187
INDEXING OF MID-RESOLUTION SATELLITE IMAGESWITH STRUCTURAL
ATTRIBUTES
A. Bhattacharya a,b , M. Roux a , H. Maître 3 ,1. Jermyn b , X. Descombes b , J. Zerubia b
“Institut TELECOM, TELECOM ParisTech, CNRS UMR 5141 LTCI,75013 Paris, France -
(avik.bhattacharya, michel.roux, henri.maitre)@telecom-paristech.fr
b Ariana (joint research group INRIA/I3S) INRIA, BP 93, 06902 Sophia Antipolis, Cedex France -
(Ian.Jermyn, Xavier.Descombes, Josiane.Zerubia)@inria.fr
Commission IV, WG IV/2
KEY WORDS: Landscape, Segmentation, Features, Extraction, Classification, Modelling, Data Mining
ABSTRACT:
Satellite image classification has been a major research field for many years with its varied applications in the field of Geography,
Geology, Archaeology, Environmental Sciences and Military purposes. Many different techniques have been proposed to classify
satellite images with color, shape and texture features. Complex indices like Vegetation index (NDVI), Brightness index (BI) or
Urban index (ISU) are used for multi-spectral or hyper-spectral satellite images. In this paper we will show the efficiency of
structural features describing man-made objects in mid-resolution satellite images to describe image content. We will then show the
state-of-the-art to classify large satellite images with structural features computed from road networks and urban regions extracted
on small image patches cut in the large image. Fisher Linear Discriminant (FLD) analysis is used for feature selection and a one-vs-
rest probabilistic Gaussian kernel Support Vector Machines (SVM) classification method is used to classify the images. The
classification probabilities associated with each subimage of the large image provide an estimate of the geographical class coverage.
1. INTRODUCTION
The growth of large image databases during the last few
decades with the advancement in image acquisition
technologies have attracted researchers from different fields to
work in the domain of image information mining systems.
These images coming from various sources must be
systematically analyzed to render important information which
are often less relevant to human perception. The technologically
advanced satellite sensors and the new storage systems have
made image data too vast and complex. The manual annotation
to describe a complex image completely is not feasible.
Indexing and retrieval from remote sensing image databases
relies on the extraction of appropriate information from the data
about the entity of interest (Daschiel and Datcu, 2005).
Indexing satellite images (Maitre, 2007) depends on the choice
of features which in turn are dependent on the type and
resolution of the sensors. For instance SIFT descriptors are
widely used in the domain of multimedia (Lowe, 2004).
Complex indices like Vegetation index (NDVI), Brightness
index (BI) or Urban index (ISU) are used for multi-spectral or
hyper-spectral images. Texture features are known to be highly
discriminative for low resolution panchromatic images
(Schroeder et al., 1998). Structural features describing man
made objects in mid-resolution images are most efficient to
describe image content (Bhattacharya et al., 2007). The road
network contained in an image is one example. The properties
of road networks vary considerably from one geographical
environment to another. The structural features computed from
them can therefore be used to classify and retrieve such
environments (Bhattacharya et al., 2007). In order to compute
the structural features of the road network, we first need to
extract the road network from the image and then convert the
output to an appropriate representation. This representation
must be absolutely independent of any extraction method. The
road extraction methods are in general resolution dependent. An
optimal road network extraction algorithm to accurately
delineate road structures for all practical purposes is very hard
to achieve. The methods used in our study are robust on many
such road characteristics but they often failed to extract the
narrow and finely structured road networks which are almost
hidden in small urban areas. This failure of the extraction
methods and hence the features computed from road networks
poorly classify images containing such areas. In order to obtain
some meaningful information from these regions, we need to
segment such areas occurring in the images. A new set of
structural features computed on segmented urban areas
combined with the existing road network features provided an
improved classification of the geographical environments.
In images, pixels provide the most basic level of information.
The pixel values are the measurements of the satellite sensors of
a region on the Earth surface. The information from these pixels
are at a level far below the semantic meaning of the desired
object or region. The classification of images based on the pixel
values is tedious and expensive and hence is not an efficient
strategy. In this paper we present a novel methodology to
classify large satellite images with patches of images extracted
from them. This is a novel idea in the sense that the patches
considered contain a significant coverage of a particular type of
geographical class. A one-vs-rest probabilistic Gaussian kernel
Support Vector Machines (SVM) classification method is used
to classify the images. In the work presented in this paper we
have defined 7 such classes. These classes can be categorized as
follows: 2 urban classes consisting of “Urban USA” and “Urban
Europe”; 3 rural classes consisting of “Villages”, “Mountains”
and Fields; an “Airports” class and a “Common” class (this can
be considered as a rejection class indicating in particular images
from seas).