Item

Abstract

The prediction of global sedimentation rates is a critical aspect of comprehending various phenomena, such as carbon sequestration and seafloor stability. With the advent of machine learning, there has been an increasing trend towards utilizing these techniques to predict quantities on the seafloor, including porosity and total organic carbon. Several studies have been conducted to predict sedimentation rates[cm/ka] and mass accumulation rates[Mt/yr] using techniques such as k Nearest Neighbors and neural networks. However, a significant issue with these methods is the observed tendency for the models to over-predict sediment budget, a crucial parameter in evaluating the validity of predictions. This bias may stem from the paucity of low sedimentation rate data points on the continental shelves, where most scientific expeditions concentrate on high sedimentation rate areas. This bias disregards valuable data on low sedimentation rates, which would be useful in developing a comprehensive global map of sedimentation rates. To address this issue, there are several potential methods. The first involves identifying physical parameters, such as porosity, sediment grain size and bottom currents, that correlate with low sedimentation rates and can be utilized to identify these areas. The second involves utilizing few-shot learning, a machine learning technique that allows for training with limited low sedimentation rate labels. The third method involves the use of numerical models in small spatial areas, which can offer further insights into the processes behind low sedimentation rates and provide more data for analysis.