Emergency managers require real‑time knowledge of flood extent, depth, and short‑term evolution across an entire basin. In practice, only a handful of stream gauges provide observations during a flood, and rerunning high‑resolution hydrodynamic models is too costly for rapid updates or large ensembles. C‑STRIDE is an observation‑driven AI digital twin that converts short records from a few gauges, together with terrain and rainfall inputs, into basin‑wide water‑depth maps and extends predictions up to one day ahead. The model is trained on simulations from a calibrated two‑dimensional hydrodynamic model and does not need a separate data‑assimilation step.
A case study in the Des Plaines River basin near Chicago used six gauges to inform predictions over 4.2 million 30‑m grid cells. Terrain contributed the most to accuracy gains, while rainfall prevented error growth over longer horizons; together they reduced errors by roughly 40% compared with gauge data alone. When future rainfall is known, one‑day‑ahead errors stay near 15%, versus nearly 40% without rainfall. Using real gauge records instead of simulated ones, the model shifts its outputs toward observed hydrographs at three of the six stations without retraining, and runs about 150 times faster than the traditional hydrodynamic model.
These findings demonstrate that sparse gauges, terrain, and rainfall can be fused into fast, continuously updated flood forecasts, moving toward operational flood digital twins, though further testing with real‑time data and rainfall forecasts is required.
Review