24-01-2013, 11:20 AM
IMAGE FUSION AND SPECTRAL UNMIXING OF HYPERSPECTRAL IMAGES FOR SPATIAL IMPROVEMENT OF CLASSIFICATION MAPS
1IMAGE FUSION.pdf (Size: 1.2 MB / Downloads: 56)
ABSTRACT
In this paper we propose a new approach for the improvement
of the spatial resolution of hyperspectral image classification
maps combining both spectral unmixing and pansharpening
approaches. The main idea is to use a spectral unmixing algorithm
based on neural networks to retrieve the abundances of
the endmembers present in the scene, and then use the spatial
information retrieved from the pansharpened image to find
the location of each endmember within the enhanced pixel
according to the endmembers abundances. The proposed approach
has been applied both to real and synthetic datasets.
INTRODUCTION
Hyperspectral imaging technology provides images composed
of several narrow and contiguous bands, resulting in
a very high spectral resolution, from the visible to the infrared
region. This main feature permits to better characterize
materials on the observed scene. However, satellite based
hyperspectral sensors have generally a low spatial resolution
as a result of the fundamental tradeoff between spatial resolution,
spectral resolution, and radiometric sensitivity in the
design of electro-optical sensor systems. This limitation can
affect the performances of the algorithms used to process
hyperspectral data. In the particular case of image classification,
low spatial resolution can produce mixed pixels, which
spectra is a combination of two or more pure spectra, representing
unmixed classes, called endmembers [1].
Spectral unmixing
Conventional image classification techniques may provide
a poor representation of the distribution of the land cover,
especially in the case of images containing mixture of pure
classes. In some applications it is therefore desireable to
unmix pixels into their single component parts. A range of
spectral mixture models have been developed for this task.
Among these, the linear mixture models are the most widely
used. These models, however, may not always be appropriate
when the task is to obtain a land cover map. A possible
solution to this problem is to soften the classification output.
Differently from the standard classification, a softened
classification output represents the grade of membership of a
pixel to each class. In [7] Foody proposed the use of Neural
Networks (NN) to perform the soften classification, while
in [8] a NN algorithm has been used to estimate the abundances
of endmembers in mixed pixels from hyperspectral
images. In both papers a standard backpropagation NN has
been trained using samples of pure pixels, to perform a classification
task.
Chris-Proba dataset
A similar approach has been applied to a dataset obtained
by fusing a CHRIS-Proba hyperspectral image with a Quick-
Bird Panchromatic one. In this case the spectral distortion
in the resulting enhanced image is influenced also by the differences
in terms of angles of view, dates of acquisition and
registration error. From a quantitative point of view, it can
be noted by inspecting Fig. 2 that even if a spatially enhanced
result can be obtained with the classification of the
enhanced image, the spectral distortion introduced by the fusion
process leads to incorrect classified pixels resulting in an
overall accuracy of about 90%. On the other hand, the proposed
method presents an improvement of the overall accuracy
(97%). Moreover, the distortions present in the enhanced
image lead to the wrong classification of several pixels, while
the proposed method, based on the original spectral information,
leads to the detection of the correct classes.
CONCLUSIONS
In this paper we proposed a novel technique for the retrieval of
the spatial distribution of different land cover types obtained
by a spectral unmixing approach. In fact, even if spectral unmixing
is a useful instrument to estimate the abundance of
each endmember within each pixel, it does not provide any
information on the spatial distribution of the detected types.
The proposed method tries to overcome this limitation using
an enhanced image obtained by fusing the original image with
a panchromatic one, according to the Indusion approach. The
intra-pixel distribution is obtained by the comparison of the
spectrum of each endmember with the spectra of the corresponding
pixels in the enhanced image.