Machine Learning Approach To Geomorphometry-Extreme Flood Links In The Lower Colorado River Basin
Presenter: Lin Ji1
Co-Author(s): Hoshin V. Gupta, Ty Ferré, Tao Liu, D. Philip Guertin
Advisor(s): Dr. Victor R. Baker
1Department of Hydrology and Atmospheric Sciences, University of Arizona
Extreme flood hazards are common in the Lower Colorado River basin (LCRB) due to the complex terrain and entrenched river channels. However, the relationships among extreme floods and various basin morphometric factors have heretofore remained significantly uncertain in the LCRB. In this study, we extract 41 basin morphometric parameters (including basin geometry features, drainage network features, drainage texture features and basin relief features) for 371 watersheds in the LCRB from a 10-m DEM. A hybrid of K-means clustering, and Random Forest (RF) approach is conducted to estimate maximum annual peak discharge (MAP) and maximum annual peak discharge per unit area (UP). The results also indicate that the most consequential variables for predicting the MAP are the basin geometry features and the drainage texture, while the most important factors contributing UP are basin relief characteristics and stream networks. Three unsupervised flood areas are generated using K-means methods. Regional regression results indicate that the hydrological response is better for each flood zone than for the hydrological regime of the whole region. The ML approach can effectively discover multidimensional patterns in the available big datasets, further revealing that basin morphometry can aid understanding of the physical behavior of watersheds with respect to extreme flood events. The resultant links are essential for promoting working hypotheses for understanding regional large flood behavior and thereby improving flood risk analyses and provide critical data for informative action in flood mitigation for the Southwestern U.S.