The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
Figure 5: Smoothing and simplifying hyperspectral imagery
(©Norsk Elektro Optikk). First column: A 3D view of the ini
tial hypercube (top) and a zoom on the band number #94 (mid
dle). Second and third column: the resulting hypercubes and the
corresponding bands after applying the ADF (second column)
and the proposed AML (third column). Contrary to ADF, which
smoothed without preserving image flat zones, AML simplified
and stayed constantly close to the initial hypercube intensity val
ues and structure.
adaptive time step At = Almax/raacc(|trace(TH)|) or on an
other selected one. In cases where just a single simplified hy
percube is needed, the coefficients can be customized accord
ingly. The proposed, thus, vectorial leveling AML takes its final
form when in Equation (4) the family of markers h n are derived
from the anisotropic diffused markers X n of Equation (8). Such
anisotropic markers, which do respect hyperspectral data speci
ficities, can naturally ameliorate the simplification process, with
out, in addition, demanding a search for selecting the appropriate
structure element size and type, as the classical morphological
operators do.
4 EXPERIMENTAL RESULTS - EVALUATION
The developed vectorial Anisotropic Morphological Leveling AML
was applied to a number of hyperspectral images and its evalua
tion was carried out by both a qualitative and a quantitative as
sessment. Datasets from the HySpex VNIR-1600 airborne sensor
(©Norsk Elektro Optikk A/S) with 160 channels (400-1000nm),
from the CASI-1500 airborne sensor (©ITRES) with 36 chan
nels (380-1050nm) and from EOS-1 Hyperion (©USGS) space-
borne sensor with 220 channels were available. Throughout the
evaluation procedure the compared ADF was the same with the
one that was used for the construction of the AML and each scale
n was derived after three iterations t. Both were also compared
with the classical ML after a standard channel by channel process
to the resulting ADF hypercube. For the quantitative evaluation
apart from the standard RMSE and NMSE measures -which give
a quantitative sense for the extent of variation between the inten
sity values of the compared images- the recently proposed com
plementary quality measure of SSIM [Wang et al., 2004] was,
also, employed because it is able to compare effectively local
patterns of pixel intensities under a perceived visual quality. The
lower RMSE and NMSE and the bigger SSIM values designate
the better filtering result.
Table 1 : Quantitative Evaluation
Test Data
Type of
Quantitative Measures
Filter
RMSE || NMSE |
SSIM
Figure 1
ADF
0.012
0.009
0.996
ML
0.009
0.004
0.998
Hypercube
AML
0.006
0.002
0.999
Figure 1
ADF
0.097
0.156
0.985
ML
0.035
0.020
0.996
Band #33
AML
0.034
0.018
0.998
Figure 1
ADF
0.068
0.021
0.944
ML
0.055
0.013
0.974
Band #87
AML
0.049
0.011
0.974
Figure 2
ADF
0.147
0.093
0.982
ML
0.049
0.010
0.997
Band #100
AML
0.041
0.007
0.998
Figure 5
ADF
0.009
0.004
0.998
ML
0.004
0.001
0.999
Hypercube
AML
0.003
0.001
1.000
Figure 5
ADF
0.052
0.018
0.973
ML
0.025
0.005
0.992
Band #94
AML
0.020
0.003
0.995
Noisy
ADF
0.013
0.012
0.996
ML
0.009
0.007
0.997
Hypercube
AML
0.008
0.004
0.998
In Figure 1, 3D views of the initial hypercube and the resulting
ones from the ADF and the AML are presented, together with
two corresponding bands #33 and #87 (filtering scale n=3).
The ADF smoothed strongly the data and created some intensity
shifts. In contrast the AML simplified the data but kept a closer
relation with the initial hypercube intensity values. This can be
more clearly verified by a close look at Figures 3 and 4, where
cross sections along the spatial y-axis and the spectral axis are
presented, respectively. One can observe that even thought all the
compared filters did not displace edges, the AML almost every
where stayed closer to the initial hypercube. AML simplified the
image in the spatial directions by enlarging or creating new flat
zones (levelled regions with constant intensity values), retaining
all its 2D scale space properties. In the spectral direction it ac
counted for large intensity variations (spike-like features) and at
the same time stayed close to the initial hypercube values. The
above observations can be further confirmed by the performed
quantitative evaluation (Table 1). In all cases (Figure 1), the AML
resulted to the lower RMSE and NMSE values and to the larger
structural similarity with the original image (SSIM).
In Figure 2, the initial and three of the resulting AML scale space
images are presented (scales n=2, 3 and 4). The increasingly
simplified versions of the original spatial image structure can be
observed. The quantitative comparison between AML’s result (at
scale n=4) with the corresponding ML and ADF (Table 1), indi
cate that the AML scored better in all measures. Furthermore and
evaluating the compared filtering techniques in another dataset
(shown in Figure 5), approximately the same conclusions were
derived. In Figure 5, 3D views of the initial hypercube and the
ones resulting from the ADF and the AML are shown, together
with the corresponding band #94. By comparing qualitatively,
all filtering results in the same scale (n=6), it can be observed
that the difference between diffusing (smoothing with ADF) and
simplifying (AML) adjacent intensity variations, is that a more el
egantly enhanced version of the original image is obtained from
the AML. Both methods respect image edges but the proposed
AML enforces the creation of flat regions instead of diffusing
inside them. This process obliges, also, AML to follow more
constantly the original hypercube’s intensity. The above obser
vations can be confirmed by the quantitative measures in Table
1 which indicate that the AML scored better in all measures,
229