PERFORMANCE OF MACHINE LEARNING TECHNIQUES IN RECOVERING GEOMORPHIC HISTORY FROM STRATIGRAPHY
Our dataset is derived from fluvio-deltaic experiment TDB-17-1, performed at Tulane University’s Sediment Dynamics Laboratory. This experiment occurred in a 4.2 m x 2.8 m x 0.65 m basin, with a constant sediment and water discharge, and a steady, uniform base level rise. The sediment mixture included analogs for coarse and fine grain sizes along with minor admixtures to reproduce cohesive behaviors. The main phase of the experiment was purely aggradational. Frequent laser scans of the experimental surface captured a point cloud of topography that was converted into a digital elevation model with mm-scale resolution horizontally and < 1 mm scale resolution vertically. This provides a precise time series of the erosion, deposition, and stasis geomorphic events embedded behind each stratigraphic bed as well as the knowledge of the depositional sub-environment (e.g., channel, lobe, wet/dry overbank).
A variety of machine learning techniques (i.e., random forests, neural networks, support vector machines) were applied with the goal of predicting the proportion of time represented by deposition, erosion, and stasis behind a given stratigraphic bed. Predictor variables included sedimentation rate, bed thickness, depositional sub-environment, and binary “tags” for a bed’s association with deep scours or long stasis. We find prediction errors generally of ~10% for the proportion of time spent depositing, eroding, or in stasis behind a given bed. We hypothesize, if appropriately applied, these techniques could potentially improve basin age models on the scale of kyrs to tens of kyrs.