Robotics and AI weave surface process narratives from rock geomorphology
Abstract
We share our experiences with data-driven discovery for geomorphology, enabled by robotics and AI. Motivated by our prior work in automated fruit counting for agriculture, we can detect and segment rocks in different types of geologic environments over a variety of spatial scales. We leverage drones for data collection, structure from motion algorithms for image registration and point cloud generation, and deep neural networks to process orthorectified maps estimated from the terrain point clouds. Distributions of 2D rock traits such as size, major-axis length, and orientation offer insights to the geomorphologic history of a field site and particularly on the nature of past ground motion in seismically active areas, given the three following considerations. First, the ground sampling resolution must remain consistent to avoid heteroscedasticity in rock traits. Failure to do so may result in errors in rock trait distributions through aggregation of incorrect detections or segmentation. The solution is active mapping and terrain-relative navigation when using drones, to ensure images are acquired at a constant distance from terrain, with optimal camera poses. Second, while 2D particle analysis has enabled scaling up granulometry, there are limitations in insights. For applications of particle transport, downslope grain size change, and fragile geologic features such as precariously balanced rocks, 3D is needed. 3D particle traits include the centroid or higher moments, contact relationships, surface area, volume, etc. Third, semantic segmentation of scenes delivers large volumes of georeferenced objects with uniform and objective traits. However, scientific validity must be evaluated before application to surface processes and geologic hazard. A major challenge is to develop metrics that constrain erosion rates, transport rates, ground shaking, and fault slip rates. Tools from computer graphics and robotics are important for understanding episodic movement and change in active landscapes. Physics engines enable realistic and accurate behavior (including rigid body dynamics) of objects. They are used for gaming environments delivering cognitive realism, and in robotics for algorithm development, and system testing. These tools can weave an accurate, rich, and dense surface process narrative .
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMS057...03D
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1914 Data mining;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS