Machine learning for the assessment of socio-economic impacts of geophysical hazards
Abstract
As natural resources are increasingly pursued in previously uncharted territories, the number of accidents due to the novelty of the exploration are bound to increase. Simultaneously, as population both increases in density at a few stress points and spreads, perhaps thinly, over a much wider area, the socio-economic consequences of these accidents are multiplied. Mitigating the risk from these accidents is challenged by a lack of ability to properly predict their evolution and their impact. This is less due to data availability, and more a consequence of the lack of uniformity in analyzing this class of problems. In particular, while satellite imagery, economic and social data are available, and while detailed physics-based models are also available for predicting the evolution of the hazard (such as oil spill, volcano lava, mudslide, or fire), the coupling between physics-based and social-based models is typically very coarse, limiting their individual and collective predictive powers. We posit that this interface is governed by rules of conduct that, while hidden from an analyst, are universal to within modeling errors. While we do not seek an explicit expression of the rules, we will discover them by simultaneously and judiciously mining joint data about the physical, natural and human components. Once this structure is characterized, we will use it to constrain the task of conditional regression. This will provide the required mechanism to update predictions of both nature and human dynamics in a manner that is intrinsically consistent with learned rules of behavior.
We will focus our attention in this presentation on accidents describing oil spills in the Gulf of Mexico. The intrinsic structure of the human-nature interface will be characterized using diffusion maps and sampling along the associated manifold will be carried out using a projected Ito equation. We will demonstrate the utility of this construction to decision-making for risk assessment and mitigation.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH052...04G
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1830 Groundwater/surface water interaction;
- HYDROLOGY;
- 1832 Groundwater transport;
- HYDROLOGY;
- 1849 Numerical approximations and analysis;
- HYDROLOGY