Could you briefly describe your position within the CanSISE Network?
I am a PhD student at the University of British Columbia working under William Hsieh, a funded CanSISE investigator.
What is your primary research goal?
The primary goal of my research is to develop high-resolution maps of snow across British Columbia (BC). The value I'm interested in is snow water equivalent (SWE), which tells you how much water is in the snowpack. Knowing the amount of snow water on a regional scale (100s of km) helps hydrologists and water resource managers to predict how much water will be released during the melt season as river runoff. This is especially important in regions like BC, which depends on this water to make it through the drier summer months.
Can you tell us a bit more about regional snow estimation?
There are different types of models or "gridded products" that attempt to represent snow, among other physical properties, on large scales. These include land data assimilation systems, reanalyses, observation-based products and others that use blended techniques. While such products work well in flat, open areas that don't see much snow, the conditions in many parts of BC are quite different -- mountainous, heavily forested, and receiving many metres of snow. These conditions make it difficult to make good estimates of snow across the entire region.
Please explain how you investigate regional snow distribution?
I use machine learning methods such as artificial neural networks (ANNs) to attempt to estimate the snowpack. The neural networks build statistical representations of the regional snowpack using many inputs. Manual snow survey stations provide on the ground measurements of the actual snowpack, whereas the gridded products show estimates on a larger scale. Additional information like elevation, latitude, longitude and time of year (known as "predictors") are also fed into the model, which can then estimate snow at other locations and times. To check how well the model is doing, I use "cross-validation", a technique in which I leave out a fraction of survey stations spread throughout the province to use later as test stations and build the model using only data from the non-test stations. The model performance is then measured at the test stations. The process is repeated so all survey stations are utilized as test stations.
Tell us about something you've discovered.
The ANN model that uses only off-the-shelf gridded products plus the additional information as inputs (ANNG in Figure 1) has a lower mean station root-mean-square error (RMSE) and therefore does a better job of estimating the snowpack than the individual products themselves. The ANN using the gridded products even does better than the high-resolution Variable Infiltration Capacity (VIC) hydrologic model that has been run by the Pacific Climate Impact Consortium for 4 major BC watersheds (see https://pacificclimate.org/data/gridded-hydrologic-model-output for details on the VIC runs).
Figure 1: Mean Station RMSE for 3 evaluated gridded products plus VIC. Shown from left to right are the 3 best performing gridded SWE products (GLDAS2, ERALand and MERRA); mean, multiple linear regression (MLR) and ANN of those gridded products; SnowMelt model; MLR and ANN using gridded products and the SnowMelt model; and VIC. Error bars represent 95% confidence intervals.
I also setup a snow model (EcoHydRology SnowMelt) driven by a high-resolution temperature and precipitation data set (bias-corrected ANUSPLIN). The snow output from this model used as an extra predictor in the ANN model further enhances its performance (ANNGS), such that the new ANN performs significantly better than VIC.
What implications does this have for future efforts to estimate the amount of snow?
Machine learning methods are a tool to bring together many sources of information and most likely better estimate snow water equivalent than individual methods or products. In areas where it is very difficult to estimate snow, such as BC, such methods could become an important part of water planning and decision making.
How does your research fit into the broader scope of the CanSISE Network?
My project is part of research area A, focused on improving seasonal to multidecadal snow and sea ice prediction. This model will be used to downscale and bias correct a prediction system (CanSIPS) that will lead to improved snow water forecasts.
What are your next steps?
Additional data sources will be incorporated into the model to try to gain further improvements. Satellite microwave data, used to estimate snow water equivalent with some success in less complex environments, along with visible wavelength data that show forests and snow lines will be tested.