erties 
h the 
irical 
st be 
ectra 
Oups. 
and 
e to 
, the 
10te 
any. 
ical 
sing 
ters 
Bouzidi, Sonia 
  
LAND USE CLASSIFICATION AT MESO-SCALE USING REMOTELY SENSED DATA 
S. Bouzidi! , F. Lahoche!, I. Herlin!, V. Hochschild? and H. Staudenrausch? 
! INRIA - Domaine de Voluceau - B.P. 105 - Rocquencourt - 78153 Le Chesnay Cedex, France 
Sonia.Bouzidi @inria.fr, Fabien.Lahoche @inria. fr, Isabelle. Herlin @inria.fr 
2 Institute for Geography, University of Jena, D-07743 Jena, Germany 
c5voho 0 geogr.uni-jena.de, c5sthe @ geogr.uni-jena.de 
ABSTRACT 
In this paper we present a framework to generate a land cover classification from coarse spatial resolution remotely sensed 
data acquired by NOAA-AVHRR sensor. We define a model for the pixels’ content and a process allowing to compute the 
individual proportions of the different land cover types for each pixel. The method is based on a linear mixture model of 
reflectances and exploits the good temporal frequency of NOAA acquisitions. The result provides a description in terms 
of land covers percentage within each NOAA pixel. A quality evaluation is performed on a test area for which high spatial 
resolution and temporal NOAA data are simultaneously available. 
1 INTRODUCTION 
Image classification is a widely used method for extracting information on surface land cover from remotely sensed 
images. The resulting cartography is helping decision makers in different research fields such as hydrology, agriculture 
and deforestation. 
At local scale a classification of images at high spatial resolution (as Landsat TM or SPOT data) improved with ground 
truthing gives a good information about land cover. However, high spatial resolution data or ground truthing are not always 
available at least for important surfaces. On these large areas, the NOAA/AVHRR sensors provide daily acquisitions with 
an important coverage; but due to the size of a pixel (1.1 x 1.1km), which is generally larger than a single field, each 
pixel includes several covers. 
In order to extract information from such daily data, we propose a method to access the sub-pixel content. We define a 
process having as input a low spatial resolution temporal sequence of images (a NOAA sequence). It exploits the high 
temporal frequency of these acquisitions in order to estimate within each pixel the surface percentage of each land cover 
type. The approach is based on a mixture modeling of the pixel’s response. This process will be called in this paper 
“NOAA unmixing”. The result of this analysis is not a map of labels corresponding to different land covers (thematic 
classification) but a set of images (with the same size as the original NOAA image), giving the percentage of each land 
cover type for each pixel. 
The mixture model is described in section 2. The inversion of this model to estimate land cover proportions is explained 
in section 3. Lastly results are presented and evaluated in section 4. 
2 THE MIXTURE MODEL 
The basic physical assumption underlying the linear mixture model is that the different occupations present in a pixel 
contribute independently to its reflectance. The reflectance of that pixel can be considered as a sum of the individual 
land covers reflectances weighted by their area proportions. This model was used in different studies to un-mix pix- 
els (Quarmby et al., 1992, Settle and Drake, 1993). The quoted authors consider the pixel's response in different spectral 
bands of satellite images and resolve a system of nb linear equations where nb corresponds to the number of bands. This 
approach restricts the number of studied land cover types. In fact, only the case for which we have fewer interesting com- 
ponents than the number of acquisition bands can be evaluated. In other words, if we are dealing with NOAA/AVHRR 
data, only three land cover types can be studied because of the only three available channels measuring the reflectance in 
the visible and near infrared bands. To avoid this restriction, we propose to exploit the temporal information offered by 
the NOAA/AVHRR sensor. We make use of the whole sequence with a daily frequency and covering the vegetative cycle. 
The number of linear equations is then important and we can study all land covers of a scene. At each date à we consider 
the linear relation between the pixel's reflectance and the individual reflectances of its components in the visible (channel 
1) and near infrared (channel 2) channels: 
N. 
Ré) =) n,Rilt) hz.) (1) 
j=l 
where: 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 205