Exploring the submerged valley of lake Guerlédan using multibeam echosounder watercolumn data and a deep learning network - neural network weigths

Guerlédan Lake is a large artificial reservoir located in central Brittany (France) created by flooding a valley and its associated infrastructures. As a consequence, the water column contains a wide variety of natural and anthropogenic targets, including submerged trees, buildings, and hydraulic structures such as old locks on the former river (Blavet). Because the lake depth varies in time (hydroelectric dam), these submerged obstructions sometimes represent a significant risk for surface and underwater navigation safety. Besides, their detection is of high interest for underwater archaeology and for automatic analysis of fishery echosounders data (fish schools versus trees). To detect these targets a Kongsberg EM2040C multibeam echosounder was mounted on ENSTA hydrographic survey vessel Mélité to acquire bathymetric and water-column data on the lake. Its very high ping rates in shallow waters (40 m deep) generate an extremely large dataset, making manual interpretation time-consuming, which motivates the development of automated detection approaches based on deep learning algorithms. We apply the YOLOv5 object detection framework to the water-column dataset. The aim is to detect acoustic signatures associated with trees and anthropogenic structures' echoes, to situate them geographically and also enable their subsequent three-dimensional representation to estimate their spreading in the water column. This dataset contains models that have been trained to detect obstructions (such as trees, houses and old locks) in different areas of Guerlédan Lake (Pouldu, Caurel and Kergoff). The neural networks are named after their training, validation and testing sets. For example, the C_K_P network was trained using the Caurel dataset for the training set, validated using the Kergoff dataset, and tested using the Pouldu dataset. Inference code (YOLOv5 with G3D files) is available on github repository.