Machine learning is reshaping what's possible in hydrology — and this paper, authored in part by our CTO, Alden Keefe Sampson, makes the case that the field needs to take that transformation seriously.
Published in Water Resources Research and adapted from a keynote at Google's 2020 Flood Forecasting Meets Machine Learning Workshop, the paper is a call to action for the hydrology community. Its central argument: large-scale hydrological datasets contain far more information than traditional models have been able to capture, and deep learning is beginning to unlock it. Recent experiments with LSTM-based rainfall-runoff models show that a single model trained across hundreds of basins can outperform traditional models calibrated individually for each one — a result that challenges long-held assumptions about "uniqueness of place" as a barrier to generalized hydrological theory.
The authors argue that the hydrology community should stop treating process understanding and machine learning as competing philosophies, and instead focus on rigorously understanding where each approach adds value. The future of accurate, scalable water prediction likely requires both.
