Evaluating and Improving Snow in the National Water Model, Using Observations from the New York State Mesonet

Abstract This study leverages observations from the New York State Mesonet to evaluate and improve the representation of snow within the National Water Model (NWM) and its associated land surface model (Noah-MP). To do so, we run and analyze distributed NWM simulations, forced by gridded meteorological analyses, and Noah-MP point simulations, forced by New York State Mesonet (NYSM) observations. Distributed NWM runs, with a baseline configuration, show substantial SWE biases caused by biases in the meteorological forcing used, imperfect representation of snow processes, and mismatches between the land cover in the model and at NYSM station locations. Noah-MP point simulations, using the baseline configuration, reveal a systematic positive bias in SWE accumulation. Sensitivity experiments show that this bias can be mitigated by using an alternative precipitation phase partitioning method. Noah-MP point simulations, with improved precipitation phase partitioning, reveal a systematic negative bias in SWE ablation rates. Sensitivity experiments highlight uncertain parameters within Noah-MP that strongly affect ablation rates and show particularly large sensitivity to the snow albedo decay time-scale (TAU0) parameter, which modulates snow albedo decay rates. Reducing TAU0 improves ablation rates by reducing biases in surface albedo. Distributed NWM experiments, with precipitation phase partitioning and TAU0 adjusted based on Noah-MP point simulation results, show qualitatively similar sensitivities. However, the distributed experiments do not show clear improvements when compared to SWE and streamflow observations. This is likely due to some combination of the abovementioned sources of bias in the baseline-distributed run and (for streamflow) biases in other parameterized processes unrelated to snow in the NWM. Significance Statement Simulating snow on the ground is important for forecasting streamflow. This study compares snow simulations from a streamflow forecast model to observations from an advanced weather station network. Errors in simulated snow are caused by errors in the weather inputs to the model and errors in how the model represents complex physical processes. Some of these errors can be reduced by changing how the model classifies rain versus snow and how the model determines how much sunlight snow reflects. However, using these results to improve the forecast model is tricky, due to errors in the weather data inputs, errors in the representation of other complex processes, and differences between vegetation assumed in the model and that at the weather station locations.