Global automated extraction of bathymetric photons from ICESat-2 data based on a PointNet++ model

Abstract

We developed and tested a fully automated method to extract bathymetric photons globally from ICESat-2 ATLAS data by leveraging the PointNet++ model, which performs well for classification and segmentation of point clouds. Training data was collected from areas spanning > 100 degrees of latitude and encompassing a wide range of seafloor characteristics and variations in ICESat-2 data features. We explored a range of data compositions and model settings, and optimized them for model training. The final model obtained precision, recall and F1 scores of 0.9291, 0.9315 and 0.9303, respectively, and achieved an Intersection over Union (IoU) of 0.6351 for sites that contain detected seafloor. Model performance varies between test sites, with most errors occurring when there is a high density of noise photons near the seafloor. This study provides evidence for the global applicability of a trained PointNet++ model to automatically extract bathymetric photons from ICESat-2 data. The model is publicly available for use and further development.

Keywords

• Bathymetry
• ICESat-2
• Deep learning
• PointNet++