Statistics, Lasers, and Pillows: Leveraging historical patterns and observations to predict distributed 1 April SWE

Abstract

This dissertation investigates statistical and satellite-based approaches for estimating snow water equivalent (SWE) and its spatial distribution in mountain regions to inform water management. In the western United States, near-real-time SWE estimates can have low domain-wide bias but tend to misrepresent the spatial distribution of SWE, which controls snowmelt runoff magnitude and timing. A retrospective SWE reanalysis product (UCLA Western U.S. SWE Reanalysis) more accurately represents the spatial distribution of SWE, but is not available in real-time. This motivates the question: can we leverage the more accurate historical spatial information contained in SWE reanalysis products to improve near-real-time SWE estimation? We focus on 1 April SWE, a benchmark for peak seasonal accumulation in northern hemisphere mid-latitude basins. Throughout this dissertation, we combine historical SWE patterns with various current-year observations to develop and test methods for estimating near-real-time distributed mountain SWE. We test these methods in the western U.S. to evaluate their potential for use in data-sparse regions globally. In Chapter 2, we combined real-time in-situ snow station observations with a SWE reanalysis product in the Upper Colorado River Basin using standard anomaly (SA) and quantile-based methods. Within a leave-one-out framework, these statistical approaches produced distributed SWE estimates with higher spatial correlation and lower bias relative to the reanalysis than did physically-based modeling systems. However, the residual errors revealed elevational and regional gradients not captured by the SA framework, particularly in anomalous years. In Chapter 3, recognizing that long-term in-situ station networks are not available everywhere, we evaluated ICESat-2 altimetry as an additional, potentially global snow depth data source. We custom processed ICESat-2 elevations and accessed bare earth elevations from the USGS 3D Elevation Program using the SlideRule Earth tool. To calculate snow depth, we subtracted the bare earth elevations from the ICESat-2 measurements. We excluded measurements over steep slopes and aggregated the remaining ICESat-2 snow depth estimates by date across the California Sierra Nevada (the domain). Daily domain-median ICESat-2 snow depth and the running 10-day median of ICESat-2 snow depth showed high correlations and minimal biases relative to the domain-wide reanalysis snow depth. This large-scale aggregation of ICESat-2 snow depths yielded accurate snow quantity estimates that could extend snow observations globally where high-quality snow-off terrain models exist. In Chapter 4, we extended the SA framework that was used in Chapter 2 by additionally applying an empirical orthogonal function (EOF) analysis to the 1 April SA values to identify dominant modes of spatial and temporal variability in Sierra Nevada SWE. We used data available in near-real-time from the ERA5 reanalysis product and ICESat-2 snow depth estimates to develop linear regressions to the time-varying principle components (PCs) that resulted from the EOF analysis, and combined these with the time-static spatial EOF patterns to estimate distributed SWE within a leave-one-out framework. We compared SWE estimates from the SA method, EOF method, and an operational model that assimilates in-situ observations relative to the SWE reanalysis product. In most years, statistical approaches performed comparably to or better than the operational product. However, in an extreme low snow year the operational product performed best, indicating that additional observations are needed to constrain estimates in low snow years. In places without in-situ SWE observations, the EOF method could be used with meteorological information from ERA5 to estimate SWE distributions. In places without dense meteorological observations for assimilation in the ERA5 reanalysis, additional observations from ICESat-2 or other satellite products will likely become more important. Collectively, this dissertation demonstrates that accurate historical SWE distributions can be leveraged to produce computationally efficient, near-real-time estimates of distributed mountain SWE. The ERA5 reanalysis product is available globally, so at minimum the techniques used here require a reliable SWE reanalysis product. The inclusion of additional observations, for example from ICESat-2, could provide added value in low-snow years and in regions with sparse in-situ observations for assimilation in the ERA5 reanalysis.