The spatial distribution of snow and its melt over large river basins cannot be directly observed, complicating decisions for competing priorities of flood protection and resource use. This work aims to advance our ability to estimate snow distribution through improvements in the satellite observation of snow-covered area (SCA) and the modeling of snow water equivalent (SWE). I analyze the accuracies of three methods to map SCA from MODIS, a binary and a fractional method (MOD10A1) based on the normalized difference snow index (NDSI), and another fractional method based on spectral mixture analyses, MODSCAG. I find that fractional SCA estimates from MODSCAG perform best over a range of snow and vegetation conditions and the model maintains its retrieval ability in the transitional periods during accumulation and melt. Forest obstructs viewable snow cover, and satellite-viewing geometry has a larger effect on the fractional NDSI-based approach than MODSCAG.

Adjustments to SCA to account for forest canopies permit MODSCAG to perform very well in all but the densest canopies while MOD10A1 shows early melt out, when compared with ground sensors, in sparse to moderately dense canopy. Combining satellite observed SCA and albedo with an energy balance model, I estimate spatially distributed snow water equivalent for 12 years in the Sierra Nevada. The model shows that the relationship of SWE with elevation is significantly different for wet, mean and dry years, between drainages and at different latitudes. SWE and SCA become increasingly correlated from March 1st to July 1st, such that real time SCA observations may be sufficient for SWE prediction. I compare spatially integrated volumes of snow water equivalent from this model and two real-time models with full natural flow estimates in 18 Sierra Nevada basins covering 70% of the streamflow range in the last 80 years.

Forecasting errors that are based on statistical relationships between point measurements of snow and streamflow in these basins can reach 25% to 70% in one out of five years, a large error where reservoirs hold only the annual streamflow. The retrospective model best estimates the unimpaired streamflow and provides a way to evaluate the real-time models.