components should be physically viable, well-behaved,
positive, have additive weights, etc. This level of analysis
can be achieved by Bayesian linear de-mixing methods,
specifically an algorithm conglomerate introduced by
Dobigeon et al. [43].
The main advantage of these methods is a quantitative,
interpretable result where the final endmembers are
non-negative, in the units of input data, and with all of
the respective abundances adding up to 1. Therefore, at
each location, the data is decomposed into a linear combination
of spectra where each pixel in the probed grid
consists of a number of components (i.e., conducting behaviors)
present in a corresponding proportion. Note
that these constraints allow a direct transition from statistical
analysis to physical behavior. By making the abundances
additive and the endmembers positive, we can
begin assigning physical behavior to the shape and nature
of the endmember curves. By extension, analysis of the
endmember loading maps adds the spatial component
to the behavior that non-statistical methods of analysis
lack entirely.
Following the experiments at 0%, 58%, and 87% humidity,
we performed Bayesian de-mixing of the current signal
into four components. The reasons for choosing four
components and the supporting arguments are discussed
in detail by Strelcov et al. [58]. De-mixed vectors for all
three experiments as well as loadings for the 0% and 87%
humidity cases are shown in Figure 9. The de-mixed components
correspond to 1) electronic transport through a
potential barrier (Figure 9a) active in the central and outer