Input feature selection for automatic target recognition of temporal data

Prior to engaging hostile targets, USAF doctrine requires a high level of confidence to ‘label’ each target correctly. To increase ‘label’ accuracy, combat identification may fuse data from multiple sensors through time. Automatic target recognition algorithms may then be required to fuse sensor data that are highly correlated. The authors suggest the use of a ‘one big net’ neural network model to fuse all sensor information. To improve classification accuracy state-of-the art feature selection methods are compared for a temporal neural network. A reduced set of input features is then observed to reduce classification accuracy variance while retaining the mean classification performance.