Background clutter rejection using generalized regression neural networks

Advances in focal plane array technology has led to the development of “staring” sensors for a number of remote sensing application. Here the sensor line-of-sight (LOS) is fixed to a background point and stares at that point while radiometric measurements are collected. Inadvertent motions of the LOS result in unwanted time signals (clutter) that corrupt the measurements. This paper develops a technique that estimates these signals in the output of each focal plane detector by employing a Generalized Regression Neural Network (GRNN). The GRNN is an optimal estimator that is based on the well-known statistical concept of conditional probability. Two implementations are evaluated for removing the background. The first technique estimates the clutter signal in each detectors output based on the previous measurements from that detector. The second method trains the GRNN with the measurements from the surrounding spatial pixels on the current data frame. Both techniques were evaluated using measurement sets from an existing staring space sensor. The results show the GRNN estimates and follows the clutter signal very well with a rms error <3% which is within the variation of the random sensor noise