Sediment types identification and seafloor geomorphology classification through machine learning in Buzzards Bay, Massachusetts

Haoran Liu, Louisiana State University, Department of Oceanography and Coastal Sciences, Baton Rouge, LA, United States, Kehui Xu, Louisiana State University, Oceanography and Coastal Sciences, Baton Rouge, LA, United States and Bin Li, Louisiana State University, Department of Experimental statistic, Baton Rouge, LA, United States
Abstract:
Traditional interpolation of acoustic seafloor mapping data are expensive and sparse, which also has large uncertainty due to resolution limits. Some bathymetric and geologic features may be lost, especially when detailed topography information is needed. This study compares the performance of a list of supervised and unsupervised classification methods for predicting and mapping seabed types with bathymetry and backscatter data. The study site is Buzzards Bay, southeastern Massachusetts, USA, an embayment with mean water depth of 5 to 10 meters. Previously published mapping methods for homogeneous regions rely on the contouring changes of grain size. However, the substrate variations in the Buzzards Bay area is too complicated to be resolved by traditional interpolation and griding methods. High-resolution bathymetry and backscatter datasets with full sea-floor coverage, supplemented with sediment samples and bottom photographs, were collected from 2005 to 2011 by the U.S. Geological Survey (USGS), in cooperation with the Massachusetts Office of Coastal Zone Management (CZM). In terms of grain-size statistics, a total of 870 sediment samples are classified as rock, gravel, sand, and mud.

This study first conduct the calculations on primary bathymetry and backscatter data and a range of secondary derived features (Curvature, Bathymetric position index, Roughness, Northness, Eastness, Moran’s I and Sobel filter). Five supervised and unsupervised classification techniques were tested in the study area: Trees, Random Forest, Support Vector Machines, Convolutional Neural Network, and K-Nearest Neighbour. The models were trained, respectively, by using three different inputs, which included i) only primary data; ii) primary data and derived features of seafloor geomorphology; iii) selected important features. The model performances were validated through separate test data with both sediment samples and bottom photographs via image analysis. The distribution map of sediment texture and seafloor geomorphology map generated from Random Forest provides more detailed information of heterogeneous sea-floor character in Buzzards Bay when compared to traditional interpolation-based techniques.