IEEE Geoscience and Remote Sensing Magazine - December 2021 - 248

" Unmixing Methods That Use Spectral Libraries " section to
be widely applicable. Fortunately, several techniques have
been proposed to address both of these problems, which we
discuss in detail in this section.
HOW TO CONSTRUCT SPECTRAL LIBRARIES
Many library-based SU works assume that spectral libraries
are manually obtained from in situ and controlled laboratory
measurements [37], [260], which may be complicated in
practical applications. Moreover, existing libraries may have
been acquired in conditions that do not reflect those actually
observed in the scene, which introduces errors in the SU
process [37], [103], [261]. Even the spatial resolution at which
the hyperspectral image was acquired was found to have a
considerable impact on the results of SU with MESMA in urban
environments when the library was fixed a priori [262].
Traditional EM extraction algorithms (EEAs), on the other
hand, typically consider only a single spectral signature
per material and are thus unable to appropriately address
spectral variability [33], [263]. These shortcomings make
the construction of spectral libraries one of the main challenges
of library-based SU methods [260]. A simple and reliable
method that has been employed to construct spectral
libraries depends on expert knowledge to manually select
pure pixels of each material from the hyperspectral image
[262], [264]. However, there has been growing interest in
developing methods that can reduce the amount of user supervision
and automatically extract libraries directly from
observed hyperspectral images. Three main general lines of
research can be identified in this direction:
1) Extract multiple pure pixels from the observed hyperspectral
image to generate a candidate library, and then
cluster the extracted signatures into their respective material
classes.
2) Generate libraries using radiative transfer models that
represent EM variability mathematically.
3) Extract pure pixels while keeping information about
their spatial locations, and apply an interpolation algorithm
to generate EM signatures for each image pixel.
Figure 15 provides an overview of the key ideas underlying
each of these approaches, which are reviewed in the following.
IMAGE-BASED LIBRARY CONSTRUCTION
IMAGE-BASED LIBRARY EXTRACTION
Image-based library extraction is characterized by the following:
◗ It enables spectral libraries to be extracted with signatures that have
the same conditions as the image pixels.
◗ It can benefit from expert knowledge to reliably identify pure pixels in
the image.
◗ It strongly depends on the presence of pure pixels.
◗ The observed image should not be too small.
◗ Mixed pixels may be included in the library by mistake.
◗ Clustering the extracted signatures into their correct material classes
is challenging.
248
The simplest approaches for the construction of imagebased
spectral libraries are completely supervised. Image
pixels are included in the library based on their correlation
to some initial EMs manually selected as the extreme
points of the PCA of the observed image [97], [265] or
simply by manually screening a large number of pure pixels
extracted from the image by using expert knowledge
about the spectral characteristic of the materials in the
scene [264]. Pure pixels were also extracted from multiple
hyperspectral images of the same scene acquired at different
spatial resolutions to increase the diversity of the
resulting spectral
library in urban environments [262].
Other work used only partially labeled data to reduce the
amount of domain knowledge that is required [266]. Recent
strategies attempted to automate this process by extending
EEAs for the extraction of multiple signatures of
each material in the observed image. The first work in this
direction proposed to apply traditional EEAs to random
subsets of pixels that are sampled from the hyperspectral
image (with or without replacement) [267]. Different
sets of EM signatures are generated through this method.
All the extracted signatures are then grouped into different
sets corresponding to the material classes by using a
clustering algorithm (e.g., k-means). The size of the image
subsets, however, must be selected with great care for the
EEAs to work satisfactorily [187], and the clustering step
can be challenging.
Later works proposed different strategies for the extraction
and selection of multiple pure pixels and EM candidates
from the observed image. One simple iterative strategy
consists of including in the library all pixels that are within
a given spectral distance of reference EMs [268]. This process
is performed iteratively, with the reference EMs initialized
using a standard EEA and then updated as the mean values
of the library signatures at the previous iteration. Besides
being very simple, this procedure does not require the library
signatures to be clustered afterward. A related strategy
worked in a reverse way by iteratively removing pure pixels
from a large initial set of candidate signatures to obtain the
final spectral library [269]. A pixel candidate is removed if
it can be represented with little error as a convex combination
of the remaining signatures in the library. A clustering
procedure is then performed to group the selected spectra
into EM classes.
Recent works have proposed more involved empirical
approaches to differentiate between spectrally similar materials
when extracting and clustering the EM signatures
and to remove mixed pixels from the constructed library.
For instance, in [270], EM extraction was performed multiple
times for different subsets of the spectral bands constructed
at multiple spectral scales and intervals. These
signatures were then clustered into EM classes based on a
metric constructed from features derived by individually
applying clustering algorithms to the spectral scales and
intervals used previously. A related strategy considered the
extraction and clustering of the library signatures based on
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021
