Comparison of some feature subset selection methods for use in remote sensing image analysis

As feature subset selection constitutes an important aspect of data fusion in general, this paper compares different measures of goodness and their influence on the classification results. These measures are 1) Fukunaga's (1990) criterion, and 2) the ML criterion with a user-specified upper limit for the total error (Smits). Results are presented using publicly available multi-spectral/multi-sensor and hyperspectral images, and it is concluded that the ML criterion with a user-specified upper limit for the total error is a valid alternative to classical methods.