Key words: Neural Networks, Classification, Genetic Algorithms, Data Mining
This paper describes some preliminary results concerning the robustness and generalisation capabilities of parametric methods versus machine-learning methods with respect to band selection and the subsequent classification of hyperspectral images. Specifically, we compare a genetic algorithm-based approach to a correlation coefficient-based approach in the selection of data used for classifiying a HYDICE image. Additionally, we use non-parametric statistics and an image-rendering tool for quantitative and qualitative analysis of neural network topology and image value classification. We then discuss the generalisation of these techniques and tools for other neural network applications. Results from a set of experiments concerning the parameter settings for both the genetic algorithm and backpropagation neural network are presented and discussed.