Automatic identification of sources recorded by the hydroacoustic stations of the International Monitoring System

As part of the Comprehensive Nuclear Test Ban Treaty Organisation (CTBTO), six hydroacoustic stations were installed. Although few in number, they record underwater acoustic waves that propagate over long distances via the SOFAR (SOund Fixing And Ranging) channel. Low-frequency coherent waves (< 40 Hz) are detected automatically by the PMCC (Progressive Multi-Channel Correlation) method. An average of 306 daily detections, with a Maximum Amplitude (MaxAmp) above 1 Pa, are reported. At this point, there is no identification made by any analyst of the source type (e.g. earthquake, volcanism, cryosphere, whales, airgun, anthropophonic explosion). Thus, the aim of this study is to develop an automatic source discrimination tool to support operational monitoring.

We analyze PMCC extractions from stations in the Atlantic (HA10), Indian (HA01, HA04, HA08) and Pacific (HA03, HA11) oceans over a period spanning January to Decembre 2023. The association to a source type is made in two stages. (i) We apply wave parameter and acoustic indices conditional statements to select typical signals with MaxAmp above 1.5 Pa for each type of source, except for airgun with MaxAmp of 1 Pa. (ii) The resulting catalog of extracted records are used to train a convolutional neural network of two layers and calibrate it by conformal prediction with Least Ambiguous set-value Classifier (LAC) score and a nominal error level of 0.05. All detections with a MaxAmp greater than 1 Pa are associated with one or more source types.

Over the year 2023, 111,260 coherent waves were extracted by PMCC on the 8 hydrophone triplets, of which 14,028 were associated to a source type using the ad hoc conditional statements. These records are associated to the right source type by the trained neural network at 92.5%. Overall, the classifier associated 75 ± 6% of records with one source. Significant differences in performance were observed between the hydrophone triplets. Results were lowest at hydrophone triplet HA10N (< 65%), while they were highest at HA04N (> 80%). This difference is due to the soundscape, with certain sources (earthquakes, volcanoes and croyspheres) being more difficult to discriminate. The criteria used to compile the reference catalog need to be improved to discriminate more accurately detections by source type.