Bayesian de-mixing
Bayesian de-mixing is a very powerful technique that
shines where PCA and ICA fall short. First and foremost,
Bayesian de-mixing returns a quantitative result with the
units of de-mixed spectra being the units of the input data.
The de-mixed vectors are also always positive and sum
to one, which makes the transition from statistics to science
quite natural. There are many optional parameters
that can be tweaked within the Bayesian code, but typically
at least the number of independent components is
required. The disadvantage of the Bayesian method is
speed, and additional insight is necessary to optimize the
algorithm. Typically, in our analysis flow, we start with
PCA and ICA to identify the parameter space; once the
region of interesting solutions or phenomena is identified,
we perform Bayesian de-mixing.
While a plethora of Bayesian-based statistics methods
exist, we have found the algorithm provided by Dobigeon
et al. to be the fastest and easiest to use [43]. The Bayesian
approach assumes data in a Y = MA + N form, where
observations Y are a linear combination of positionindependent
endmembers, M, each weighted with respective
relative abundances, A, and corrupted by an additive
Gaussian noise N. This approach features the following:
the endmembers and the abundance coefficients are nonnegative,
fully additive, and sum-to-one [44-47].
The algorithm operates by estimating the initial projection
of endmembers in a reduced subspace via the NFINDR
[48] algorithm that finds a simplex of the maximum
volume that can be inscribed within the hyperspectral data
set using a non-linear inversion. The endmember abundance
priors along with noise variance priors are picked
from a multivariate Gaussian distribution found within the
data, whereas the posterior distribution is based on endmember
independence calculated by Markov Chain Monte
Carlo, with asymptotically distributed samples probed by
the Gibbs sampling strategy. An additional, unique aspect
of Bayesian analysis is that the endmember spectra and
abundances are estimated jointly, in a single step, unlike
multiple least square regression methods where initial spectra
should be known [43].
Clustering
A very natural way to analyze data is to cluster it. There
are many algorithms available that have a variety of
built-in assumptions about the data and as such could
predict the optimal clustering value, order clusters based
on variance, or other distance metrics, etc. We present a
method, k-means clustering, which is rather flexible and
easy to find on a variety of platforms and in many programming
languages. The only required input value for
k-means is the number of clusters; however, additional
variables such as the distance metric, number of iterations,
how the initial sample is calculated, and how to