Automated Stepwise Selection of Hyperspectral Hypertemporal Features for Target Detection

There is a need to identify and extract the most useful information from spectral and temporal features, where usefulness is measured in terms of signal classification or target detection. For this reason, instead of using the entire spectral and temporal feature space, pertinent features are extracted to reduce dimensionality. The classification accuracy increases if the distributions of the classes are statistically more separate in the feature space. One method to measure the ability of a feature to discriminate between two classes is to calculate the area under the feature\´s receiver operating characteristics (ROC) curves. To estimate the classification capabilities of the different spectro-temporal features, ROC areas were calculated for each feature in the spectro-temporal feature vector for the two-class system. An algorithm was then designed and implemented to obtain the best combination of the individual features to serve as a "best feature". Once the best features were determined, the system was tested to estimate the accuracy of target detection.