Taking the identification on urban vegetation categories as an example

this study discussed feature set optimization methods to improve the efficiency of objected-oriented classification. Considering the characteristics of urban vegetations from IKONOS

31 features were primarily selected

including 6 shape indices

2 location features

17 spectral and 6 texture features. Firstly

the features with low entropy and strong correlation with others were removed from the primary feature set

and the dimension of feature set was cut down to 23. From the point of identification on urban vegetations

the minimum and mean J-M distance were used to select the optimum feature set from 2 to 23 dimensions using 220 samples of vegetation patches

and the dimension of feature set was decreased to 14. K-L transformation was used to further decrease the dimension of feature set

in which deviation matrix between the target categories substituted the covariance matrix between different features

and the results showed that K-L transformation to the whole feature set compressed 70% of features and K-L transformation to the subgroup feature set compressed 50% of features

respectively. Comparing with the classification rules derived through CART

K-L transformation to subgroup feature set achieved the training accuracy 12% higher than the transformation to the whole feature set

and 1% lower than that without K-L transformation

respectively. The classification accuracy also showed that the total accuracy and Kappa coefficient using K-L transformation with subgroups decreased only 1.5% and 2.3%

but its feature set dimension decreased 50%.