Global streamflow modelling using process-informed machine learning

Abstract

We present a novel hybrid framework that incorporates information from the process-based global hydrological model (GHM) PCR-GLOBWB, to reduce prediction errors in streamflow simulations. In addition to catchment attributes and meteorological data, our methodology employs simulated streamflow and state variables from PCR-GLOBWB as predictors of observed river discharge. These outputs are used in a random forest, trained on a global database of streamflow measurements, to improve estimates of simulated river discharge across the globe. PCR-GLOBWB was run for the years 1979–2019 at 30 arcmin and its inputs and outputs were upscaled from daily to monthly time steps. A single random forest model was trained with these state variables, meteorological data and catchment attributes, as predictors of observed streamflow from 2,286 stations worldwide. Model performance was evaluated using Kling–Gupta efficiency (KGE). Results based on cross-validation show that the model is capable of discerning between a variety of hydroclimatic conditions and river flow dynamics, improving KGE of PCR-GLOBWB simulations at more than 80% of testing locations and increasing median KGE from −0.02 in uncalibrated runs to 0.52 after post-processing. Performance boosts are usually independent of the availability of streamflow data, making our method a potential candidate in addressing prediction in poorly gauged and ungauged basins.

Enhancing Streamflow Prediction Physically Consistently Using Process-Based Modeling and Domain Knowledge: A Review

“…In the realm of utilizing PBM outputs, it is important to acknowledge the widespread use of both calibrated and uncalibrated model outputs in various studies. Intriguingly, while specific research [118] underscores the superiority of calibrated model outputs, a body of evidence [94,140], suggests that both calibrated and uncalibrated PBM outputs can contribute to the enhancement of SFP. This situation raises a pertinent question: Should our priority lie in using calibrated models, or should we opt for uncalibrated ones?…”