IEEE Geoscience and Remote Sensing Magazine - March 2022 - 85

sensitivity to the dual-window size because OSP extracts
sample information by using feature changes to resist abnormal
measurement pollution, and the CNND adopts a
smoothing strategy to weaken the influence of anomalies.
In addition, the best results from the three data sets are
obtained by the CRD algorithm with the optimal dual
window, which indicates that the CRD has potential to
achieve optimal performance.
UNIVERSALITY OF ALGORITHMS
FOR DIFFERENT DATA SETS
For the RX, RPCA, LRASR, and AED global anomaly detection
methods, it is clear in Table 4 that RPCA can achieve
good results for all data sets and is more suitable for unknown
scenes. The RX and AED have better detection on data
anomalies with simple background distributions, while
the LRASR is more suitable for small targets because its
background dictionary design considers only those. For the
LRX, SVDD, CRD, OSPAD, and CNND local anomaly detection
methods, Figure 15 shows that the OSPAD and CNND
have a relatively stable effect for different data sets with different
parameters, but they have better location for large targets
in simple backgrounds. The LRX, SVDD, and CRD can
obtain more stable and good results with small targets under
different parameters. This is because small targets have
fewer requirements for the size of the inner window of the
double window, so it is easy to avoid abnormal pollution
and achieve high precision. In addition, from the perspective
of algorithm running time, global anomaly detection
methods, such as RPCA, should be preferred for large scenes
and targets. For small scenes and targets, local anomaly detection
methods, such as the CRD, can detect anomalies
more effectively.
SUMMARY AND FUTURE CHALLENGES
SUMMARY
Hyperspectral images contain abundant spectral information
about ground objects, which makes them unique in
target detection applications. Anomaly detection is more
practical when detailed target features are unknown. In the
past few decades, many algorithms have been developed
for hyperspectral anomaly detection. In this article, hyperspectral
anomaly detection algorithms were divided into
seven categories, based on statistics, distance, reconstruction,
subspace, spatial-spectral characteristics, deep learning,
and real-time processing. Anomaly detection algorithms
based on statistics and reconstruction have attracted extensive
attention, and many excellent versions have been
developed. Among them, the idea of statistical anomaly
detection is simple to implement, and anomaly detection
based on reconstruction is reasonable, with a strong generalization
ability. In recent years, spatial-spectral- and deep
learning-based methods have become research hotspots.
They improve algorithm performance by mining image
features. In addition, the hyperspectral anomaly detection
MARCH 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
evaluation indices and commonly used real image data sets
were widely investigated in this article.
In the experiments, nine representative algorithms were
selected for anomaly detection on three typical data sets
(the small target image HYDICE Urban, large target image
San Diego airport, and complex scene National Mall sets),
and the outcomes were analyzed and discussed in depth.
The results show that the optimal algorithms for different
images vary. In general, after a large amount of parameter
debugging, the CRD has the best effect, followed
by RPCA, the SVDD, and the LRX. However, considering
the running time, parameter analysis, and universality
of the algorithms, it can be concluded that in practical
applications, for images with simple background distributions,
the RX and AED can quickly obtain relatively
reliable results. For small target images, the CRD has the
best anomaly detection, which is not only insensitive to
window settings but faster. For large target images, global
anomaly detection methods
are more reliable. RPCA is
a good choice for situations
where the context and objectives
are unknown.
FUTURE CHALLENGES
Anomaly detection is an important
hyperspectral remote
sensing application. However,
the algorithms that have been
widely recognized and applied
are still dominated by the classical RX algorithm.
Therefore, in addition to the development of new algorithms,
it is necessary to solve key technical problems in real
applications. Future challenges are summarized in three aspects:
algorithm universality, parameter tuning, and timeliness.
It should be noted that data universality refers only to
the applicability of different scenarios and does not include
data corruption caused by other factors.
Algorithm universality is a common problem in hyperIN
ADDITION TO THE
DEVELOPMENT OF NEW
ALGORITHMS, IT IS
NECESSARY TO SOLVE KEY
TECHNICAL PROBLEMS IN
REAL APPLICATIONS.
spectral remote sensing image processing. Hyperspectral
image scenes can be simply divided into four cases: simple
backgrounds and pixel-level targets, simple backgrounds
and planar targets, complex backgrounds and pixel-level
targets, and complex backgrounds and planar targets. Relatively
speaking, the more complex the background and the
smaller the target, the less likely the target is to be detected.
Therefore, in practical applications, unknown scene situations
will lead to great differences in detection. To solve this
problem, multiple features may be used to improve the universality
of an algorithm to the data, and the most effective
features can be adaptively selected.
Parameter tuning is an indispensable step in algorithm
research. The quality of parameter debugging dictates
the detection accuracy of algorithms. In general, an
algorithm has different parameter settings for various
data. In practical applications, users are most likely to be
85
IEEE Geoscience and Remote Sensing Magazine - March 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2022
