Flood Early Warning Systems (FEWS) are among the most effective non-structural measures for reducing flood risk, particularly in data-scarce regions with rapid hydrological responses. However, designing efficient FEWS requires balancing forecasting skill with the economic costs of dense monitoring networks. Identifying the most influential observation points is therefore essential for reliable flood forecasting with minimal instrumentation.In this study, we propose a data-driven framework to identify critical sub-basins whose monitoring provides the greatest benefit for flood early warning. The framework integrates long-term stochastic rainfall simulation, semi-distributed hydrological modeling, machine learning, and feature importance analysis. High-resolution synthetic rainfall time series are generated using a multifractal-based stochastic approach and used to drive a hydrological model, resulting in an extensive virtual database of flood events across multiple sub-basins. Simulated sub-basin discharges are then used as predictors in a Random Forest model to forecast outlet discharge at different lead times.Feature Importance Measures (FIM) quantify the relative contribution of each sub-basin to flood forecasting performance, enabling identification of a reduced set of hydrologically dominant sub-basins. The methodology is demonstrated in the semi-arid, mountainous Torghabeh River Basin (northeastern Iran), where limited hydrometric infrastructure and short response times pose significant challenges for flood monitoring. Results show that only a subset of sub-basins exerts dominant control on outlet flood response, while many others contribute marginally. The identified influential sub-basins vary with the forecasting lead time, highlighting the importance of tailoring FEWS design to operational objectives.Overall, the proposed framework offers a flexible approach for optimizing FEWS design, supporting evidence-based decisions on sensor placement and providing new insights into the internal organization of flood-generating processes.
Flood early warning systems (FEWSs) in data-scarce and rapidly responding catchments require monitoring networks that provide sufficient forecasting information with minimal operational cost. This study evaluates the applicability of a data-driven sensor-placement framework to the semi-arid Torghabeh River Basin in northeastern Iran, where flood-producing tributaries affect the city of Torghabeh and downstream developed areas. A 2000-year stochastic rainfall series was generated to represent a wide range of hydrological conditions and was used to drive a basin-scale HEC-HMS model parameterized from catchment physiographic characteristics and available hydrological information. After applying the adopted flood-event extraction and filtering criteria, 466 relevant synthetic events were retained for the sensor-placement analysis. Sub-basin hydrographs were then used as predictors in a Random Forest forecasting model, while permutation feature importance, SHAP, and derivative-based importance were combined to identify the most informative monitoring locations. For a 1 h lead time, the progressive predictor-selection analysis identified five of 24 candidate sub-basins as a parsimonious configuration associated with the performance plateau, corresponding to an approximately 80% reduction in monitoring locations. A different but similarly parsimonious configuration emerged for a 2 h lead time, demonstrating that optimal sensor placement is lead-time-dependent. These results demonstrate the potential of explainable machine learning tools to support parsimonious FEWS monitoring-network design under the evaluated synthetic modeling conditions, while operational application requires further validation using independent observed hydrological data.