Deep Learning Reach-level Estimates of Mean River Depth at the Conterminous United States Spatial Scale

Abstract: Estimates of riverine channel geometry play a vital role in the physical representation of stream networks in models used to predict flood and drought conditions, manage water resources, and increase our knowledge of fluvial conditions under a changing climate. A well established body of literature exists that explains the relationship between channel geometry parameters width, depth, and velocity to instantaneous river discharge using a log-log linear power-law regression. In this study, a state-of-the-art deep learning regression model is presented and compared against the power-law method to evaluate their ability to estimate cross-sectional mean river depth. Results reveal three key findings, the neural network: (1) decreases RMSE by 22% verse a CONUS scale power-law equation, (2) reduces prediction variance across Strahler stream orders, and (3) generally outperforms regional power-law equations with an average decrease in RMSE of 8.7% Lastly, a reach-level CONUS dataset of estimated mean river depth is delivered.

The deep learning model was trained using the following features:

• AI - Mean aridity index of unit catchment - Trabucco and Zomer, 2019
• area - Upstream drainage area (km2) - P. Lin et al., 2020
• CLY - Mean clay content (mass percentage, %) of unit catchment - Hengl et al., 2017
• DOR - Stream segment degree of dam regulation (Scale 0. – 100.) - Grill et al., 2019
• Elev - Stream segment mean elevation - P. Lin et al., 2020
• K - Mean bedrock permeability of unit catchment surrounding stream segment - Huscroft et al., 2018
• LAI - Mean leaf area index of unit catchment - Zhu et al., 2013
• order - Strahler-Horton stream order - P. Lin et al., 2020
• P - Mean bedrock porosity of unit catchment - Huscroft et al., 2018
• QMEAN - Stream segment mean annual discharge (m3/s) - P. Lin et al., 2019
• Sin - Stream segment sinuosity - P. Lin et al., 2020
• Slp - Stream segment mean longitudinal slope - P. Lin et al., 2020
• SLT - Mean silt content (mass percentage, %) of unit catchment - Hengl et al., 2017
• SND - Mean sand content (mass percentage, %) of unit catchment - Hengl et al., 2017
• stream_wdth_va - Measured stream cross-sectional width (m) - Canova et al., 2016
• Urb - Mean urban fraction of unit catchment - Liu et al., 2018

The deep learning model was trained using the following label:

• mean_depth_va - Measured stream mean depth (m) - Canova et al., 2016

Predictions of mean depth were made by replacing stream_wdth_va from Canova et al., (2016) with bankfull width estimates (width_m) from P. Lin et al., (2020). Missing records from the P. Lin et al., (2020) dataset were excluded when making predictions, thus there are missing reaches in the dataset.