Aerospace and Electronic Systems Magazine July 2017 Tutorial XI - 19

Acito et al.

Figure 8.

EX-2. Test statistic obtained by applying the: (a) SACD, A1 and A2 are two nontemporal
anomalies that exhibit very high values; (b) HACD, in this case A1 and A2 have very low
values; (c) SDADC; (d) SDHACD; (e) SVACD. In the last three cases just one of the two
targets (A2) has very high values. Results in this figure were obtained by using the CE for
RE and the LCRA approach.

JULY 2017, Part II of II

IEEE A&E SYSTEMS MAGAZINE

gorithms is well below that obtained by including
the LCRA in the processing chain.
Fig. 6 also confirms that the local processing adopted by SVACD makes the algorithms
more robust to the false changes generated by the
RMRE. In particular, the AC-ROC shows that the
false changes due to nontemporal anomalies do
not increase significantly.
We conclude the discussion of EX-1 with an
example that illustrate the concept of nontemporal anomaly. For this purpose, in Fig. 7, we show
a zoomed version of the area around the nontemporal anomaly labeled as A3 in Fig. 4 (a). Fig.
7 (a) shows the RGB for the test and the reference image, while Fig. 7 (b) shows the RX test
statistics. The presence of the anomalous objects
is clearly visible in both the cases. The spectral
radiances of such anomalous target in the two images are plotted in Fig.7 (c). The two plots are
very close thus indicating that the anomalies have
the same spectrum. This suggests that A3 is an
anomalous object placed in the same position of
the scene in the two looks: this is what we call a
nontemporal anomaly.
Fig. 8 shows the test statistics obtained in
the second experiment (EX-2). Such results are
carried out by considering the complete processing chain that includes the CE and the LCRA
approach. Here again we have applied a simple
contrast enhancement technique to improve the
visual appearance and interpretability of the images.
Fig. 8 (a)-(e) show that the statistics provided
by all the algorithms have high values in correspondence of most of the anomalous changes
included in the CRM of Fig. 3 (c). However, as
expected, the SACD exhibits high values also in
those pixels that represent nontemporal anomalies. As an example, consider the two objects labeled as A1 and A2 that are anomalous in both the
test and the reference image but are not temporally anomalous. SACD provides very high values
for both the targets. Conversely, A1 and A2 have
very low values (lower than zero) in the HACD
statistic. This confirms the weakness of SACD in
discriminating the temporal from the nontemporal anomalies. Note, also in this case, that HACD
gives high positive values (the brightest levels) to
temporal anomalies and negative values to most
of the nontemporal anomalies (included A1 and
A2) in the NARM. Here again we can conclude
that HACD is able to separate most of the nontemporal anomalies from the background pixels
as well as from the temporal anomalies. SDACD,
SDHACD and SVACD provide high values for
some of the nontemporal anomalous pixels in the
19

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