Improving Prediction Capability in Machine Learning via Spatial Wavenumber Analyses of Observational and Training Datasets

Abstract:

Machine learning has proven to be a valuable tool for predicting global seafloor properties - most notably in locations where physical measurements have not been made. Currently, it is widely accepted that meaningful seafloor property predictions can be made at high resolution using natively lower resolution training datasets. While these predictions remain adequate for relatively low spatial wavenumber predictands, higher wavenumber predictands result in poor predictions with little-to-no predictive capability. We demonstrate the significance of the spatial wavenumber of the training predictors, with respect to the observational dataset, for improving predictive capability using machine learning algorithms. Results show that for relatively high wavenumber predictands, such as magnetic and/or gravity anomaly datasets, predictions using training datasets with natively higher resolutions outperform those of lower resolution. This means that the plethora of training datasets available, most of which were collected at the 1°x1° resolution, will result in poor predictive capability, even when sampled at higher resolutions. In order to perform quality predictions using high wavenumber datasets, high wavenumber predictors must be acquired, signifying the need for increased resolution of global seafloor physical property grids, as well as global geologic property grids.