Aerospace and Electronic Systems Magazine July 2017 Tutorial XI - 15

Acito et al.
present in the reference image (Fig. 3 (b)).
Obviously, such differences cannot be effectively managed by RE algorithms such
as the CC and the CE, which assume space
invariant acquisition conditions within the
entire scene.
According to the processing chain depicted in Fig. 1, in both the experiments,
the data were preprocessed in order to reduce the noise in the images. Specifically,
residual striping noise reduction was obtained by means of the Multi-Linear Regression (MLR) based algorithm proposed
in [4]. Random noise reduction is obtained
by binning four contiguous spectral bands
in a way that they form a unique channel.
The resulting filtered data have spectral
dimension equal to 127 (L = 127). It is
important to remark that the data obtained
after striping reduction and binning along
with the CRMs adopted in the two experiments, are available and freely downloadable from http://rsipg.dii.unipi.it/.
Each experiment consists in applying the five ACD algorithms presented in
this paper to the selected multitemporal
hyperspectral image pair. Results are obtained by adopting the CE for the RE and
the LCRA approach. To discuss the impact
of the RE and the LCRA on the various
ACD algorithms' performance, additional
results are derived by disabling the RE or
the LCRA blocks.
To apply the LCRA approach, the
RMRE magnitude estimation is obtained
Figure 3.
by the method proposed in [8]. Such a
EX-2. (a) Test image D2F22H2, the red rectangles point out two shadowed areas that are not presmethod provides the estimates of the mean
ent in the reference image and that make a graphical evidence of the differences in the acquisition
values μx, μy and the standard deviations
conditions. (b) Reference image D1F12H1. (c) CRM (red dots) and NARM (blue dots), the targets
σx, σy of the RMRE in the two spatial dilabeled as S1 and S2 are two distinct examples of temporal-spectral anomalies. (d) Average of the
rections.
The specific values of those estitest image and the reference image.
mates for the two image pairs considered
in EX-1 and EX-2 are reported in Table III.
The uncertainty window is obtained according to the relationship
in (26) with α = 2.
Table III.
We conclude this subsection dedicated to the experiments setup
by
specifying the size of the two sliding windows Ωμ(i, j) and
RMRE estimates for the two image pairs.
ΩΓ(i, j) adopted by the SVACD algorithm. Both the windows have
a squared shape with size (in pixel units) 3 × 3 and 27 × 27, respecRMRE
tively. The corresponding number of secondary data is Nμ = 9 and
estimates
EX-1
EX-2
(pixels)
NΓ = 729 (≅ 5.6 · L).
μx

0.87

0.62

μy

-1.94

2.20

σx

1.08

1.77

σy

0.48

0.84

JULY 2017, Part II of II

PERFORMANCE COMPARISON PROCEDURE
To compare the performance of the ACD algorithms we use the object based experimental receiver operating characteristics simply
denoted as ROC, hereinafter.
IEEE A&E SYSTEMS MAGAZINE

http://rsipg.dii.unipi.it/

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine July 2017 Tutorial XI