Machine Learning-Derived Channel Width and Depth for the National Hydrologic Geospatial Fabric in CONUS

This dataset focuses on reach-averaged estimation of river channel geometry, including top-width and depth, crucial for water flow prediction and flood mapping. Leveraging HYDRoacoustic data from the Surface Water Oceanographic Topography (HYDRoSWOT) program, we develop a machine learning model to predict channel geometry using data from the National Water Model, National Hydrologic Geospatial Fabric network, and other geospatial datasets. Our model demonstrates good fit within the Continental United States, with better performance observed in flatter regions. Covering nearly 2.7 million reaches in the US, this dataset is indexed to the National Hydrologic Geospatial Fabric network. However, in estuaries, particularly near river mouths where it widens into the coastal zone, there are no recorded Acoustic Doppler Current Profiler (ADCP) measurements in HYDRoSWOT, leading to unreliable model accuracy. Additionally, limitations in the training dataset, particularly the primary significant feature of ML models—100% annual exceedance probability discharge derived from the NWM—diminish skill in this exceedance probability, impacting the overall model goodness-of-fit. We provide estimates of channel geometry for two conditions: 100% and 50% annual exceedance probability, based on NWM historical retrospective data..