A Neural Network Approach to Sensitivity Analysis of AVIRIS Spectral Bands

The purpose of the project described in this paper was to perform sensitivity analysis on the 224 bands collected by the Advanced Very High Resolution Imaging Spectrometer (AVIRIS) sensor. The sensitivity analysis was conducted utilizing artificial neural network technology. A baseline was established by performing partial training of a neural network using the equivalent six non-thermal TM bands as input. The remaining AVIRIS data was divided into nine groups of contiguous bands. The first, last and middle bands of each group were added to the baseline inputs and used to partially train a separate neural network using parameters identical to the baseline network. While several of the groups demonstrated a small (or even negative) impact on pixel classification, the presence of other groups improved the performance of the neural network. The results obtained support the viability of the neural network approach in ascertaining the sensitivity of band groups within the AVIRIS data. Introduction The advent of the hyper-spectral sensor has enabled the collection of data associated with spectral bands that have never before been analyzed with respect to ground cover classification. This new technology raises questions concerning the contribution of these new bands with respect to the classification of ground cover. Traditionally, the data available from the LandSat or the Spot series of satellites have bandwidths that are not sufficiently narrow to provide highly discrete information concerning the scene contents. Hyper-spectral data offers the potential to utilize narrow band passes that could potentially reveal much greater separability of landcover classes. The commercial potential for this capability is considerable. The forest industry represents a large market that would benefit greatly from processed data that could accurately separate