Neural network clutter filter for large-array mosaic sensors

Summary form only given, as follows. Tracking moving targets with satellite-mounted large-array mosaic sensors requires either large on-board digital computers or wide-band data links (for ground processing). Use of high detection thresholds, to avoid saturating the computer or communication link, results in low false-alarm rates but high missed-detection rates. The authors show that a neural network (NN) can be used for analog filtering based on spatial-temporal correlation among track signals, with improved performance over simple thresholding. Feedback from interconnected nodes effectively lowers detection thresholds when signals are present and raises thresholds when signals are absent. A significant finding in this research was that the behavior of this single-layer neural network can be interpreted in terms of least-squares estimation, and its interconnection weights are analytically related to conditional probabilities, which can be determined from Monte-Carlo ´training´ simulations.<>