New Approach for Underwater Dunes Segmentation Using Deep Learning

In bathymetry, the detection and mapping of underwater dune areas are crucial for coastal development planning, maritime navigation, coastal resource management, and the preservation of coastal biodiversity. Digital Bathymetric Models (DBMs) are frequently used to address this issue. However, segmenting dunes from such data presents a significant challenge due to data insufficiency and their specific characteristics. Moreover, current solutions in the literature often lack generalization and require substantial adaptation for each study area. We refined a Mask R-CNN model with pretrained weights using our dune data from the Saint Lawrence River estuarine zone. Our focus was primarily on data preparation and adapting the network to maximize extracting underwater characteristics. Our approach shows promising results in terms of generalizing a large amount of data and improving the quality of feature maps extracted from the backbone. At this stage, we conclude that having a substantial amount of available data and utilizing transfer learning are advantageous for ensuring meticulous segmentation of underwater dunes.