Constraining Remote River Discharge Estimation Using Reach-Scale Geomorphology

Recent advances in remote sensing and the upcoming launch of the joint NASA/CNES/CSA/UKSA Surface Water and Ocean Topography (SWOT) satellite point toward improved river discharge estimates in ungauged basins. Existing discharge methods rely on prior river knowledge to infer parameters not directly measured from space. Here, we show that discharge estimation is improved by classifying and parameterizing rivers based on their unique geomorphology and hydraulics. Using over 370,000 in situ hydraulic observations as training data, we test unsupervised learning and an expert method to assign these hydraulics and geomorphology to rivers via remote sensing. This intervention, along with updates to model physics, constitutes a new method we term geoBAM, an update of the Bayesian At-many-stations hydraulic geometry-Manning's (BAM) algorithm. We tested geoBAM on Landsat imagery over more than 7,500 rivers (108 are gauged) in Canada's Mackenzie River basin and on simulated hydraulic data for 19 rivers that mimic SWOT observations without measurement error. geoBAM yielded considerable improvement over BAM, improving the median Nash-Sutcliffe efficiency (NSE) for the Mackenzie River from -0.05 to 0.26 and from 0.16 to 0.46 for the SWOT rivers. Further, NSE improved by at least 0.10 in 78/108 gauged Mackenzie rivers and 8/19 SWOT rivers. We attribute geoBAM improvement to parameterizing rivers by type rather than globally, but prediction accuracy worsens if parameters are misassigned. This method is easily mapped to rivers at the global scale and paves the way for improving future discharge estimates, especially when coupled with hydrologic models.