Speaker
Description
The low attenuation in seawater and the low-velocity layer SOFAR (Sound Fixing and Ranging) channel enable the hydroacoustic stations of the International Monitoring System (IMS) to record acoustic waves over long distances. The French National Data Center uses the Progressive Multi-Channel Correlation (PMCC) method to detect low-frequency coherent waves (< 40 Hz). These detected underwater acoustic waves are emitted by various sources (e.g. earthquake, volcanism, cryosphere, whales, airgun, anthropophonic explosion). High-amplitude PMCC detections can be associated with one type of these sources with DTK-Diva, but their revision is not systematically carried out by an analyst. We present here a method to discriminate these detections by source type. We analyze recordings from hydrophone triplets in the Atlantic (HA10), Indian (HA1, HA4, HA8) and Pacific (HA3, HA11) oceans over the period from January to December 2023. This represents 111,233 detections with a maximum amplitude above 1 Pa. We train a convolutional neural network and calibrate it using conformal prediction. On average, 75 ± 6% of detections are associated with one source type with a confidence level of 95%. A consolidated event catalogue would enhance the discrimination performance of the classifier by source type.
[email protected] |