Researchers built an AI model that predicts flood depth across an entire river basin using data from just a handful of stream gauges.
The system, called C-STRIDE, is trained on simulations from a calibrated two-dimensional hydrodynamic model of the Des Plaines River basin near Chicago. Six stream gauges feed it real-time readings, which it combines with terrain and rainfall data to generate flood depth maps across 4.2 million 30-meter grid cells, projecting up to 24 hours ahead. Terrain data improves accuracy the most, while rainfall data keeps prediction errors from compounding over longer forecast windows. Together, those two inputs cut errors by about 40% compared with using gauge records alone.
Traditional hydrodynamic models are too slow and expensive to rerun every time new gauge data comes in, which is exactly when emergency managers need fresh numbers most. C-STRIDE runs about 150 times faster than the physics-based model it was trained on, and when fed real gauge data instead of simulated data, it adjusted its predictions at three of six gauges without needing to be retrained.
It is still a research prototype trained on one basin's simulations, not a tool tested against live rainfall forecasts, so treat the speed and accuracy numbers as a promising first lap rather than a finished product.