of the criteria, deterministic measures. They are spectral distance, 
classical techniques (Jia 1996). Spatial based image processing spectral angle and correlation. A more sophisticated method, 
techniques are not able to extract the existing information in the called Constrained Energy Minimizing, has also been 
hyperspectral cube (Chang 2002). Therefore another type of = investigated: it is a linear operator which maximizes the 
technique with basic in signal processing has come to response difference between target and non-target pixels. 
hyperspectral image analysis. As it will be shown, each method has some facilities and 
In hyperspectral image analysis, there is a class of techniques benefits beside of its limitations. Then a fusion strategy has 
that efficiently use the information of the spectral signature. been developed at the decision level to obtain better results. 
They are called Spectral Matching methods. These techniques 
belong to supervised pattern recognition approaches. They 
define some kind of similarity measures between an unknown 
pixel and a reference target. In these techniques, we use the 
49