Improved Target Identification of Correlated Input Data Using Recurrent Neural Networks and Feature Selection

摘要

For non-cooperative targets, combat ID may be accomplished by fusing data obtained from multiple sensors taken across time periods using ATR algorithms. With some ambiguity existing amongst fusion models, definitions are first developed to identify the specific type of fusion to be performed. Since input features extracted from sensor data for ATR algorithms are likely to contain significant correlation, models such as artificial neural networks that do not assume independent input data are a viable approach for fusion. An experiment was designed to assign generated temporal data with significant autocorrelation, cross correlation and noise into one of two classes. This feasibility study assesses use of an Elman recurrent neural network to perform fusion of multiple sensors with multiple looks to accomplish target identification. To improve classification accuracy, feature saliency screening was performed to select a subset of eight candidate input features with a signal-to-noise ratio and a network output sensitivity based measure. Both measures indicate a subset of about three of the original eight features should be retained. When comparing the two methods, both selection and ranking of salient features is consistent. Numerical results show the parsimonious subset of features improved generalization by significantly reducing the classification accuracy variance across multiple data sets and through time periods. Additionally, the reduced feature set yields an increase in the observed classification accuracy for the last time period of the external validation set.