Particle Filter data assimilation of streamflow in basins with seasonal snow for initializing short- to medium-range streamflow forecasts

Abstract

Although short- to medium-range streamflow forecasting is vital for many water management decisions the quality of streamflow forecasts in the United States has not improved over time. A possible explanation for this is that the current forecast systems, which rely on manual intervention on the part of forecasters, do not easily allow for testing of potential system upgrades and new methodologies. To transition from a semi-manual procedure to a fully automated procedure—and thereby allow hindcasting experiments to test new methods—forecast system must use an automated data assimilation (DA) framework. In this dissertation, I evaluate the capabilities of one DA method—the particle filter (PF) —for the assimilation of streamflow observations in basins with seasonal snow cover. Very few studies have explored the use of DA based solely on streamflow to update snow states. Studying DA based solely on streamflow in basins with seasonal snow cover is important because streamflow observations are much more widely available than observations of snow cover extent, depth or water content. I first use a synthetic experiment to examine the impacts of such DA on streamflow, snow states, and soil moisture states. PF-DA almost always improves simulated soil moisture and streamflow in two Pacific Northwest basins with seasonal snow, but it degrades the quality of snow water equivalent estimates during the mid-winter and in the basin for which snow is less of a control on runoff. Next, I evaluate to what extent the improved initial hydrologic conditions lead to improvements in the 1- to 7-day lead-time forecasts. I find that PF of streamflow observations in basins with seasonal snow improves forecast performance in terms of accuracy during the spring and summer and in terms of reliability during spring and fall. Finally, I propose an Analog Resampling (AR) method, for use in the PF, that shows potential for expanding the spread of particles when the particle sample degenerates (i.e., when most weight is assigned to few particles). A set of exploratory analyses using AR suggests areas for future development of this method.