We present a knowledge-guided machine learning framework for operational hydrologic forecasting at the catchment scale. Our approach, a Factorized Hierarchical Neural Network (FHNN), has two main components: inverse and forward models. The inverse model uses observed precipitation, temperature, and streamflow data to generate a representation of the current underlying catchment state. The forward model predicts streamflow using the learned catchment state. The FHNN architecture is designed to model multi-scale processes and capture their interactions while providing explainability and interpretability. FHNN also improves forecasts based on real-time data through an inference-based data integration approach. FHNN’s data integration approach improves forecasts in response to observed data more efficiently than data assimilation methods (e.g., ensemble Kalman filtering) that require computationally intensive optimization. Once an inverse model is trained, it can quickly infer catchment states directly based on data in real-time. To show the operational performance of FHNN, we compare the FHNN forecasts with that of an expert human hydrologic forecaster using a physics-based model where both use the same imperfectly known future precipitation forecast in their modeling. The expert human forecaster creates a more accurate forecast within the first 18 hours of a forecast’s issuance, but FHNN has significantly better predictions at longer lead times. Additionally, FHNN internal states correlate strongly with internal physics-based model states, such as soil moisture, in a synthetic case. This research lays the groundwork for leveraging the predictive performance of AI-based models with the expertise in forecasting agencies to produce better river forecasts at all lead times.

The catchment approach has been traditionally limited to small, experimental catchments where water fluxes can be determined with high accuracy. However, larger catchments where landscape management occurs have emergent drivers of streamflow at scale, and thus may exhibit novel responses to land cover disturbance. We used statistical models of water yield and annual maximum peak streamflow for multiple forested catchments in the low-relief glaciated region of central North America to investigate how forest disturbance may affect water yield and peak flows in similar landscapes. We utilized linear models, linear mixed effects models, and probabilistic flood-frequency analysis, with Bayesian parameter estimation in two case studies in Minnesota, USA: 1) a wildfire comprising ~30% of a 650km 2 wilderness Upper Kawishiwi catchment, and 2) 11 catchments within the St. Louis River Basin ranging from 56 to 8,880 km 2 with a patchwork disturbance regime wherein ~0.25% to 1% of the catchment is harvested or converted to non-forest land use each year. We also assessed for the most likely hydrological recovery year after forest disturbance, and the relative importance of stationary and nonstationary drivers of streamflow. We found forest disturbance correlated with declines in water yield for low-level disturbance regimes, but that water yield increased in response to the large-scale wildfire. Positive and negative associations of forest disturbance with peak flows were observed, generally with low confidence. Hydrologic recovery time ranged from 5 to 12 years for water yield and peak flows following disturbance. Despite these effects of forest disturbance on streamflow, effects of climate variability and stationary catchment size factors were more prominent drivers of streamflow. Basins larger than ~50 km 2 in low-relief glaciated regions were resilient to forest cover change when it comprised <30% of basin area, but climate change may have a larger effect than could be mitigated by land management.