Information Technology Reference
In-Depth Information
The known methods employ various mathematical tools including the vector
quantization [9], the principal component analysis [10], the orthogonal trans-
forms [11] and combined approaches where similar transforms (like 3-D wavelets
or 3-D JPEG) are used for the spectral and the spatial decorrelation of the data
[12]. Whilst the independent and the principal component analysis (ICA and
PCA) methods have been basically recommended for the spectral decorrelation
of bands, other orthogonal (different wavelet and discrete cosine) transforms
have been mainly exploited for decreasing the spatial redundancy in HSI. This
is explained by the fact that the ICA and the PCA techniques are more common
in classification based on spectral features, whilst DCT and wavelets are put into
basis of the modern standards JPEG and JPEG2000 used for 2-D data (image)
lossy compression [13], [14].
An important item in the lossy compression of HSI is to take into account
the fact that the original images are noisy and the signal-to-noise ratio (SNR)
is considerably different in different sub-band images [15]. Then, if losses mainly
relate to the noise removal (image filtering), such lossy compression can be useful
in two senses. First, the data size reduction is provided. Second, images are
filtered [16] and this leads to a better classification of the decompressed HSI.
Note that similar approaches have been considered in astronomy [17] and it has
been demonstrated that the lossy compression under certain conditions does not
lead to the degradation of object parameters measurements for the compressed
images.
In general, there are two options in compressing AVIRIS and other hyperspec-
tral data. One option is to compress the radiance data and the other variant is
to apply coding to the reflectance data. Below we considered the latter approach
since it has been shown in [18] that it leads to smaller degradations.
Two basic requirements are to be satisfied in the lossy compression. The
statistical and spatial correlation characteristics of the noise are to be carefully
taken into consideration to introduce minimal losses in the image content [19],
[20]. Besides, it is desirable to carry out the compression in an automatic manner,
i.e., in a non-interactive mode. Note that a provided CR depends on both the
noise level (the type and the statistical characteristics) and the image content
[19], [21]. The requirement to increase the CR as possible remains important as
well.
Fortunately, there exist methods for the blind evaluation of the noise statistics
[22], [23]. Methods that operate in the spectral domain are able to evaluate
the variance of additive i.i.d. noise quite accurately even if the image is rather
textural [22], [25], [27]. However, these methods produce biased estimates in cases
of the spatially correlated noise. The estimates might be considerably smaller
than true values of the noise variance. Note that the noise in AVIRIS images is
not i.i.d. [23].
The aforementioned properties of the estimates of the noise statistics are taken
into consideration in design of the modified method for the automatic lossy
compression of the AVIRIS images. In fact, below we show how it is possible to
improve the performance of the method earlier proposed in [16]. The positive
Search WWH ::
Custom Search