Hydrology has a dangerous sample size problem, and our Intelligent Systems are ignoring it. We are training AI to predict catastrophic 10,000-year events using datasets that are statistically anemic often spanning just 40 years. This isn't just a data gap; it’s a recipe for massive epistemic uncertainty.
Most models still assume stationarity, but with urbanization spiking runoff potential and climate change shifting rainfall intensities, the baseline is a moving target.
If we don't force AI to respect the physical limits of catchment-scale mass balance, we’re just building high-tech masks for low-probability disasters. No amount of data optimization can fix a model that hallucinates patterns in short-term noise.
I think it is really important to train AI to respect the hard physics of water balance instead of just waiting the next extreme event to shatter our predictions
