Hyperspectral dimensionality reduction via localized discriminant bases

To overcome the dimensionality curse of hyperspectral data, the authors of the paper have investigated the use of grouping the spectral bands along with localized discriminant bases, followed by decision fusion to develop an ATR system for data reduction and enhanced classification of hyperspectral data. The proposed system is robust to the availability of limited training data. Initially, the entire span of spectral bands in the hyperspectral data is subdivided into subspaces or groups based on a performance metric. The groups are not allowed to grow beyond what is supported by the amount of available training data. Feature extraction is done using supervised methods as well as unsupervised methods. Further, decision level fusion is applied to the features extracted from each group. To reduce the effect of conflicting decisions by individual groups, a voting scheme called Qualified Majority Voting is adopted to combine decisions. The effectiveness of the proposed system is tested using a data set consisting of hyperspectral signatures of a target class, Cogongrass (Imperata Cylindrica), and a non-target class, Johnsongrass (Sorghum halepense). Cogongrass is an invasive species of plant whose monitoring has become important due to the extensive ecosystem damage that it causes. A comparison of target detection accuracies by the proposed system before and after decision fusion is done to illustrate the effect of the influence of each group of spectral bands on the final decision and to illustrate the benefit of using decision fusion with multiclassifiers.