Machine learning model selection for predicting bathymetry

This research investigates the viability of using Machine Learning (ML) for predicting bathymetry. We built and trained several models using ocean features aggregated from multiple sources and predicted bathymetry from the ETOPO dataset at a 2-min resolution. Each model was evaluated to identify a global best fit, however we found that none performed well on a global scale. When training on subsets of the world, we observed that some models performed significantly better, which led to developing a novel model selection technique that identifies the best performing model and most relevant features for a given geospatial coverage. This leads to improved predictions and more reliable results. This model selection technique can be generalized to be applied to any set of models.