Global Marine Gas Hydrate Occurrence Using Random Decision Forest Prediction

Abstract:

We have applied machine learning, specifically the technique of random decision forests (RDF), to predict densely spaced values of sparsely sampled seafloor sediment attributes relevant to gas hydrate occurrence. The results of global gas hydrate stability models using these new grids are similar to previously published predictions (the newly derived heat flow alone changes pore space volume in the global gas hydrate stability zone by ~3%), but our model inputs are statistically rigorous estimates (including uncertainties) of sub-seafloor sediment properties. Specifically we use as input recently updated, sparsely sampled, yet globally extensive datasets of seafloor temperature, salinity, porosity, organic carbon content, and fluid flux. The RDF estimate is based on empirical statistical relationships between the relevant parameters and other parameters for which we have more densely sampled estimates (e.g. water depth, seafloor temperature, mixed layer depth, sediment thickness, sediment grain type and crustal age). We create additional attributes by applying statistical analyses and physical models to existing densely sampled attributes. These statistics include mean, median, variance, and other parameters, over a variety of ranges from 5 to 500km. The physical models include established models of compaction, heat conduction, and diagenesis, as well as recently derived estimates of fluid flux at convergent margins. Over 600 densely sampled attributes are used in each prediction, and for each predicted grid, we calculate the relative importance of each input attribute. The RDF technique and resulting sediment model also show promise for global models outside the discipline of gas hydrates.