The Effect of Geographical Bias in Streamflow Gauge Distribution for Global Flood Forecasting

The "Prediction in Ungauged Basins" (PUB) problem remains a central challenge in global hydrology, as the accuracy of rainfall-runoff models is fundamentally constrained by the availability of local streamflow observations for training and calibration. While recent advancements in data-driven modeling have improved our ability to generalize across catchments, the global distribution of streamflow gauges is characterized by severe geographical and socio-economic biases. Most available data are concentrated in the Global North, leaving vast regions, particularly in the Global South, functionally "ungauged" or under-represented in the training sets of global models.

In this study, we shift the focus from simply counting the fraction of ungauged watersheds to estimating the quantitative effect of this geographical bias on global flood forecasting skill. Using a large-sample machine learning framework based on Google’s flood forecasting model, we quantify the relationship between gauge network density (specifically upstream and downstream coverage fractions) and predictive performance. We utilize cross-validation experiments to isolate the information loss associated with geographical distance and hydrological connectivity from gauged locations.

Our analysis indicates that hydrological factors are the main driver of predictive performance, with basin aridity being a larger factor in model skill than whether a basin is gauged or ungauged. However, if streamflow gauges were hypothetically installed in all of the world’s watersheds, we could expect a 20% increase in the Nash-Sutcliffe Efficiency (NSE) skill score for state-of-the-art global models, including causing almost half of the basins globally currently scoring below NSE = 0.50 to rise above that threshold, with an average skill improvement of about ΔNSE = 0.1. Critically, this potential for improvement is not uniform, with Africa being the continent where our model predicts that largest overall skill improvement with higher density gauging networks. 

These findings emphasize that the path toward equitable global flood safety requires not just better algorithms, but a concerted effort to address the structural biases in the global hydrological data ecosystem.