Machine Learning Augmented Time‐Lapse Bathymetric Surveys: A Case Study From the Mississippi River Delta Front

• Published in: Geophysical research letters 2020-05, Vol.47 (10), p.n/a
• Main Authors: Obelcz, Jeffrey, Wood, Warren T., Phrampus, Benjamin J., Lee, Taylor R.
• Format: Article
• Language: English
• Subjects: Bathymetric surveys; bathymetry; Cables; Communication; Communication cables; Data; Deltas; Depth; depth change; geohazards; Geological hazards; Learning algorithms; Machine learning; Ocean floor; Polls & surveys; Rivers; Spatial variability; Spatial variations; Stability analysis; Submarine pipelines; Training; Water depth
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2020GL087857

The subaqueous Mississippi River Delta Front is prone to seabed instabilities >1 m of vertical bathymetric change per year, but the ability to predict the location and magnitude of instability‐driven depth change is limited. Here we demonstrate that data‐driven geospatial models can predict MRDF depth change from a small amount (1% of full coverage) of training data. We predict depth change at 100 m2 resolution between 2005 and 2017 over a ~100 km2 area. Models trained on ~1% of full‐coverage depth change data produce comparable and relatively low average predicted depth change errors (1–2 cm). K‐nearest neighbors best reproduce the spatial variability of depth change and can interpolate and extrapolate from training data. This approach has immediate applications for geohazard monitoring on the MRDF and other geologically similar settings and can be applied in other settings if the drivers of depth change variance are well known. Plain Language Summary Many river deltas worldwide have unstable seabeds that move vertically and laterally on yearly intervals, threatening seafloor infrastructure such as communication cables and pipelines. A common method to measure seabed movement is with repeat bathymetric surveys, which are expensive and require considerable expertise. Here we introduce a new method that greatly reduces the resurvey data required to assess seabed instability: using geospatial models to estimate depth change where it is not measured. We test three different models trained on a small amount (1% of complete resurvey coverage) of Mississippi River Delta Front (MRDF) depth change data and find the predicted depth change is comparably accurate (1–2 cm error) for all methods. We recommend the K‐nearest neighbors machine learning method of the three, as it produces the most geologically realistic estimates and can predict depth change from multiple types of resurvey data. This method is applicable to MRDF‐like deltas and can be modified to predict depth change in different geologic environments as well. Key Points Data‐driven geospatial models are used to predict depth change between 2005 and 2017 over ~100 km2 of the Mississippi River Delta Front Training models on 1% of available data resulted in 1–2 cm average predicted error depth change estimates for three methods tested The K‐nearest neighbor machine learning algorithm is recommended for geologically realistic depth change models