Reconstructing catchment precipitation with deep learning and a large network of streamflow gauges

Abstract. Global precipitation estimates are uncertain, particularly in sparsely gauged forested areas and mountainous terrain. Improving precipitation estimates in such regions is important for studies of the global water cycle. Previous studies in individual catchments have demonstrated the potential of inferring precipitation from streamflow. Using the Multi-Attribute Catchment Hydrometeorological (MACH) dataset, we shared the parameters of an entity-aware long short-term memory network (EA-LSTM) across catchments and used local weather and streamflow records to reconstruct daily areal precipitation for 906 catchments in the conterminous United States. Each reconstruction used observations through seven days after the target date, with no daily precipitation input. The reference precipitation and weather variables came from Daymet. With runoff included, adding relative humidity, vapour pressure deficit and downward shortwave radiation to temperature improved daily Kling–Gupta efficiency (KGE) in 895 catchments and raised its median from 0.626 to 0.771. Median KGE reached 0.834 and 0.845 for three- and five-day totals. The median fraction of reference events with a correctly identified peak date was 73.3 %. Median amount bias on reference heavy days was −32.0 %, indicating that the magnitude of extreme precipitation remains more difficult to reconstruct. Grouped by the relative size of the two surface-flow components, median exact peak-date success was 73.8 % in infiltration-excess catchments and 67.5 % in saturation-excess catchments. The median paired gain from adding runoff was 1.9 and 4.5 percentage points, respectively. These results show that learning across catchments can recover multi-day precipitation totals from streamflow and weather records. Future research could improve transfer across catchments and the representation of extreme precipitation amounts.