Due to advances in sensor technology, it is now possible to acquire hypersp
ectral data simultaneously in hundreds of bands. Algorithms that both reduc
e the dimensionality of the data sets and handle highly correlated bands ar
e required to exploit the information in these data sets effectively. We pr
opose a set of best-bases feature extraction algorithms that are simple, fa
st, and highly effective for classification of hyperspectral data. These te
chniques intelligently combine subsets of adjacent bands into a smaller num
ber of features. Both top-down and bottom-up algorithms are proposed. The t
op-down algorithm recursively partitions the bands into two (not necessaril
y equal) sets of hands and then replaces each final set of bands by its mea
n value. The bottom-up algorithm builds an agglomerative tree by merging hi
ghly correlated adjacent bands and projecting them onto their Fisher direct
ion, yielding high discrimination among classes. Both these algorithms are
used in a pairwise classifier framework where the original C-class problem
is divided into a set of ((C)(2)) two-class problems.
The new algorithms 1) find variable length bases localized in wavelength, 2
) favor grouping highly correlated adjacent bands that, when merged either
by taking their mean or Fisher linear projection, yield maximum discriminat
ion, and 3) seek orthogonal bases for each of the ((C)(2)) two-class proble
ms into which a C-class 2 problem can be decomposed. Experiments on an AVIR
IS data set for a 12-class problem show significant improvements in classif
ication accuracies while using a much smaller number of features. Moreover,
the proposed methodology facilitates the extraction of valuable domain kno
wledge regarding the importance of certain bands for discriminating specifi
c groups of classes.