| Autor: |
Salin, Mikhail, Ponomarenko, Alexander |
| Předmět: |
|
| Zdroj: |
Proceedings of Meetings on Acoustics; 6/20/2021, Vol. 44 Issue 1, p1-9, 9p |
| Abstrakt: |
This paper is devoted to methods of processing hydrophone records in order to identify specific signals, produced by marine mammals. The aim of processing is to detect a signal according to a certain pattern against the background of non-stationary noise. Development of a robust detection algorithm is of great interest because it could help to assess marine mammal occurrence and distribution at greater temporal and spatial scales by excluding laborious task of manually analyzing acoustic data. In addition, such algorithm may be implemented in automatic systems designed to warn vessel crew to avoid potential collision with whales. The data, studied in this paper, contained the calls of North Atlantic right whales, that were collected in the Gulf of St. Lawrence. Machine learning is exploited to solve a signal detection problem. The model was trained by a set of manually labelled signal fragments. Each fragment is labelled whether a whale call is present or not in the fragment. The proposed detection method is based on the two-level stacked gradient boosting models (XGBoost algorithm). The first-level model deals with spectral features, extracted from the short signal frames. The second model accumulates estimated output for a group of subsequent frames. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
