Using artificial neural networks to quantify erosion by mega-floods and the implications for paleoflood discharge reconstruction
Abstract
Pleistocene outburst floods from Glacial Lake Missoula spilled across the Columbia Plateau and generated an anabranching network of channels and bedrock canyons in eastern Washington, USA. Downstream at the Washington-Oregon border, the flood waters converged and flowed through a single channel within the Columbia Gorge, making it an ideal location to quantify paleo-flood discharge. Geomorphic evidence, such as scabland topography and dissected lava flows, indicate bedrock erosion enlarged the gorge as the largest floods in Earth's history were routed through it. However, the magnitude of the increase in canyon cross-sectional area is unknown, which convolutes our ability to reconstruct paleo-flood discharge. To address the unknown initial gorge topography encountered by floods, we used an artificial neural network machine learning technique developed by Mey et al. (2015) to reconstruct the pre-flood topography along 150 km of the Columbia Gorge. The artificial neural network was trained using topographic data above the maximum elevation of flood-induced erosion and the training data were used to predict the paleo-topography in the lower-elevation part of the gorge by reconstructing valley walls. We used a 2D hydrodynamic model to route flood discharges through the modern and reconstructed Columbia Gorge topography to assess differences in stage-discharge relationships. Preliminary results indicate that floods through the reconstructed paleo-topography can require up to 30% less discharge to reach the same elevations as floods through the modern, flood-eroded topography. For example, a flood discharge of 10x106 m3s-1 through the reconstructed paleo-topography produces water surface elevations approximately 30 m higher than the same discharge routed through the modern topography. Prior work has shown that a 10 m change in flood stage corresponds approximately to a 1x106 m3s-1 change in discharge in the Columbia Gorge. Hence our findings show that accounting for flood-induced topographic changes can revise paleo-flood discharge estimates by several million m3s-1. Our study highlights how machine learning can advance understanding of surface processes, as the complexity of the Columbia Gorge topography precludes the application of more routine methods of topographic reconstruction.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMEP54C..12D
- Keywords:
-
- 1824 Geomorphology: general;
- HYDROLOGY;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS;
- 4217 Coastal processes;
- OCEANOGRAPHY: GENERAL