222
9: Chintan A. Shah
Step 6: Update the bias b j by using the pixel vectors from r = 1 to t, where tis
the index of the current pixel vector being processed
t
L xrP (wjlxn B )
bj = _r=_~ _ _ _ _ _
(9.11)
L P (wjlxn B)
r=1
End For
Step 7: If the log-likelihood of the observed hyperspectral data, given as
(9.12)
remains unchanged, go to Step 9, else continue.
Step 8: Repeat Step 2 to Step 7.
Step 9: Use Bayes decision rule to determine class allocation for pixel vector
as,
(9.13)
The ICAMM classification algorithm as described above is an iterative algorithm, where the adaptation loop (Step 2 to Step 7) is repeated until the
convergence is achieved. The convergence is measured in terms of the loglikelihood (i. e. log [p (XIB)]) of the observed data. The adaptation is ceased
once the log-likelihood of the observed data stabilizes with increasing number
of iterations. This adaptation procedure, however, becomes computationally
intensive with an increase in data dimensionality (Shah 2003; Shah et al. 2004).
In the next section, we will outline the proposed experimental methodology in classifying the hyperspectral datasets. We will also discuss the feature
extraction techniques employed for preprocessing the hyperspectral data in
order to satisfy the assumption N = M.
9.3
Experimental Methodology
The steps involved in our experimental methodology for classification of hyperspectral data using the ICAMM algorithm are depicted in Fig. 9.1. First, we
remove the water absorption bands and some noisy bands as observed from
visual inspection of the dataset. Next, we perform preprocessing of data, also
known as feature extraction to reduce the dimensionality further in order to
satisfy N = M assumption for implementing the ICAMM algorithm.
9: Chintan A. Shah
Step 6: Update the bias b j by using the pixel vectors from r = 1 to t, where tis
the index of the current pixel vector being processed
t
L xrP (wjlxn B )
bj = _r=_~ _ _ _ _ _
(9.11)
L P (wjlxn B)
r=1
End For
Step 7: If the log-likelihood of the observed hyperspectral data, given as
(9.12)
remains unchanged, go to Step 9, else continue.
Step 8: Repeat Step 2 to Step 7.
Step 9: Use Bayes decision rule to determine class allocation for pixel vector
as,
(9.13)
The ICAMM classification algorithm as described above is an iterative algorithm, where the adaptation loop (Step 2 to Step 7) is repeated until the
convergence is achieved. The convergence is measured in terms of the loglikelihood (i. e. log [p (XIB)]) of the observed data. The adaptation is ceased
once the log-likelihood of the observed data stabilizes with increasing number
of iterations. This adaptation procedure, however, becomes computationally
intensive with an increase in data dimensionality (Shah 2003; Shah et al. 2004).
In the next section, we will outline the proposed experimental methodology in classifying the hyperspectral datasets. We will also discuss the feature
extraction techniques employed for preprocessing the hyperspectral data in
order to satisfy the assumption N = M.
9.3
Experimental Methodology
The steps involved in our experimental methodology for classification of hyperspectral data using the ICAMM algorithm are depicted in Fig. 9.1. First, we
remove the water absorption bands and some noisy bands as observed from
visual inspection of the dataset. Next, we perform preprocessing of data, also
known as feature extraction to reduce the dimensionality further in order to
satisfy N = M assumption for implementing the ICAMM algorithm.