Understanding Ocean Bathymetry using Machine Learning

Abstract:

For the last half century, geophysicists have been trying to understand the relationship between seafloor bathymetry and age. Dozens of physical models have been proposed, such as the half-space cooling model (Davis and Lister, 1974), PSM (Parsons and Sclater, 1977), and GDH1 (Stein and Stein, 1992). While these physical models provide an intuitive explanation for seafloor subsidence, there exist large-scale unexplained bathymetric discrepancies.

In this presentation, we introduce an empirical model with a data-driven machine learning approach. Besides seafloor ages, we also use the CRUST 1.0 dataset (Laske et al., 2013), which contains boundary thickness and density for a 9-layer model of the Earth: water; ice; upper, middle, and lower sediments; upper, middle, and lower crystalline crust; and below the Moho. We use the CRUST 1.0 crust type add-on to remove atypical oceanic crust as well as continental crust, leaving around 30,000 data points to train our model.

We compare the performance of several different machine learning models, including linear regression (OLS), support vector regression (SVR), and a simple multi-layer perceptron (MLP) neural network. Our preliminary tests using the raw CRUST 1.0 bathymetry data show that, of these models, SVR achieved the best performance, with an RMSE or 0.458 and an R2 of 0.735 on the test set. In comparison, PSM has an RMSE of 1.14 and an R2 of -0.61, while GDH1 has an RMSE of 1.04 and an R2 of -0.32.

Our machine learning model provides several benefits over traditional physical models. First, it allows a nonlinear function to be fit to the data using dozens of input parameters instead of just age and/or heat flux. This provides a theoretical upper limit on the accuracy that any physical model could achieve. Second, it allows us to eliminate all bathymetry that can be explained by thermal isostasy following seafloor age, leaving behind a more precise residual topography for quantifying other processes such as dynamic topography. Third, machine learning models do not rely on physics, and as such, don't suffer from potential biases in our current understanding of lithosphere structure. Although machine learning models aren't as interpretable as physical models, they provide multiple benefits for furthering research into this important topic.