|
A divergence-based feature selection scheme was implemented in
our study. MSSimagery of SPOT satellite and their textureal
features were used for landuseclassification of Tsengwen
Reservoir watershed located in southern Taiwan.Our feature
selection schemem involves calculation of average divergence
forselected classification features. New features were
sequentially added to thegroup of already-selected features
based on the largest divergence incrementin each calculation
iteration. The ratio of divergence(DR) of selected features to
that of all features was used to determine the minimum number of
features.Then from the ranked feature sequence, we identified
those features that shouldbe selected for later landuse
calssification. Our results showed that at DR=0.9only 6 out of
12 features were needed in landuse classification. Since
featureswere sequentially selected, the class-specific increment
of classificationaccuracy contributed by the feature under
consideration could be observed. Wefound that for landuse class
of betel nut, red band was added after infrared andgreen bands
but still largely increased the classification accuracy.
Thisindicates that although for most landuse classes the red and
green bandreflectance are highly correlated, betel nut spectral
features of red and greenbands are not.
|