Research
I research the roles that snow and glacier ice play in modifying the timing and magnitude of water availability for supporting critical downstream uses such as municipal water supply, salmon habitat, agriculture, and hydropower, particularly in the context of climate change and extreme events like heatwaves. As regions like Western North America stand to lose a majority of their glacier ice over the coming decades, my work underscores a fundamental truth about life in Western Canada: ice matters.
Community water resource vulnerability to deglaciation
Using a case study of the province of Alberta, Canada, we identify vulnerable water resources by detailing signatures of glacier runoff in August streamflow, and we determine the relative importance of glacier runoff at the local scale.
We take three key steps:
Identify signatures of glacier runoff in historical August streamflow data
Develop a dataset of municipal water source locations for >500 communities in Alberta
Determine the water sources that overlap with strong glacier runoff signals
We find four locations that are both important for water supply and projected to change substantially under a loss of glacier ice: the hamlet of Lake Louise, the town of Hinton, the Bighorn Dam, and the town of Rocky Mountain House. Critically, the Bighorn Dam is the largest reservoir in Alberta, and over 1 million people source their water supply downstream.
Reference:
Sam Anderson and Valentina Radic. "Identification of local water resource vulnerability to rapid deglaciation in Alberta." Nature Climate Change. (2020)
Deep machine learning for regional streamflow prediction
Using a case study in the hydrologically diverse region of southwestern British Columbia and Alberta, Canada, we train, fine-tune, and evaluate convolutional long short-term memory (CNN-LSTM) neural network models across six streamflow regimes.
We find that the model works well overall, with the best performance in snow-dominated mainland British Columbia and the worst performance in the Prairie region of Alberta. We also find that fine-tuning models on sub-regional streamflow regimes improves model performance, especially in rain-dominated regimes.
Reference:
Sam Anderson and Valentina Radic. "Evaluation and interpretation of convolutional long short-term memory networks for regional hydrological modelling". Hydrology and Earth System Sciences. (2022)
Interpreting deep machine learning hydrological models
Deep machine learning models can be used to predict streamflow across a range of flow regimes, but it isn't well understood what the models are actually learning. We design a set of experiments to investigate:
Where is the model focusing in space when predicting streamflow?
Which variables are most important for model decision making?
What weather drives modelled runoff generation?
How are glaciers represented within the model?
Our results reveal that the decision-making process of the deep-learning model is interpretable and consistent with the known drivers of streamflow.
Reference:
Anderson, Sam and Valentina Radic. “Interpreting deep machine learning for streamflow modelling across glacial, nival, and pluvial regimes in southwestern Canada” . Frontiers in Water. (2022)
