• hyperspectral images
• binary decision theory
• anomalous change detection
• mis-registration noise estimation

In latest years, the possibility to exploit the high amount of spectral information
has made hyperspectral remote sensing a very promising approach to detect changes
occurred in multi-temporal images. Detection of changes in images of the same area
collected at different times is of crucial interest in military and civilian applications,
spanning from wide area surveillance and damage assessment to geology and land
cover. In military operations, the interest is in rapid location and tracking of objects of
interest, people, vehicles or equipment that pose a potential threat. In civilian contexts,
changes of interest may include different types of natural or manmade threats, such as
the path of an impending storm or the source of a hazardous material spill.
In this PhD thesis, the focus is on Anomalous Change Detection (ACD) in airborne
hyperspectral images. The goal is the detection of small changes occurred in two images
of the same scene, i.e. changes having size comparable with the sensor ground
resolution. The objects of interest typically occupy few pixels of the image and change detection must be accomplished in a pixel-wise
fashion. Moreover, since the images are in general not radiometrically comparable,
because illumination, atmospheric and environmental conditions change from one
acquisition to the other, pervasive and uninteresting changes must be accounted for in
developing ACD strategies.
ACD process can be distinguished into two main phases: a pre-processing step, which
includes radiometric correction, image co-registration and noise filtering, and a
detection step, where the pre-processed images are compared according to a defined
criterion in order to derive a statistical ACD map highlighting the anomalous changes
occurred in the scene. In the literature, ACD has been widely investigated providing
valuable methods in order to cope with these problems. In this work, a general overview
of ACD methods is given reviewing the most known pre-processing and detection
methods proposed in the literature. The analysis has been conducted unifying different
techniques in a common framework based on binary decision theory, where one has to
test the two competing hypotheses H0 (change absent) and H1 (change present) on the
basis of an observation vector derived from the radiance measured on each pixel of the
two images.
Particular emphasis has been posed on statistical approaches, where ACD is derived in
the framework of Neymann Pearson theory and the decision rule is carried out on the
basis of the statistical properties assumed for the two hypotheses distribution, the
observation vector space and the secondary data exploited for the estimation of the
unknown parameters. Typically, ACD techniques assume that the observation
represents the realization of jointly Gaussian spatially stationary random process.
Though such assumption is adopted because of its mathematical tractability, it may be
quite simplistic to model the multimodality usually met in real data. A more appropriate
model is that adopted to derive the well known RX anomaly detector which assumes the
local Gaussianity of the hyperspectral data. In this framework, a new statistical ACD
method has been proposed considering the local Gaussianity of the hyperspectral data.
The assumption of local stationarity for the observations in the two hypotheses is taken
into account by considering two different models, leading to two different detectors.
In addition, when data are collected by airborne platforms, perfect co-registration
between images is very difficult to achieve. As a consequence, a residual misregistration
(RMR) error should be taken into account in developing ACD techniques.
Different techniques have been proposed to cope with the performance degradation
problem due to the RMR, embedding the a priori knowledge on the statistical properties
of the RMR in the change detection scheme. In this context, a new method has been
proposed for the estimation of the first and second order statistics of the RMR. The
technique is based on a sequential strategy that exploits the Scale Invariant Feature
Transform (SIFT) algorithm cascaded with the Minimum Covariance Determinant
algorithm. The proposed method adapts the SIFT procedure to hyperspectral images and
improves the robustness of the outliers filtering by means of a highly robust estimator of
multivariate location.
Then, the attention has been focused on noise filtering techniques aimed at enforcing
the consistency of the ACD process. To this purpose, a new method has been proposed
to mitigate the negative effects due to random noise. In particular, this is achieved by
means of a band selection technique aimed at discarding spectral channels whose useful
signal content is low compared with the noise contribution. Band selection is performed
on a per-pixel basis by exploiting the estimates of the noise variance accounting also for
the presence of the signal dependent noise component.
Finally, the effectiveness of the proposed techniques has been extensively evaluated by
employing different real hyperspectral datasets containing anomalous changes collected
in different acquisition conditions and on different scenarios, highlighting advantages
and drawbacks of each method.
In summary, the main issues related to ACD in multi-temporal hyperspectral images
have been examined in this PhD thesis. With reference to the pre-processing step, two
original contributions have been offered: i) an unsupervised technique for the estimation
of the RMR noise affecting hyperspectral images, and ii) an adaptive approach for ACD
which mitigates the negative effects due to random noise. As to the detection step, a
survey of the existing techniques has been carried out, highlighting the major drawbacks
and disadvantages, and a novel contribution has been offered by presenting a new
statistical ACD method which considers the local Gaussianity of the hyperspectral data.