66 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 22, NO. 1, JANUARY 1997 Active Impulsive Echo Discrimination in Shallow Water by Mapping Target Physics-Derived Features to Classifiers Frances B. Shin, Member, IEEE, David H. Kil, and Richard F. Wayland, Associate Member, IEEE Abstract—One of the most difficult challenges in shallow-water active sonar processing is false-alarm rate reduction via active classification. In impulsive-echo-range processing, an additional challenge is dealing with stochastic impulsive source variability. The goal of active classification is to remove as much clutter as possible while maintaining an acceptable detection performance. Clutter in this context refers to any non-target, threshold-crossing cluster event. In this paper, we present a clutter-reduction al- gorithm using an integrated pattern-recognition paradigm that spans a wide spectrum of signal and image processing—target physics, exploration of projection spaces, feature optimization, and mapping the decision architecture to the underlying good- feature distribution. This approach is analogous to a classify- before-detect strategy that utilizes multiple informations to arrive at the detection decision. After a thorough algorithm evaluation with real active sonar data, we achieved over an order of mag- nitude performance improvement in clutter reduction with our methodology over that of the baseline processing. Index Terms—Active classification, feature optimization, image compression, integrated pattern recognition paradigm, target physics. I. INTRODUCTION I N SHALLOW-WATER active sonar processing, clutter encompasses echoes from seamounts, bottom features, wrecks, schools of fish, and so on. These returns can have high signal-to-noise ratios (SNR’s), possess complicated structures, and be confused with actual target echoes. As a result, conventional active sonar systems that rely on amplitude- based thresholding will suffer from degraded detection performance and high false-alarm rate, which can overload even experienced operators. Another complicating factor in impulsive-echo-range (IER) processing is that underwater sound sources are stochastic and impulsive. They generate very short, high-power wideband shock waves and bubble pulses. Their pulse characteristics vary as a function of source type, depth, range, and frequency. Not only does their stochastic nature preclude the use of a replica correlator in detection, but the combination of high power and nondirectionality contributes to a high level of reverberation that can obscure the desired echo structure [1]. Manuscript received November 2, 1995; revised November 15, 1996. This work was supported by the Naval Air Warfare Center under Contract N62269- 94-C-1179 and by the ONR under Project RJ14C42. F. B. Shin and D. H. Kil are with the Advanced Concepts & Development, Lockheed Martin-AZ, Litchfield Park, AZ 85340-0085 USA. R. F. Wayland is with the Naval Air Warfare Center, Aircraft Division, Code 45544, Patuxent River, MD 20670 USA. Publisher Item Identifier S 0364-9059(97)01419-2. In active classification, the main objective is to fuse multiple informations to improve detection performance instead of relying on simple parameters, such as amplitude and pulse width, for detection [2]. The three critical elements in IER classification are: 1) exploration of various projection spaces subject to source variability and echo-formation phenomenol- ogy; 2) classification clue optimization; and 3) identification of an appropriate decision logic. The stochastic sound sources contribute to more echo- signature variabilities than directional coherent sources that transmit repeatable pulses with a low level of reverbera- tion. This disadvantage makes it imperative that features be derived from projection spaces that represent the desired signal 1 compactly while suppressing background noise. That is, compact signal representation is desirable to overcome the adverse impact of source-induced echo-signature variabilities on detection. In general, transformation or data projection allows one to characterize the desired signal using as few basis functions as possible while undesirable components, such as background noise and clutter, are spread out into many basis functions (energy compaction for improved SNR) and/or occupy different basis functions (subspace filtering or signal discrimination). Fig. 1 illustrates the utility of data projection. For example, the fast Fourier transform (FFT) utilizes the sinusoidal basis functions to separate multiple signals based on their spectral contents. For infinitely long narrow-band signals, the SNR im- provement or the level of energy compaction is proportional to the FFT length. However, the FFT is not a suitable projection operator for exponentially damped sinusoids since it cannot compactly represent them. In this case, the oversampled Gabor transform would be a logical candidate [3]. This observation forms the backbone of projection-space exploration. Based on our understanding of low-frequency target physics, both target echoes and clutter are assumed to consist of many components that are spread out in time. The time-domain representation, i.e., time samples, of these components is transformed onto various projection spaces from which features are extracted. Furthermore, each projection space ideally focuses on different signal attributes. This implies that fusing multiple, orthogonal or mutually reinforcing informations from various projection spaces should improve decision making. In short, it is desirable to extract classification clues or features from projection spaces 1 In this context, the desired signal includes both target and clutter echoes that exceed the amplitude-based detection threshold. 0364–9059/97$10.00  1997 IEEE