Enhancing runoff forecasting through the integration of satellite precipitation data and hydrological knowledge into machine learning models

In this study, we use feature engineering (FE) strategies to enhance the performance of machine learning (ML) models in forecasting runoff and peak runoff. We selected a 300-km2 tropical Andean catchment, representative of rapid response systems where hourly runoff forecasting is particularly challenging. The selected FE strategies aim to integrate ground-based and satellite precipitation (PERSIANN-CCS) and to incorporate hydrological knowledge into the Random Forest (RF) model. Although the evaluation of the satellite product (microcatchment-wide and hourly scales) was initially discouraging (correlation of R = 0.21), our approach proved to be effective. We achieved Nash-Sutcliffe efficiencies (NSE) ranging from 0.95 to 0.61 for varying lead times from 1 to 12 h. Moreover, the inclusion of satellite data improved efficiencies at all lead times, with gains of up to 0.15 in NSE compared to RF models using ground-based precipitation alone. In addition, an extreme event analysis demonstrated the utility of the developed models in capturing peak runoff 98% of the time, despite a systematic underestimation as lead time increased. We highlight the ability of the RF models to forecast lead times up to three times the concentration time of the catchment. This has direct implications for enhancing flood risk management in complex hydrological settings where conventional data acquisition methods are insufficient. This study also underscores the value of testing hydrological hypotheses and leveraging computational advances through ML models in operational hydrology.