315
NEW QUALITY REPRESENTATION FOR HYPERSPECTRAL IMAGES
Emmanuel Christophe*, Dominique Léger*, and Corinne Mailhes*
* CNES DCT/SI/AP Bpi 1219, 18 av E. Belin 31401 Toulouse Cedex 09
emmanuel.christophe@cnes.fr
* Optics Department of ONERA, 2 avenue E. Belin BP 4025, 31055 Toulouse Cedex 4
* University of Toulouse, TESA - IRIT - ENSEEIHT, 14 port St Etienne 31000 Toulouse, FRANCE
Commission VII/3
KEY WORDS: Hyperspectral, quality criteria, evaluation, compression
ABSTRACT:
Assessing the quality of a hyperspectral image is a difficult task. However, this assessment is required at different levels of the
instrument design: evaluation of the signal to noise ratio necessary for a particular application, determining the acceptable level of
losses from compression algorithms for example. It has been shown previously that a combination of five quality criteria can provide
a good evaluation of the impact of some degradation on applications, such as classification algorithms for example. This paper refines
this concept, providing a representation of the degradation which allows predicting the impact on applications.
1 INTRODUCTION
Quality criteria should be easily applicable to measure the loss of
information caused by compression or by any other forms of pro
cessing. In the case of ordinary 2D images, a quality criterion has
often to reflect the visual perception of a human observer. This
is not the case for hyperspectral images, which are first aimed
to be used through classification or detection algorithms. There
fore, quality criteria have to be relevant to these corresponding
applications. For example, some papers ( [Ryan and Arnold,
1997], [Hu et al., 2004], [Qian, 2004]) address the problem of
evaluating compression impact on specific hyperspectral applica
tions. However, quality evaluations within the context of specific
applications are heavy to conduct as they require in-depth knowl
edge of these applications.
Figure 1: Different hyperspectral images used during the experi
ments. (a) and (b) are different parts from the f970620t01p02_r03
run from AVIRIS sensor on Moffett Field site, (a) presents uni
form spatial area with strong spectral features, (b) is mixed area
with city (strong spatial frequency features), (c) is from the
f010903t01p01_r03 AVIRIS run over Harvard Forest, it contains
mostly vegetation whose spectrum contrasts with man-made ob
jects.
In a previous work [Christophe et al., 2005], different degrada
tions were applied to hyperspectral images: additive white noise,
smoothing (spectral and/or spatial) with a lowpass filtering of
the data, Gibbs effect (ringing around sharp changes) and JPEG
2000 compression [Taubman and Marcellin, 2002] using a mul
ticomponent transform. Different images from the NASA/JPL
AVIRIS hyperspectral airborne sensor were used for the exper
iments (Fig. 1). Finally, five quality criteria have been selected
to give a valuable representation of the degradations affecting the
hyperspectral data and their impacts on three different classifica
tion algorithms. These five quality criteria were found to be a
good combination to discriminate between data degradation and
appear to be almost orthogonal to each other. One advantage of
this combination is the mix between local and global criteria for
both spatial and spectral dimensions, thus enabling the detection
of local and global degradations.
2 QUALITY REPRESENTATION
2.1 Efficient representation for five criteria
The five quality criteria retained in [Christophe et al., 2005] were
the MAD, MAE, RRMSE, F\ and Q( x , y ). Denoting I(x,y, A)
the original image, I(x, y, A) the degraded image and ej = I — I
(x, y being spatial dimensions and A the spectral one), these five
quality criteria are defined as
• Maximum Absolute Difference
MAD = £oo(7 — I) = max {|ej-(a;, y, A) |} . (1)
However, the way to use these measures was not detailed. The
present paper proposes in the following section a graphic repre
sentation of the chosen quality criteria. A numerical method is
then derived in section 3 from this representation to provide a
way to identify the degradation nature if unknown and to predict
its impact on a specific application. The interest of the proposed
representation is finally illustrated and compared with traditional
SNR-based measure.
• Mean Absolute Error
MAE =
Tlx Tly Tï\
l
TixTlyTl X
Y \e T (x,y,X)\.
x,y,\
(2)
where n x , n y and n\ are the number of pixels for each di
mension.