Abstract
A new subband-based classification scheme is developed for classifying underwater mines and mine-like targets from the acoustic backscattered signals. The system consists of a feature extractor using wavelet packets, a feature selection scheme, and a backpropagation neural network classifier. The data set used consists of the backscattered signals for seven frequency bands and six different objects: two mine-like targets and four non-targets. The targets are insonified at 72 aspect angles from 0 to 355 degrees with 5 degree increment. Simulation results on ten different realizations of this data set and for signal-to-noise ratio of 12 dB are presented. The receiver operating characteristic curve of the classifier generated based on these results demonstrates excellent classification performance of the system. In addition, the generalization ability of the trained network is demonstrated by computing the error and classification rate statistics on a large data set consists of 50 different realizations.
Introduction
Identifying underwater mines using sonar is a classic "needle in a haystack" problem. Acoustic backscattered signals, or echoes, are highly volatile; a mine’s acoustic signature changes drastically depending on the aspect angle—the angle at which the sonar hits the target. Furthermore, the underwater environment is chaotic, filled with reverberation from the sea floor and surface, as well as biological noise.
Traditional methods often relied on simple spectral features or matched filtering, which struggled with non-stationary noise and the geometric complexity of targets. This research presents a robust classification system designed to handle the non-repeatability of these signatures while maintaining a low false-alarm rate—a critical requirement when the "false alarm" might be a deadly mine.
Methodology
The proposed system utilizes a multi-stage pipeline combining multi-resolution signal analysis with neural networks to extract discriminatory features from sonar echoes.
1. Preprocessing
To ensure the system wasn't "cheating" by using timing artifacts or signal strength, the data was normalized. Inverse filtering was applied to remove secondary reflections (multipath) caused by the experimental tank walls. Synthesized reverberation was added to simulate a 12 dB Signal-to-Reverberation Ratio (SRR).
2. Feature Extraction
The core innovation lies in the use of Wavelet Packet Decomposition and Linear Predictive Coding (LPC).
- Wavelet Packets: Unlike a standard Fourier Transform, which loses time information, Wavelet Packets allow for a flexible, subband-based decomposition. This captures subtle tonal features across different frequency bands.
- LPC Modeling: For each subband, a 4th-order LPC model was fitted. LPC serves as a compact way to describe the spectral envelope or "shape" of the echo in each subband.
3. Feature Selection and Classification
The researchers used the Fisher discriminant function to prune the initial 30 features down to the 22 most discriminatory ones. These were then fed into a two-layer backpropagation neural network (BPNN) with a 22-42-2 architecture, trained to distinguish between "mine-like" and "non-target" objects.
Experimental Setup and Datasets
The study utilized a dataset provided by the Coastal Systems Station (CSS) in Panama City, FL. The environment was a controlled water tank approximately 6 meters deep. Six objects were tested: two mine-like targets (a metallic blunt-bullet shape and a plastic truncated cone) and four non-targets (a water-filled drum, a limestone rock, a granite rock, and a wooden log). Each object was rotated and insonified at 72 different aspect angles in 5° increments.
Key Results
The system's performance was evaluated using Receiver Operating Characteristic (ROC) curves. For single-aspect classification (using just one sonar ping), the system achieved a 92.64% correct classification rate with a 7.43% false alarm rate.
Multiaspect Fusion
Recognizing that a single "look" might be ambiguous, the researchers implemented a multiaspect fusion scheme. By combining the outputs of the network from three sequential pings (e.g., at 0°, 10°, and 20°), the performance improved significantly. A nonlinear fusion approach boosted accuracy to 99.03% with a remarkably low 0.97% false alarm rate.
Conclusion
This subband-based approach demonstrates that combining Wavelet Packets with neural networks provides a powerful tool for underwater target recognition. The use of LPC coefficients effectively compresses the acoustic data while retaining essential physical characteristics. Most importantly, the results highlight the power of "multiple looks"; while a rock might mimic a mine from one angle, it is highly unlikely to do so across multiple sequential observations.
Summary updated