Coastal systems are extremely dynamic environments exposed to many hazards, making accurate and regular monitoring a major challenge, particularly in the context of global change and sea level rise. In this frame of reference, high-frequency, high-resolution coastal Video Monitoring Systems (VMS) have been installed on three megatidal (tidal amplitude > 9 m) sites of Normandy (France) including a sandy beach at Villers-sur-Mer, a pebble beach at Etretat and a composite beach at Hautot-sur-Mer. This article proposes the use of Mask R–CNN to process images acquired at these sites and perform the automatic segmentation of the visible bodies of water in order to extract the waterline. The extracted waterlines are associated with a measured water level, which makes it possible to reconstruct the topography of the beaches at the scale of the tidal cycle. After training the neural network on manually labeled data, the segmentation by Mask R–CNN is very efficient by achieving a satisfactory segmentation on 69.87% of the images of Villers-sur-Mer, on 67.11% at Hautot-sur-Mer, and on 97.33% at Etretat. Once the waterlines have been extracted and georeferenced, the reproduction of the beaches’ morphology is satisfactory (averaged vertical RMSE = 28 cm). These results confirm that segmentation by Mask R–CNN is a particularly powerful tool that allows efficient and low-cost monitoring of the evolution of beach morphology, particularly in response to marine conditions. Its capabilities to detect and segment bodies of water while not being affected by the various sources of noise make it a notably effective tool for coastal science applications.

沿海系统是一个极其动态的环境，暴露在许多危险之中，这使得准确和定期监测成为一项重大挑战，特别是在全球变化和海平面上升的背景下。在此参考框架中，已在诺曼底（法国）的三个巨潮（潮汐幅度 > 9 m）地点安装了高频、高分辨率沿海视频监控系统 (VMS)，包括 Villers-sur-Mer 的沙滩， Etretat 的卵石海滩和 Hautot-sur-Mer 的复合海滩。本文提出使用 Mask R-CNN 来处理在这些站点获取的图像，并执行可见水体的自动分割以提取水线。提取的水线与测量的水位相关联，这样就可以在潮汐周期的尺度上重建海滩的地形。在人工标记的数据上训练神经网络后，Mask R-CNN 的分割非常有效，对 Villers-sur-Mer 的 69.87% 的图像、Hautot-sur-Mer 的 67.11% 以及在97.33% 在埃特尔塔。一旦水线被提取和地理参考，海滩形态的再现是令人满意的（平均垂直 RMSE = 28 厘米）。这些结果证实，Mask R-CNN 的分割是一种特别强大的工具，可以高效且低成本地监测海滩形态的演变，尤其是在对海洋条件的响应中。