A semi-supervised learning approach to polymetallic nodule parameter modeling

We explore spatial modeling of polymetallic nodule parameters in the Clarion-Clipperton zone (CCZ) using a semisupervised learning approach. Spatial models that utilize the factors affecting nodule formation or their proxy variables, are a useful tool for characterizing the CCZ nodule deposits and their economic potential. Some such models based on neural networks have been explored in the literature, but these approaches have employed supervised learning. These rely on hand-modeled features to incorporate the effect of topography, one of the key factors affecting nodule formation. We employ unsupervised learning via auto-encoders or principal component analysis to efficiently capture the effect of local topography around a point being modeled, and express it in terms of a few features. We then use these features for subsequent supervised learningbased nodule parameter modeling. Thus, this is overall a semisupervised approach. We show that the efficient incorporation of bathymetric information into these features yields better modeling performance as compared to using hand-modeled topographic features.