Accurate reservoir representation in large‐scale river models remains challenging owing to limited access to data on reservoir operations. We contribute to model development by introducing a global machine‐learning based flood storage capacity (FSC) data set and a satellite‐based target storage reservoir operation scheme (SBTS). The FSC data set for 1,178 flood control reservoirs is constructed using multiple reservoir attributes and reported FSC data. Integrating these FSCs into SBTS enables its global applicability with generic formulations of reservoir zoning. Then, we develop SBTS by introducing monthly median values of satellite storage data as target storage parameters. With these seasonal patterns as constrains, improvements in simulation results are achieved. When simulated with observed inflow, SBTS performed significantly better (median Kling‐Gupta efficiency values of 0.52 and 0.17 for outflow and storage simulations among 289 reservoirs), compared to the previous reservoir operation scheme with linearly interpolated target storage parameter (0.41 and −0.19). Compared to two existing global schemes without seasonal target storages, SBTS demonstrates improved performance for many reservoirs whose inflow seasonal pattern is more regular. When coupled with a global river model, it improved discharge simulations across 293 downstream gauges, with overall performance, peak, and low flow improving at 40%, 21%, and 35% of gauges, respectively, compared to simulations without reservoirs. However, reservoir simulations do not improve notably due to the biases in simulated inflow to reservoirs. We demonstrated that machine‐learning FSC and satellite observations help improve reservoir parameterizations, and found that improvements in other aspects of river modeling are essential for accurately reproducing discharge patterns.
