where a and b are the unknown proportions and
xj, yj, zj and
wj are the reflectances of the pure and the mixed class
in band j. For given values of xj, yj,
zj and wj, equation (14) represents a straight
line in a coordinate system (a,b) where we measure the values
of proportions a and b along the two axes. If we consider the
equations for two bands, we can solve for the two unknowns a and
b. This way, a quadruple of pixels (i1, i2, i3, i4)
with values (xi1j, yi2j, zi3j,
wi4j) in band j and values (xi1k,
yi2k, zi3k, wi4k) in band k defines
a point in parameter space (a,b). This process is called the Hough
transform.
Imagine now that we use a grid in the parameter space
(a,b), i.e. imagine that we discretise the values of parameters
a and b and create discrete cells each of which corresponds
to a particular set of ranges of values of a and b. Every time
we find a pair (a,b) that is within the range of a cell, the value
of the cell will be incremented by 1.
Imagine then that we pick at random quadruples of pixels, one from each pure
class and one from the mixed class. For every quadruple of pixels and for
every pair of bands we identify a point (a,b). We accumulate the value
of the corresponding cell by 1. We repeat the process for all quadruples
of pixels and all pairs of bands. The cell in the (a,b) space that
will end up having the maximum value will indicate the correct values of
the mixing proportions a and b.
This approach was proposed by Bosdogianni, Petrou and Kittler (Mixed Pixel
Classification with Robust Statistics, IEEE Transactions on Geoscience and
Remote Sensing, vol 35,pp 551-559, 1997 ). Figure 11 shows an example
of an accumulator array with the peak clearly identified. |