Dietmar Müller1, Adriana Dutkiewicz1, Joyce Yu2
1EarthByte Group, School of Geosciences, University of Sydney, Sydney, Australia, 2CSIRO, Data61, Sydney, Australia
Deep-sea sediments record changes in climate, paleogeography, surface environments and biogeochemical cycles through geological time. Yet, their composition and distribution through time is not well known due to the sparse distribution of ocean drilling sites. We have developed a neural network-based methodology that allows us to produce deep sea lithology maps from sparse samples in combination with other features that can be reconstructed through geological time. These include bathymetry and features derived from paleo-climate models including sea surface temperature, salinity and upwelling intensity as a proxy for productivity. We test this model based on a present-day lithology, bathymetry and oceanography data set to map seafloor lithology with a probabilistic multi-class neural network model trained with about 12,500 sample points. We reduce seafloor lithologies to five end-member classes: (1) gravel/sand/silt, (2) clay, (3) calcareous ooze and other fine-grained calcareous sediments, (4) radiolarian ooze and (5) diatom ooze. K-fold cross validation techniques were adopted to avoid overfitting. A final accuracy score of 86% was achieved in performance for our classification model. We compare this with a previously applied support vector machine (SVM) approach. When applied to our five end-member classes its classification accuracy is 79%. Unlike the SVM, our neural network model also provides probability estimates, which helps quantify the level of uncertainty in predicting each lithology. Our method represents a low-cost way of obtaining lithology and paleo-lithology maps, and has the potential to be extended to produce sediment accumulation rate maps, allowing an analysis of the deep ocean sediment flux through time in a context of changing paleogeography and climate.
Dietmar Müller is Professor of Geophysics at the School of Geosciences, University of Sydney. He received his PhD in Earth Science from the Scripps Institution of Oceanography in 1993. With his EarthByte research group he combines tectonics, geodynamics, and sedimentology to build virtual Earth models from global to basin scales.