Aerospace and Electronic Systems Magazine July 2017 Tutorial XI - 21

Acito et al.

Figure 11.

EX-2. (a) RGB representations of the test and the reference image in a neighborhood of the target S1. (b) RGB representations of the test and the reference image in a neighborhood of the target S2. (c) Decision statistics of the RX anomaly detector (logarithmic scale) obtained on the test and the reference image (neighborhood of the target S1). (d) Decision statistics of the RX anomaly detector (logarithmic scale) obtained on the test and the reference
image (neighborhood of the target S2).

Figure 12.

EX-2. (a) Spectral radiance of the target S1 in the two images (after RE via CE). (b) Spectral radiance of the target S2 in the two images (after RE via CE).

a linear transformation is applied to the data. Conversely, as expected, the performance of the other algorithms are negatively
influenced by the lack of RE. In this case, RE is more important
than in the EX-1, because the two images were collected in two
subsequent days and with significant difference in the acquisition
conditions.
Fig. 10 shows that the RMRE has indeed a negative impact on
the performance of all the algorithms. According to the remarks
in Subsection II-D, RMRE is responsible of the dramatic performance reduction for SDACD, SDHACD, and HACD that are subject to the increase of false alarms due to both the edge pixels and
the nontemporal anomalies. This latter effect is made evident by
the performance deterioration in terms of AC-ROC obtained by
the three algorithms. Due to the poor performance of the SACD in
discriminating between temporal andnon-temporal anomalies, its
AC-ROC does not actually change.
Fig. 10 also shows that the SVACD takes advantage from the
local estimation strategy adopted that makes the algorithm more
JULY 2017, Part II of II

robust to the false changes introduced by the RMRE. This is not
true in absolute, in that it depends on the particular relationship
among the RMRE strength, the size of the estimation windows,
and the spatial correlation of the pixels in the scene.
To conclude this section, we recall that the particular form of
the test statistic for the HACD algorithm might reduce its sensitivity to the temporal-spectral anomalies (Subsection II-D). However,
we cannot state that HACD is always unable to reveal the presence
of such kind of anomalies. Experimental evidence of this can be
found in the following. Let us consider the two targets labeled as
S1 and S2 in Fig. 3 (c). They are two distinct examples of temporal-spectral anomalies. In Fig. 11, focusing on a neighborhood of
each of the two targets, we show the RGB for the test and the reference image along with the RX test statistics. The presence of the
anomalous objects is clearly visible in both the cases. The spectral
radiances of such anomalous targets in the two images, after the
application of the CE, are plotted in Fig. 12. Specifically, Fig. 12
(a) compares the spectral radiance of the target S1 in the test image

IEEE A&E SYSTEMS MAGAZINE

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