Streamflow forecasting in ungauged basins with CNN-LSTM and radar-based precipitation

Predicting high-resolution streamflow in ungauged basins remains a fundamental challenge in hydrology. This study addresses this challenge by developing a novel dual-stream CNN-LSTM architecture that separately processes dynamic meteorological inputs and static watershed characteristics to capture complex spatiotemporal hydrological processes. The model was evaluated across 35 natural watersheds in South Korea using 1 km resolution radar-based precipitation and watershed data at 10-minute intervals. Our approach achieved a mean Nash-Sutcliffe Efficiency of 0.59 (±0.12), significantly outperforming both non-CNN and lumped baseline models. Flow regime analysis revealed consistent improvements across all flow conditions, though challenges in peak flow prediction remain. Water balance analysis demonstrated improved physical consistency compared to lumped model. Statistical analysis identified hydrological variability as the primary performance-limiting factor, while input sensitivity testing confirmed flow accumulation raster data as the most critical spatial variable. Through controlled experiments, we demonstrated that the model can capture fundamental hydrological processes and physically plausible spatial runoff patterns, even without being given explicit information about the underlying physical phenomena.