Snapping Shrimp Noise Detection Methods Based on Linear Prediction Analysis

Snapping shrimps (SSs) are a species that inhabits warm shallow waters of sea and frequently generates high amplitude short-duration signals. The sound generated by SS is a main source of underwater noise that impacts signal detection and communication. In order to reduce the effects of SS noise (SSN), it is important to detect the exact snap interval. Therefore, we propose features based on linear prediction (LP) analysis, a technique for predicting the next sample by the linearly weighted sum of previous samples, for SSN detection in this article. The characteristic of SSN, which occurs suddenly and disappears rapidly, is extremely transient. Thus, the error between the value predicted by LP analysis and the true measured value is large and features based on this error result in excellent performance for SSN detection. In addition, we further improve the performance of SSN detection by incorporating a constant false alarm rate (CFAR) detector to the proposed LP analysis-based features. For evaluation, we compare the proposed methods with multilayer wavelet packet decomposition (ML-WPD), known as the state-of-the-art in SSN detection, using text shallow-water acoustic variability experiment 2015 (SAVEX 15) data. Through evaluation, it was confirmed that the performances of the proposed methods outperform that of ML-WPD in the aspects of the receiver operating characteristic (ROC) curve and area under the curve (AUC). In particular, the proposed LP analysis-based features achieved a higher AUC by 0.082 on average and less computation complexity than ML-WPD.