CPU Central Processing Unit CTD Conductivity Temperature Depth CW Continuous Waves ; Approches hybrides acoustique et machine learning pour l’étude du fonctionnement des écosystèmes marins ; CPU Central Processing Unit CTD Conductivity Temperature Depth CW Continuous Waves: Hybrid acoustic and machine learning approaches for studying the functioning of marine ecosystems

The objective of this internship is to develop a deep learning method to detect the bottom line in split-beam echosounder echograms. The issue at stake is to help the Institut de Recherche pour le Développement (IRD) to carry out more quickly the time-consuming treatment of the acoustic data collected during oceanographic cruises. The quantitative and qualitative analysis of acoustic signals backscattered from the seabottom up to the surface is used in the entire word for the evaluation of fish stocks i and for the monitoring of marine ecosystems. Large amounts of data are collected during scientific cruises at sea with echosounders. A set of tools developed by the IRD and the IFREMER (softwares MATECHO and MOVIES3D) are used to process this data in a semi-authomatic way. However, the tedious and time-consuming manual labelling of echograms is still needed to correct the bottom line detection errors. Using the data collected in recent years, operations of manual corrections done on these images have provided us with a database of several million of pings. The chosen method is a convolutional neural network (CNN) with a U-Net architecture that was trained with a database of labeled images from the IRD. The data used were acquired during three oceanographic cruises between 2019 and 2023 : SCOPES (nearshore Senegal), FAROFA3 (Fernando de Noronha (Brazil)) and PIRATA FR29 (Gulf of Guinea). The network was developed by training it on data acquired at 200kHz, a frequency shared by the three cruises. The network's parameters are then selected and the performance of the final network is evaluated. The final version of the network gives better results than the algorithm currently used by the IRD, both in terms of accuracy of the prediction (95.2% of accurate detection of the bottom line in 200kHz-echograms for the network against 27.9% for MATECHO) and duration of processing (two days for the network against fifteen for MATECHO). This study proved the potential of deep learning methods to process fishery acoustic data ...