Presenter: Jennifer Yin (Stanford University)

Description:
The accurate mapping of topography and bathymetry are critical for water resource management, navigation, and flood prediction. Close range remote sensing using UXVs have enabled high-resolution sensing of coastal and riverine environments at sub-meter scale. Meanwhile, space-borne methods have facilitated the development of global digital elevation models (global DEMs) albeit at lower resolution. Global DEMs are limited in their ability to estimate the elevation of a channel due to their horizontal resolution (~90m) which averages out important channel features, as well as inaccurate elevations below the water surface due to sensor limitations. As a result, global DEMs typically have elevated channel bottoms compared to that captured by close-range methods, which limits their utility. Historically, geomorphic relationships have been used to quantify this error, but with limited success due to the complexity of the problem in a wide range of environments. In this work, we propose to correct the channel depth using a theoretical framework describing the relationship between the unresolved channel depth, resolution, and channel width. We compare high-resolution LIDAR data and surveyed bathymetry data at 12 areas across the US with several global DEMs. We develop a physics-informed machine learning model using an XGBoost algorithm and environmental predictor data including land use, slope, watershed area, elevation, terrain variability, and water occurrence to predict the channel depth. The ML model shows high predictive power (R² = 0.82 and R² = 0.8 for two test sets). By using globally available topographic and hydrologic features, the ML model is globally scalable and can enhance estimates for bathymetry of channels where close-range remote sensing is unavailable.