Seafloor bedforms are sedimentary structures that can reveal the local hydrodynamics conditions as they are the result of the bottom sediments' response to the dominant flow. The studying of these dynamic bottom shapes is important given that they can provide auxiliary information for the mapping of benthic habitats and can present a risk to navigation and marine structures. Mapping of the sea bottom is often done with dense and spatially extensive 3D bathymetry data (point clouds) resulting in a more precise representation of the targeted area. However, the most common praxis in this field is to rasterize the original data for the easiness in computation and lack of yet well-established methodology for 3D bathymetric data processing. As a consequence, the rasterization process causes information loss and increase of data processing time. The advantage of points clouds is that they comprise a larger volume of information in the same file, e.g. depth, intensity, RGB, and point classes, enabling a closer representation of reality and simultaneous generation of multiple products as potential habitat maps, Landscape Information Model (LIM), and denser Digital Bathymetry Model (DBM). Therefore, the purpose of this work is to apply a modified U-Net convolutional neural network for detecting and classifying bedform types in the bathymetry point cloud. The methodology will be applied to two datasets collected on the Espirito Santo Continental Shelf: Recifes Esquecidos (RE) and Doce River (DR).The following methodological steps will be carried out: (1) data collection, (2) generation of bathymetry derivatives as slope, curvature, geomorphons, aspect, and data tiling up for data augmentation, (3) image labeling, (4) model implementation including the training, validation, and testing steps, (5) calculation of models performance metrics: Intersection over Union (IoU), mean Average Precision (mAP), recall, and precision. This study will also present a performance comparison between the modified U-Net and a Random Forest (RF) forest. The selected areas are very rich in sedimentary features, so it is expected as the final product a classified points cloud in transversal, parallel, transitional, and artifacts (errors related to the surveying method) classes. It is also expected to have better performance from the non-convolutional model, RF.
Read full abstract