Data-driven inference of snowmelt from streamflow in 110 catchments across three continents

Abstract. Direct observations of snow-related processes, and especially snowmelt rates, are difficult to acquire. In contrast, streamflow is continuously measured in many snow-influenced catchments around the world, where it contains valuable information on the timing and rate of catchment-wide snowmelt. Importantly, streamflow observations are often available in regions and periods that lack the observations commonly used for snow reanalyses, such as in-situ snow measurements and space-based observations. Here, we investigate how much information streamflow contains on snowmelt, and whether this information can be used to reconstruct snowmelt in catchments with streamflow observations but without more direct snow observations. To this end, we train bidirectional long short-term memory (BiLSTM) networks on catchment-scale snowmelt estimates derived from state-of-the-art snow water equivalent (SWE) reanalysis products across 163 snow-influenced catchments in the Western United States, Switzerland, and Chile, and evaluate them on a subset of 110 snow-dominated catchments. We compare three input configurations: a streamflow-only model, a meteorology-only model driven by ERA5-Land temperature and precipitation, and a combined model using both streamflow and meteorological forcing. We compare the results against both a climatology baseline and a benchmark derived directly from ERA5-Land SWE estimates. We find that the streamflow-only model reconstructs snowmelt substantially better than both the climatological baseline and the ERA5-Land benchmark, demonstrating that streamflow contains considerable and extractable information on catchment-scale snowmelt. This result holds especially in catchments used during training, but also in out-of-sample catchments not seen during training. However, the meteorology-only model performs considerably better, showing the strong capacity of the BiLSTM to correct the biases in the ERA5-Land forcing. Adding streamflow alongside the meteorological inputs results in only modest performance improvements, despite the substantial snowmelt information contained in streamflow. Our results demonstrate both the potential of BiLSTMs to transfer snowmelt information from existing reanalyses to new years and catchments, and the significant but limited added value of streamflow beyond meteorological forcing in this task. We recommend future work to explore the potential of streamflow to constrain catchment-wide melt rates within existing SWE reanalysis frameworks, in order to improve both the reanalyses themselves and their value as training data for future reanalysis extensions.