Assessing the robustness of neural network classifiers

A preliminary report on the first phase of a multiyear experiment designed to assess the robustness of neural network classifiers compared to human experts involving the classification of terrain features is presented and discussed. The experiment includes a definition of the problem, description of the terrain data sets, preprocessing of the imagery, a variable reduction scheme involving genetic algorithms, manual and automatic classification routines, and an assessment of the different methodologies. Preliminary results of a parametric and automatic classification of a Landsat image are also presented