Mitigating Prediction Error of Deep Learning Streamflow Models in Large Data?Sparse Regions With Ensemble Modeling and Soft Data

Predicting discharge in contiguously data-scarce or ungauged regions is needed for quantifying the global hydrologic cycle. We show that prediction (PUR) has major, underrecognized uncertainty and drastically more difficult than previous problems where basins can be represented by neighboring similar (known as basins). While deep neural networks demonstrated stellar performance streamflow predictions, nonetheless declined PUR, benchmarked here with a new stringent region-based holdout test on US data set 671 basins. tested approaches to reduce such errors, leveraging network's flexibility integrate “soft” data, satellite-based soil moisture product, daily flow distributions which improved low simulations. A novel input-selection ensemble average greatly reduced catastrophic failures. Despite challenges, showed stronger metrics PUR traditional models. They appear competitive geoscientific modeling even settings.