A GA-based Band Selection Algorithm for Hyper-spectral Image Classification

概要

論文の詳細を見る
To improve the accuracy of land-cover classification using the hyper-spectral data (i.e., hyperion data), a GA (Genetic Algorithms)-based Band Selection algorithm (termed "GBS algorithm") for hyper-spectral classification has been proposed. The required condition for improving the classification accuracy is not only to maximize the classification accuracy for the training-and the reference data sets used to evaluate the overall accuracy, but also tc minimize the error rates of omission- and commission-error. For satisfying those conditions simultaneously, the effective bands for classification are searched through GA operations. Toward the end of the run, "51 bands" out of all hyperion bands were selected as the effective bands for classification. As the results of maximum likelihood classification based on the selected bands, the average of classification accuracy with respect to training classes has increased from 80.5% to 98.1 %, while the average of error rate has decreased from 40.7% to 3.4%. Furthermore, PCC (Probability of Correct Classification) with respect to the reference data has increased from 72.2% to 81.4%. Those results indicate that GBS algorithm might be useful for selecting the effective-band for hyper spectral classification.
社団法人日本リモートセンシング学会の論文