UNRAVELING PATTERNS IN RIVER GEOMETRY: MULTI-MODEL MACHINE LEARNING FOR CONTINENTAL-SCALE PREDICTIONS

Hydraulic geometry parameters describing river hydrogeomorphic relationships are critical for determining channels’ capacity to convey water and sediment and are essential for viable, real-time forecasting. This simple but well-established power-law relationship of river geometry has been widely used to understand river systems worldwide during the past 70 years, which still remains the foundation of modern Hydraulic Geometry theories. Unquestionably, we have realized the limitations of this empirically observed relationship, including lack of precision, universal solutions, and physical principles. So where will the next advances in the development of hydraulic geometry be? We put forward that harnessing big data with Machine Learning (ML) approaches can be one solution. In this work, we have assembled a comprehensive U.S. Geological Survey (USGS) river measurement dataset, called HYDRoSWOT, and developed novel data-driven approaches to better estimate river geometries over the Continental United States (CONUS). We reported parsimonious high-quality river geometry models to understand the river cross-section width and depth under long-term median flow conditions, outperforming state-of-art architectures. Additionally, we have constructed the Continental United States River Geometry Database (USRGD) to characterize the nearly 2.7 million rivers and streams across the CONUS. This newly assembled database allows better measurements of river width and river depth, and direct quantification of with-depth ratio and total river and stream surface area (RSSA) for the very first time, to our best knowledge. This knowledge potentially paves the way for developing end-to-end models customised for many continental-scale hydrology models, such as the National Water Model’s Next Generation framework, to improve continental-scale hydrology simulations, especially flooding inundation mapping.