An Introduction to a CanSISE graduate student and their work at UBC

Hi, my name is Drew Snauffer and I'm a PhD student at the University of British Columbia with the CanSISE network. My advisor William Hsieh tackles tough environmental problems using machine learning methods, from the field of artificial intelligence (AI). My work involves using these methods to estimate snow across the province of British Columbia. Good snow estimates are crucial for water resources management and streamflow forecasts, especially in places that get much of their annual water from snowmelt, like BC.  The amount of water contained in a snowpack is called Snow Water Equivalent (SWE), discussed by my colleague Reinel in a previous post.  This can be thought of as the depth of water you'd get if you melted down all the snow in a certain area, and it gives you an indication of how much runoff there will be in streams and rivers when temperatures increase throughout the spring.

Figure 1:  Automatic Snow Pillow at Mission Ridge, BC (photo credit: Viasat Data Systems)
Figure 1: Automatic Snow Pillow at Mission Ridge, BC (photo credit: Viasat Data Systems)

There are many places throughout the province where snow is measured regularly. Two of the ways we get these measurements are Automatic Snow Pillow (ASP) stations (Figure 1) and Manual Snow Survey (MSS) stations (Figure 2). There are about 400 snow survey sites and 70 snow pillows are scattered all over BC, mostly in the southern part (a map is here). As many snow stations as there are throughout the province, it's not enough. Large gaps that exist in our survey network hamper our ability to see how much snow is there, and understand how it changes through time.

Figure 2: Surveyor measuring SWE using a snow tube on a snow course
Figure 2: Surveyor measuring SWE using a snow tube on a snow course

To be able to make good water management plans, we need good estimates of wintertime snow, but we have gaps in our measurement network of Automatic Snow Pillow (ASP) stations and Manual Snow Survey (MSS) stations.  There are a number of ways we try to fill these gaps. Methods that use satellite measurements of microwaves (like those in your microwave oven) that come off the earth have been used for many years. These methods work well in flat, open areas.  But not much of BC looks like this.  The province is known for its rugged, heavily forested mountains, and these factors make satellite microwave data hard to use.  Other methods include: 

a) Reanalyses - long term climate estimates produced using a large number of quality-checked atmospheric and land-based observations 

 

 

b) Land Data Assimilation Systems or LDAS - detailed land-surface models describing things like slope, soil type, etc. used to estimate water and energy movements according to observations

c) Snow models - estimates of snow accumulation and melt at individual points using maps of temperature and precipitation, as well as site information (I'll address these more in the future)

 

These methods are summarized in Table 1 below, along with links to detailed descriptions. They all create a map or grid of estimated snow, so I usually refer to them as "gridded products".  But basically they're large maps of snow estimates.

Table 1 SWE Gridded Product Summary

Product

Type

Resolution

Time range

GlobSnow

observational

25 km

1979-2011

CMC

observational

24 km

1998-2012

GLDAS-1

LDAS

0.25 deg

2000-2012

GLDAS-2

LDAS

0.25 deg

1979-2010

ERAInterim

reanalysis

0.75 deg

1979-2012

ERALand

reanalysis

0.75 deg

1979-2010

MERRA

reanalysis

2/3 x 1/2 deg

1979-2010

MERRALand

reanalysis

2/3 x 1/2 deg

1980-2012

 

Figure 3:  BC's Physiographic Regions

Since it's hard to look at all 470 stations individually, I group them by region. BC has 5 main physiographic regions shown in Figure 3 on the left (and described in more detail in the BC Forest Service's 1994 Forest, Range & Recreation Resource Analysis).  By grouping the stations, I'm looking for regional characteristics or trends that describe how well these methods are doing.  It's also useful to see how they do year to year, so I break them out by month (e.g. all years for January, February, etc.).  


One way to see how well the methods are doing is to look at the average difference between the measurement and the estimate.  This is called the bias.  Bias tells me how much each estimate is over or under the measurement.  If I plot the average bias for several gridded products by region and month, I get Figure 4 below. 


Figure 4:  Bias by Physiographic Region and Survey Month for BC's 5 physiographic regions: Coast Mountains and Islands (CMI), Interior Plateau (IP), Northern and Central Plateaus and Mountains (NPM), Great Plains (GP), and Columbia Mountains and Southern Rockies (CR)


In figure 4, blues represent a negative bias, meaning the estimate is lower than what is measured, and greys mean insufficient data.  Whole season average values are printed to the left of each box.  Clearly there are regional differences in the bias across the methods I'm looking at, but all are underestimating to some extent.  Compare that to Figure 5 on the right showing mean snow accumulation (dark reds mean more snow).  The bias for all of these looks quite similar to the average snow accumulation.  This makes sense.  If gridded products are underpredicting, areas with more snow are likely to have a bigger difference between the estimate and the measurement.  A more fair comparison might be to divide the difference by the measured value.  This is known as relative bias.  


A similar plot for relative bias is show in Figure 6.  From this plot, I can see that while indeed most products are underestimating, ERA-Land and GLDAS-2 have the lowest magnitude relative biases across the two regions of highest accumulation, the Coast Mountain and Columbia & Rocky Mountain regions.  However MERRA and the CMC product have more modest relative biases in the Northern Plateaus and Great Plains respectively.  That tells me that these products perform better in the northern part of the province, and I can apply this knowledge to my efforts.  The ultimate goal will be to apply machine learning techniques to select and combine the SWE gridded products with other important information (elevation, forest cover, etc) and produce better snow estimates for BC.


Figure 5:  Mean snow accumulation by region and month for BC's 5 physiographic regions (as in Figure 4)


Figure 6: Relative Bias by Physiographic Region and Survey Month for BC's 5 physiographic regions 

Comments: 0 (Discussion closed)
    There are no comments yet.

  CanSISE Twitter Feed