Deep Learning to Locate Seafloor Landslides in High Resolution Bathymetry

Abstract:

The advancement of technology in remote sensing systems has made available increasingly higher resolution elevation datasets, allowing the pursuit of innovative methods and models for characterizing terrestrial and aquatic landscapes. Additionally, machine learning (ML) capabilities increase as faster processors and more efficient algorithms are developed. ML and neural network models have proven effective in the detection of terrestrial landslides from satellite imagery, and similar methods can be used to develop a model that can learn to detect submarine landslides from high-resolution bathymetry.

In the northern Gulf of Mexico, submarine landslide events are common natural hazards that risk the structural integrity and longevity of offshore oil and gas infrastructure. Publicly available high-resolution bathymetry derived from 3D seismic analysis allows for the identification of submarine landslides and their attendant features. We present an object detection deep learning model that locates submarine landslide-related features automatically throughout the bathymetry raster. The input image composites the bathymetry layer along with derivatives that have been found to be important for submarine landslide detection including shaded relief, slope, aspect, curvature, profile curvature, and plan curvature. Training data is compiled from various open source data sets, as well as digitized and manually identified data. Using the available training data, the object detection model is trained to identify various landslide features that can be visually identified in the imagery. These include topological features that could result from a landslide event and include scarps, mounds, thalwegs, channels, flows, fans, slumps, and the landslide extent.

The results of this ML-enhanced landslide detection model will accelerate the identification and prediction of submarine landslides on the seafloor. These results can be integrated into risk assessment workflows for marine and offshore environments. This presentation includes the analytical framework from data collection to prediction and highlights the current findings and challenges to stimulate further discussion.