Updates on global sediment thickness using geospatial machine learning

Abstract:

We present here a geospatial machine learning (GML) estimate (map) of global seafloor sediment thickness based on geographically sparse two-way time picks of reflection seismic data, and subsequent depth conversion based on a robust model of sediment compression. The thickness of sediment on the seafloor, and its rate of accumulation are fundamental to understanding such diverse geologic processes as continental denudation, carbon sequestration and cycling, hydrocarbon formation and accumulation, and seafloor slope instability. However, direct, or even indirect observations of sediment thickness are quite sparse, occurring mostly in the form of two-way travel times from long-transect reflection seismic, and ocean bottom seismometer data – both typically acquired using high-sound-level (air-gun) active sources. The seafloor and crust are picked by trained interpreters, yielding a two-way-time difference which can then be depth converted and used to update the global prediction- a K-nearest-neighbor geospatial machine learning algorithm.

Transoceanic seismic lines, once relatively common, have become quite rare, but many that were acquired decades ago have been archived and are publicly available via services such as GeoMapApp. Many of these data only exist now as digital images (rather than waveforms) scanned from paper plots. A key aspect of GML is the ease by which a new prediction can be generated with every new observation – the need for extensive interpretation of new data is minimized. This has motivated our effort toward developing an additional machine learning algorithm which we train to pick the seafloor and crust in the scanned image, thus providing additional “observations” upon which to make geospatial predictions. We discuss the results of our machine learned sediment thickness estimate, and compare it with estimates made via more traditional means.