An Innovative Approach to Seafloor Classification: Applying Advanced Unsupervised Machine Learning in Complex and Variable Bottom Types with Multispectral Satellite Imagery

Machine learning and unsupervised classification have been applied to seafloor classification and benthic mapping using multibeam echosounders (MBES), Light Ranging and Detection (LiDAR), airborne hyperspectral, underwater optical cameras, and satellite imagery. Regardless of the sensor, data dimensionality reduction is an integral processing step of seabed mapping workflows . The most commonly implemented pre-classification dimensionality reduction technique is Principal Components Analysis (PCA). TCarta proposes research into a powerful alternative to PCA, Unsupervised Topological Data Analysis (UTDA) to address the established methods’ shortcomings as applied to WorldView 2/3 multispectral imagery. This research will offer a direct comparison of these two dimensionality reduction tools for three scenarios. The first scenario, with training data input and in situ control, will be St. Croix, US Virgin Islands; the second, Puerto Rico, within the same geographical region with no further training applied to determine predictive ability and accuracy; and the final test area will be a remote location with no local training applied, Kiribati. This research will determine the feasibility of this alternative to established unsupervised classification methods for seafloor classification and spectra-based data dimensionality reduction. If successful this research could lead to a potentially highly scalable and predictive tool for multi-sensor marine geospatial analysis.