A post-processing method to calibrate large-scale hydrologic models with limited historical observation data leveraging machine learning and spatial analysis

Expanding cloud computing power combined with increasingly available global satellite and remote sensing data creates new possibilities for hydrologic modeling at continental and global scales. However, emerging continental- and global-scale models face challenges in performing model calibration. Traditional calibration methods require vast historical and real-time observed data; data that often does not exist, may not be shareable for political reasons, or may otherwise be inaccessible to the model creators. We present a new method for hydrologic model calibration to remove biases in near-real-time using limited observational data. We perform bias adjustments by expanding upon an existing method that involves comparing the flow duration curves of the observed and simulated data. Our method maps simulated flows corresponding to a certain exceedance probability on the simulated flow duration curve to the flow for the corresponding exceedance probability on the observed data flow duration curve. We systematically extend these adjustments to ungauged basins by leveraging machine learning. We create K-means clusters separately for the simulated and observed discharge using Dynamic Time Warping and Euclidean Distance measures. We spatially match these two sets of clusters, creating sets of similar simulated and observed subbasins. In paired basins with observed data, we compute a scalar flow duration curve; the ratio of the observed and simulated flow duration curve at each exceedance probability. In paired basins without observed data, we apply these scalar values by mapping simulated flow exceedance probabilities to scaling factors. We refine the corrections by fitting the distribution of corrected flows to the Gumbel Type 1 distribution to estimate extreme flows. We tested this method in river basins around the world using the GEOGloWS ECMWF Streamflow Model, a global hydrologic model with an ERA5 driven 40 year streamflow simulation. Because this method involves post-processing data rather than modification of the model, it enables community level stakeholders to correct simulations in gauged and ungauged basins and thereby interact with any continental- to global-scale model with confidence. We provide python tools that implement this approach.