IEEE Geoscience and Remote Sensing Magazine - June 2019 - 24

each image pixel. Like most empirical methods, ELM does
not require absolute radiometric calibration, but it does require further adjustments if the atmospheric and illumination conditions are not stationary within the scene [150].
Whereas empirical methods are widely applied in operational scenarios [151], over the years researchers have put
their efforts toward "forever improving the physical model"
[152] to provide increasingly more accurate radiative-transfer-modeling-based AC techniques. The AC approaches
belonging to this category are based on an analytic model
for the radiation transfer in the atmosphere, such as that
explored in the "Environmentally Induced Target Variability" section, and, thus, need to estimate the various model
parameters and then invert the model to retrieve the surface reflectance. Model parameter estimation is generally
accomplished by resorting to radiative-transfer codes, such
as MODTRAN [54] or 5S [153], or exploiting the imaged radiance in certain bands to estimate visibility, spatially variable parameters (e.g., WV and aerosols), and adjacency effects. One of the first AC techniques led to the atmospheric
removal program (ATREM) [145], which did not explicitly
account for adjacency effects. This was instead incorporated
in the approach of fast line-of-sight atmospheric analysis of
spectral hypercubes (FLAASH) [154], which is a very popular commercial AC software package. More recently, the atmospheric and topographic correction algorithm (ATCOR)
[147] has been developed, which also features a correction
for topographic effects (provided that a digital elevation
map of the scene is available).
SHADOW AND CLOUD REMOVAL
Shadows and clouds are among the most nonstationary
sources of variability in hyperspectral imagery, in both
space and time. Whereas the effects of clouds are mostly
significant for high-altitude acquisitions (e.g., from spaceborne sensors), the impacts of shadows are stronger
for high spatial resolution imagery, especially in the urban environment.
Approaches to shadow removal are typically based on
specific transformations that are applied to the imaged
data [155], [32]. For instance, a transformation from Cartesian to hyperspherical coordinates allows the data to be
separated into d - 1 angular components related to spectral information content and one radial component whose
magnitude represents the overall brightness level [32]. By
segmenting the image and identifying those clusters with
the lowest magnitude of the radial component, a shadow
mask can be obtained [32]. Another approach to remove
shadows is to exploit ancillary data, such as lidar 3D point
clouds [44], [135].
Detection of clouds is essentially based on finding
pixels in which the cloud optical depth is greater than a
specific threshold [136] or evaluating a combination of
threshold tests [156]. This enables generation of a cloud
mask. When possible, the radiative effects (e.g., changes
in illumination, shadows, temperature variations, spatial
24

structure, and enhanced adjacency effects) of cloud layers
are corrected [136].
SPECTRAL REFLECTANCE DECOMPOSITION WITH
MULTIPLE GAUSSIAN CURVES
A physics-inspired spectral characterization for material
reflectance spectra was presented by Lanker et al. [137],
with the aim of suppressing variability due primarily to
material morphology and composition. Inspiration was
drawn from the Lorentz oscillator model applied to the
material dielectric function (i.e., how, as a function of
wavelength, the optical index of refraction varies). Each
Lorenz oscillator, roughly speaking, corresponds to a spectral "line" (although for solid materials, unlike gases, individual lines can be very broad); thus, a model of material
spectral reflectance with multiple Lorentz oscillators identifies multiple lines in a spectrum. In [137], the authors
chose to fit the curve of reflection versus wavelength using
Gaussian line shapes-this was found to be simpler and
less computationally intense than fitting the complex dielectric function with Lorentz oscillators. This is similar
to the curve-fitting approaches performed in mineral spectroscopy [157].
It was experimentally found that, although the Gaussian amplitudes were subject to considerable variations
upon varying atmospheric, illumination, and viewing
conditions as well as material morphology and composition, the locations and widths of the Gaussian functions
were much less affected [137], [138]. These findings suggest that alternative profiles based on the locations and
widths of the Gaussian functions used to fit the spectra
could be adopted for robust characterization of target
material spectra and potentially for background spectral
variability as well.
IN-SCENE TARGET SIGNATURE CHARACTERIZATION
In some circumstances, the problem of target variability
can be bypassed by directly estimating from the image the
most discriminative target spectral signature for a given
scenario and material by means of a multiple-instance machine-learning approach [158]. This can be accomplished
when at least imprecisely labeled training samples are available-for instance, when GPS coordinates for the targets
are known and general regions of pixels containing the targets can be identified. (GPS precision is generally limited by
the coregistration accuracy; thus, pixel-level ground truthing is hardly ever achieved.) Using a simple iterative algorithm with a closed-form solution for the update rule, the
multiple-instance learning approach is capable of retrieving the target signature that maximizes the cosine spectral
similarity (more robust in the case of mixed training data)
between the estimated signature and the positive (targetincluding) instances while minimizing the similarity with
the non-target-labeled instances. As a byproduct, the estimated signature helps uncover the discriminative spectral
characteristics and features of the target class, which are
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019
IEEE Geoscience and Remote Sensing Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2019
