My thesis is titled Investigating the spatio-temporal relationships between snow melt timing and wildfire occurrence in the US Mountain West.

I have created a new set of snowmelt timing surfaces using the MOD10A2 Snow Cover Product developed by the National Snow and Ice Data Center (NSIDC) as my data source for weekly measurements of snow cover across the continental US from Denver to the West Coast.

Please see our recent data publication with the Oak Ridge National Lab:

Snow Presence Video

In this animation blue is snow covered area, white is cloud interference. Wait for the third year to see the change detection algorithm in action.

I developed a change detection algorithm that compares these scenes to determine when each pixel changes from snow to no-snow and stores that value in a new raster image, creating the snow melt timing map shown here:

Snow Melt Timing Map



This image shows mean snow melt timing for 2001-2015, with warmer colors indicating snow loss early in the season and cooler colors indicating a snow loss later in the season. As we would expect the higher elevations of the major mountain ranges lose snow late in the season where as the low elevations through the great plains and the great basin lose their snow very early in the season. I have also developed annual snowmelt images for 2001-2015 and compared them to the mean to calculate a z-score for the melt timing of each pixel per year.


Here we see a close-up of the Sierra Nevada, including Lakes Mono and Tahoe. This image shows well the resolution of the MOD10A2 parent product (500m) and how the data quality suffers at the edges of snowy areas. The central crest of the Sierras has a rich density of snowmelt values, where as the lower foothills have spotty coverage, and therefore many edge pixels have only 1 year of recorded snowfall, and that is the only value in the average for that pixel. It is important to take into account the number of years on record for these observations:



When you compare these z-score values to wildfire area burned in each Ecoregion for each year you can look for trends. I performed a Spearman’s rank correlation for each EPA Level III Ecoregion comparing annual values of snowmelt timing and wildfire total area burned. Even though it may make sense that an early snowmelt would lead to a larger fire season this was generally not the case. Shown here is the resulting p-value from the Spearman’s rank correlation comparing snoment timing with fire on an annual basis. A low p-value (<0.10) is considered a strong correlation. The only Ecoregion reporting a p-value <0.10 is the west coast mountains, which has a positive relationship (later snowmelt leads to larger wildfire season). This may or may not be ecologically significant, and may result in part by the burning of slash piles in this heavily timber-managed Ecoregion.



If we look at these Ecoregions individually we see some general trends. First, the top-right quadrant is often vacant of points. This is the quadrant where large fire years follow late snowmelt years. Second, we see the simple linear regression is almost always negative. Note, the fire data are not normally distributed, so the simple linear regression is inappropriate and the p-values shown are deflated (look more significant than they are).

Ecoregions_Plots_lm_1Ecoregions_Plots_lm_2Ecoregions_Plots_lm_3This leads me to believe that at the regional scale there is little relationship between snowmelt timing and wildfire occurrence. Snowmelt timing would most influence soil and fuel moisture contents, and while a later snowmelt may contribute enough moisture to quench fuels, dry conditions leading up to the fire season may overcome that moisture and return the fuels to a fire-prone condition.