Recovering Geomorphic History From Stratigraphy: Application of Machine Learning Techniques to an Experimental Basin

Each bed in a stratigraphic succession encodes the set of geomorphic conditions that generated it as well as the time over which those processes occurred. However, recovering that information from the rock record is nontrivial due to the incomplete nature of stratigraphy. Herein, we evaluate the utility of machine learning approaches to recover this information using a data-rich, fluvio-deltaic experiment with precise tracking of its geomorphic surface, resultant stratigraphy, and spatiotemporal distribution of sub-environments (e.g., channels, lobes, wet overbanks, dry overbanks). Surprisingly, we find that relatively simple random forest models, with just a few predictor attributes, can reliably reconstruct the amount of sediment lost to erosion from a bed and the time missing from a bed with mean absolute errors of ~18% and ~6%, respectively. We show that even binary, categorical attributes can significantly improve model outputs. These findings are encouraging for the application of machine learning approaches to field-scale systems, if the correct non-dimensional framework can be constructed. Our initial efforts suggest it may be possible to recover the distribution of missing sediment on spatial-scales less than a meter and time-scales less than 10 4 years.