Automated classification of valid and invalid satellite derived bathymetry with random forest

Abstract

Recent decades have seen rapid growth in algorithms and workflows for generating bathymetry from multispectral satellite imagery, with the output typically referred to as satellite derived bathymetry (SDB). An inherent challenge is that, while SDB algorithms generally output a value for every pixel in each input scene, the value of any particular raster cell may not represent a valid depth. Typically, large portions of a satellite scene will correspond to optically deep waters, where SDB depth retrieval is impossible. Boats, boat wakes, breaking waves, clouds, and land can also cause erroneous depth estimates. Compounding this challenge is the fact that SDB algorithms tend to be poor at self-diagnosing invalid retrievals. As a result, SDB grids often require substantial manual editing to generate reliable bathymetric digital elevation models (DEMs). This study investigates the ability to automatically classify valid and erroneous bathymetry from SDB grids using random forest and surface features generated from the SDB grids, spectral bands values, and band indices. The trained models achieved mean intersection over union accuracies of 73–91% at five geographically-diverse test sites of differing sizes, seafloor morphologies, substrates, and wind and wave climates. These methods can be used to rapidly assess satellite image-based bathymetry grids and are currently being adapted for operational use within the National Oceanic and Atmospheric Administration’s National Ocean Service.