Valentina Radic

Despite the social, ecological, and cultural importance of glaciers and glacier-fed rivers, a quantification of key glacier controls of streamflow remain elusive and outstanding questions persist. For example: which communities’ water supplies are most vulnerable to the loss of glacier ice? By how much do glaciers modify the streamflow response to heatwaves? First, I use principal component analysis, self-organizing maps, and multivariate linear regression to provide an assessment of community vulnerability to deglaciation in Alberta, Canada, by identifying and predicting signals of glacier runoff in historical streamflow datasets. I combine these models with a new dataset of community water supply sources to find that the most vulnerable locations are the communities of Hinton, Lake Louise, and Rocky Mountain House, as well as the Bighorn Dam, which forms the largest reservoir in the province and provides water for over a million people downstream. Next, I develop an accurate and interpretable convolutional long short-term memory neural network regional hydrological model for streamflow prediction across Alberta and British Columbia, Canada. This deep machine learning model is forced by gridded ERA5 temperature and precipitation data and predicts streamflow at 226 stream gauge stations.Finally, I use this model to systematically investigate the streamflow response to heatwaves. I determine how this streamflow response varies by basin glacier coverage, as well as by heatwave timing, duration, and intensity, under both normal and warmer climate scenarios. I quantify how increasing glacier coverage is associated with both increasing streamflow generation during summer heatwaves, as well as increasing compensation in summer to the loss of snow during spring heatwaves. My results advance understanding on multiple research fronts in glaciology and hydrology: I demonstrate the importance of local-scale water resource data for glacier runoff projections; I emphasize the interpretability of deep machine learning models as a means to apply machine learning to new frontiers in hydrology; and I offer new frameworks and metrics to understand and characterize the hydrological impacts of heatwaves. My findings motivate future inter- and trans-disciplinary research to develop better deep learning hydrological models, and make progress towards answering politically and socially relevant glacio-hydrological research questions.

View record

The exchange of energy between a glacier surface and its surroundings, known as its surface energy balance (SEB), is a primary control on surface ablation rates. In the modelling of glacier SEB, parameterisation rather than direct measurement is frequently used to estimate one or more of the contributing heat fluxes, with smaller fluxes often deemed negligible. The turbulent fluxes of sensible and latent heat are commonly parameterised using forms of the bulk aerodynamic method. These techniques were developed for flat, uniform surfaces, and substantial uncertainty remains in the validity of their application over sloped, inhomogeneous glacier terrain. A multi-year field campaign was performed on two glaciers in the Purcell Mountains of British Columbia, Canada, where season-long observations of the complete SEB were obtained at multiple locations. The obtained dataset was used to drive an ablation model which showed good agreement with observed rates at seasonal, daily, and sub-daily timescales, effectively closing the energy balance. Through eddy covariance measurements, the turbulent heat fluxes were observed to be important components of SEB at each location, providing 31% of seasonal melt energy, and up to 78% of melt energy on individual days, underlining the need for their accurate estimation. The rain heat flux, often assumed negligible, was a significant contributor to melt energy on daily and sub-daily timescales during heavy rainfall (up to 20% day⁻¹). An evaluation of common turbulent flux parameterisation methods found their performance to be highly sensitive to the choice of roughness length scheme and atmospheric stability function. Observed roughness length values differed from those commonly assumed for glacier surfaces, and varied substantially between locations, highlighting the need for site-specific values. Two techniques were developed for the remote estimation of roughness using digital elevation models, and performed well when compared with in situ observations. The occurrence of shallow, katabatic surface layers with low-level wind maximums was frequently observed over the sloped, glacial test sites. Existing stability parameters and functions used in turbulent flux parameterisation were found to be unreliable in these conditions, as was the commonly employed assumption of constant turbulent flux and friction velocity with height through this layer.

View record