Machine Learning Analysis of Flood Hydrology and Fluvial Geomorphometry On Earth and Mars

Abstract

This dissertation synthesizes the findings from three key papers that leverage machine learning (ML) techniques to advance our understanding of geomorphometry's role in hydrological processes and the paleoclimate, offering novel insights into flood prediction on Earth and ancient water flows on Mars. The first paper reveals the intricate links between extreme flood events and basin morphometry within the Lower Colorado River Basin (LCRB), a region prone to significant flood hazards due to its complex terrain and entrenched river channels. By extracting 41 geomorphometric parameters across 371 watersheds from a 10-m digital elevation model (DEM), the study employs a novel hybrid approach combining K-means clustering and Random Forest (RF) analysis. This methodology successfully predicts maximum annual peak discharge (MAP) and its unit area (UP), revealing that basin geometry and drainage texture are significant in forecasting MAP, while basin relief and stream networks crucially influence UP. The research identifies three distinct flood zones, demonstrating that regional hydrological responses can be more accurately modeled through targeted analysis of these zones, thereby providing a robust framework for flood risk assessment and mitigation in the Southwestern United States. The second paper addresses the limitations of traditional linear regression methods in regional flood frequency analysis (RFFA) in the LCRB by implementing advanced machine learning (ML) models to capture the complex interactions between flood discharges and a wide array of predictor variables. Analyzing data from 280 hydrometric stations and incorporating 94 predictor variables, including geomorphometric, climatic, land use, and soil type factors, the study compares the efficacy of various ML models, including Linear Regression, Ridge Regression, Lasso Regression, K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), and Multilayer Perceptron (MLP) regression. The findings indicate superior performance of RF, GB, and MLP models, especially in estimating frequent smaller floods, highlighting the critical role of geomorphometric features combined with the climate, land use cover, and soil types in determining flood frequency. This comprehensive approach significantly enhances the accuracy of flood frequency and magnitude estimation in the LCRB and underscores the value of integrating geomorphometric analysis into hydrological modeling. In the third paper, the application of ML extends beyond Earth to analyze Martian valley networks, aiming to infer the planet's paleoclimatic conditions. Employing a global dataset for Earth and Mars, and analyzing basin networks, shapes, drainage textures, relief, longitudinal profiles, and hypsometric metrics, the study compares geomorphometric characteristics across Earth's climate regions with Martian landscapes. The research estimates that 91% of Martian landscapes, especially the extensive highland valley networks, are analogous to terrestrial arid landscapes, characterized by smaller, narrower, steeper, and younger watersheds with straighter longitudinal profiles. This analysis suggests that the Martian valley networks primarily formed under arid conditions through intermittent, concentrated precipitation events, supporting the theory of an ancient Martian hydrological cycle and the possible existence of a substantial ancient water body, theorized as “Oceanus Borealis.” These findings not only contribute to a deeper understanding of Martian climate and geomorphology but also highlight the potential for ML techniques in geomorphometry to bridge terrestrial and extraterrestrial studies. Collectively, these papers underscore the transformative potential of machine learning in geomorphometry for enhancing flood risk management on Earth and advancing our knowledge of planetary geomorphology and ancient climates, particularly on Mars.