Data-driven control reveals distributed flood adaptation priorities across large river networks under climate change

Distributed flood adaptation requires knowing where in a river network attenuation effort should concentrate and how much each reach requires, but the spatial coupling, scenario dependence, and high dimensionality of real drainage networks have kept these requirements largely unresolved. We combine data-driven dynamics learning, reduced-order modeling, and optimal control theory into a diagnostic framework that infers reach-level attenuation targets directly from process-based hydrologic simulations without iterative simulation and optimization. Proper Orthogonal Decomposition compresses the network-wide discharge field into a low-rank basis, Dynamic Mode Decomposition with control identifies a linear surrogate of precipitation-driven flood dynamics, and a Linear Quadratic Regulator solves for the spatially distributed attenuation in closed form. Applied to a large river basin under a multi-model, multi-scenario climate ensemble, the effort–residual trade-off follows a common diminishing-return structure across emission pathways, but higher-emission scenarios retain substantially greater residual flood volume at comparable effort levels. The bulk of the allocation tracks mean-flow scaling, yet the framework identifies priority reaches at tributary junctions that neither drainage area nor mean discharge can flag; these reaches retain the highest residual-to-baseline exceedance ratio after optimal control, revealing structurally stubborn bottlenecks where flooding is hardest to attenuate. Inter-scenario separation in residual risk widens progressively downstream, and ensemble agreement on effectiveness degradation distinguishes reaches where investments can proceed with confidence from those requiring flexible, adaptive strategies.