Adaptive feature selection for hyperspectral data analysis

Hyperspectral data can potentially provide greatly improved capability for discrimination between many land cover types, but new methods are required to process these data and extract the required information. Data sets are extremely large, and the data are not well distributed across these high dimensional spaces. The increased number and resolution of spectral bands, many of which are highly correlated, is problematic for supervised statistical classification techniques when the number of training samples is small relative to the dimension of the input vector. Selection of the most relevant subset of features is one means of mitigating these effects. A new algorithm based on the tabu search metaheuristic optimization technique was developed to perform subset feature selection and implemented within a binary hierarchical tree framework. Results obtained using the new approach were compared to those from a greedy common greedy selection technique and to a Fisher discriminant based feature extraction method, both of which were implemented in the same binary hierarchical tree classification scheme. The tabu search based method generally yielded higher classification accuracies with lower variability than these other methods in experiments using hyperspectral data acquired by the EO-1 Hyperion sensor over the Okavango Delta of Botswana.