Towards the use of quantile mapping and historical patterns in SWE calculations

Water managers would benefit from high-resolution, distributed snow water equivalent (SWE) estimates, but limited observations and high spatial and temporal variability make SWE difficult to estimate in real-time. For example, in-situ snow measurement stations provide current-year SWE data, but under-sample SWE spatial heterogeneity; and SWE Reanalysis products back-calculate distributed daily SWE once snow has melted to provide more accurate SWE estimates than do real-time models, but don’t provide real-time information. Recent studies have shown that many regions and years have annual snow accumulation patterns that are repeatable, so here we seek to leverage these patterns by combining sparse real-time measurements with distributed historical information to map real-time SWE. We develop and test two methods for calculating SWE, and for each method we test three criteria for deciding which station or collection of stations should be used to estimate SWE at each grid cell. We determine which Western U.S. regions have similar standardized SWE anomalies, and then test our methods in the Upper Colorado River Basin (UCRB). For our two methods, we calculate 1 April 1990 – 2021 SWE using parametric and nonparametric distributions with a leave-one-out approach to map the current-year’s position within an in-situ station’s long-term SWE distribution to the corresponding position within historical distributions at nearby SWE Reanalysis grid cells. In each of these methods, we use our three station selection criteria to calculate SWE such that we calculate six SWE products over the UCRB. These criteria are: i) the nearest-neighbor station, ii) the collection of most-correlated stations, and iii) all in-situ stations within the UCRB. We then compare these to the SWE Reanalysis product from each corresponding year and compare the accuracies of our various methods with the accuracy of SnowModel output relative to the SWE Reanaysis product. The most accurate method used the mean SNV from the collection of most-correlated in-situ stations. This produced distributed 1 April SWE with a median R value of 0.80 and a root mean squared error (RMSE) of 0.13 m, compared to SnowModel results with an R of 0.60 and RMSE of 0.18 m. The methods used here could be applied to additional data, such as updated SWE Reanalysis products that might have higher resolution and improved accuracy over the product used here.