Suppression of impulsive noise in active acoustic time series

Echolocation in shallow water is often characterized by spurious detections generated by impulsive noise processes and heavy tailed boundary reverberation. Persistent echos with durations commensurate with the transmitted pulse length that are delay/Doppler consistent over several interrogations are best detected and left to subsequent classifier stages. Random transient events in the time series increase the false alarm rate but can be suppressed by preprocessing the data stream with a non-linear filter. A simple preprocessor consisting of a linear predictor with a thresholded nonlinearity suppresses sporadic impulse-like events without affecting longer duration echos. The preprocessor is tested with two heavy tailed time series. The first is a reverberation process generated by a Rayleigh mixture model that simulates recorded high frequency in-water data. A linear FM echo signal is added and it is shown that the preprocessor restores nearly all of the matched filter ROC performance of the uncontaminated time series. The second noise process is a variable kurtosis mixture that simulates transient ambient noise events and is generated by combining Gaussian noise with a random impulsive time series