Predicting Bottom Current Deposition and Erosion
Summary
With the growing need for ocean infrastructure development, geo-resource extraction, sustainable energy solutions, and seafloor slope stability, iden- tifying zones of sediment deposition and erosion caused by ocean bottom currents is increasingly important. However, knowledge about this niche topic is limited, as only a small percentage of the ocean floor has been di- rectly mapped through seismic or sonar imaging. Sampled locations pro- vide precise vertical constraints at specific points, but extending this infor- mation beyond those points is challenging. Approaches that utilize numer- ical model solutions and sedimentological measurements have displayed promising results that have given insight on roughly where many of these zones of sediment deposition and erosion occur. The sake of this research is to expand on these ideas and turn to a machine learning approach that will predict these zones. In particular this method will implement and uti- lize a random forest model that will create predictions based on specified explanatory model input. The objective of the algorithm is to utilize the sparsely sampled observational data set alongside densely gridded predic- tors to generate statistically predictions in areas lacking physical measure- ments. The predictions will be based on three data sets (1) Shear stress of bottom currents from the HYCOM numerical ocean model; (2) Sedimenta- tion rates based on ocean lithospheric age and sediment thickness from the GlobSed Model; (3) Measured extents of bottom current deposits obtained from sonar observations. The specified region in the Atlantic Ocean con- tained most observations and therefore will be used to quantify the train- ing set for evaluating and validating the model’s accuracy. With the new approach, shear stresses and sedimentation rates can be utilized to train the model, facilitating the understanding of bottom current deposition and erosion. Specifically, the presence of sediment deposition will be measured using the model. The prediction on the training set has an overall model accuracy of 0.78, however the precision for predicting the presence of a contourite itself is only 0.08. Moreover, these metrics were calculated by leave-spatial-block-out-cross validation to test the extrapolation capacity of
the model. The final prediction on the global model gives an overall model accuracy of 0.76 and a prediction a contourite itself of 0.04. Results show some deposition predictions that coincide with established patterns rec- ognized by experts in the field. Deposition is typically observed near con- tinental shelves and other topographic obstacles, while non-deposition is common near mid-ocean ridges. Notably, some well-known drifts in the North Atlantic Ocean, such as the Hatton, Feni, and Gardar Drifts, are evi- dent in the results. However, the model appears to be overfitting. Finally, a data gap identification highlights areas where additional samples are nec- essary to enhance the overall prediction accuracy.