Data-driven prediction of global river width

Large-scale river models typically parameterize channel shapes, i.e., bankfull width and depth, based on discharge (Q) or drainage basin area (A) with a power law model, or based on stream order with look-up tables. However, these are highly simplified representations of real-world channel geometry, which reflect our limited understanding of factors controlling its spatial variability and may lead to great uncertainty in global hydrologic/hydraulic modeling. During the past few years, progress in remote sensing has made available global snapshots of river planform geometry, which can potentially inform the channel shape parameterization in global river modeling and improve our understanding of what influences channel geometry variability. In this study, we first assess how well state-of-the-art river models capture the spatial variability of global river width by jointly using three recently available global databases: (1) the Global River Width from Landsat (GRWL), (2) a global river discharge database, and (3) MERIT-Hydro, a high-accuracy global DEM at 3-arcsec resolution. We find that only ~30% of the river width spatial variability can be explained using Q, A, or stream order as the single predictor, based on ~0.2 million reach-averaged river width data samples calculated from GRWL. Then, we propose a machine learning approach to estimate river width, taking into account physical and non-physical environmental covariates for river width estimation, including channel slope, sinuosity, latitude, elevation, climatic aridity, leaf area index (LAI), soil clay/silt/sand content, human water use (e.g., irrigation, industrial, domestic withdraw), and urban fraction. A random forest (RF) regressor is trained with 5-fold cross validation by incorporating these covariates, and the explained spatial variability of river width can be increased to ~70%. Beyond Q and A, covariates related to soil and channel sinuosity are found to have high feature importance, possibly related to their roles to indicate availability of sediment supply in the catchment. Results from this study may help complement the traditional hydraulic downstream theory from a data mining viewpoint, and eventually be used to improve channel shape parameterization in global river modeling.