From Lumped to Spatially Distributed Hydrologic Modeling: A Data‐Driven Framework Evaluated Across North American Catchments

Abstract

Data-driven rainfall–runoff models have advanced rapidly, yet the majority of large-scale applications still rely on lumped inputs that smooth out spatial variability in precipitation, temperature, and landscape properties. This simplification can introduce substantial biases in flood peaks, hydrograph timing, and water balance estimates. Here, we develop a deep learning framework that ingests spatially distributed (gridded) meteorological forcings and is evaluated across 921 catchments in North America using the CAMELS-SPAT data set. We compare two strategies for fusing spatial and temporal information, namely a spatial-to-temporal (S2T) architecture and a temporal-to-spatial (T2S) architecture, and assess the added value of gridded catchment attributes and finer spatial resolutions. The T2S approach consistently achieves the best performance, increasing median Nash–Sutcliffe Efficiency (NSE) from 0.60 (lumped baseline) to 0.65 and Kling–Gupta Efficiency (KGE) from 0.63 to 0.69, with NSE improvements observed in 75.4% of catchments. Performance gains are consistent across catchment sizes and aridity regimes, with enhanced skill in low-flow, high-flow, and peak timing representations. Integrated Gradients (IG) analysis shows that the model learns physically coherent spatial behavior, highlighting grid cells near the outlet and identifying dominant meteorological drivers of peak runoff events. These results demonstrate the practical value of spatially distributed forcings for data-driven hydrologic modeling and outline a pathway toward interpretable, large-scale streamflow prediction.