As a quick follow-up on the topic of predicting breakup dates, I looked at whether breakup is more closely related to hourly temperatures above freezing rather than daily mean temperatures above freezing. This is something that reader Eric suggested a while back, and reader BJ re-emphasized that daily average temperatures could be misleading.
I pulled out hourly temperature data from 1998-2018 in Fairbanks and calculated the accumulation of thaw degree hours up until the date of breakup in Nenana, i.e. the sum of the hourly temperature excess (if any) above 32°F. Historical hourly data from Nenana is not quite complete enough for this task, so data from Fairbanks will have to do.
Here's a scatter plot of thaw degree hours (TDHs) versus thaw degree days (TDDs) accumulated through breakup; click to enlarge.
Not surprisingly, it's a very good relationship, but the best-fit line is flatter than would be expected for a one-to-one relationship. Superficially, this suggests that hourly data is indeed a better predictor of breakup date than daily data, because there is less variance in TDHs at breakup than in TDDs at breakup. (Imagine if TDHs were a perfect predictor: the best-fit line would be horizontal.) The standard deviations of the two variables demonstrate the difference: the standard deviation of TDHs at breakup is 25% of their mean value, whereas TDDs have a standard deviation of 31% of the mean value.
A bit more work confirms that the typical range of thaw degree hours at breakup is associated with a slightly smaller date window than for thaw degree days. See the chart below, which attempts to illustrate this. If we're using daily data, we find that a typical (80%) range of TDDs at breakup corresponds to a 7-day climatological window, but if we use hourly data the window is narrowed to about 6 days. It's obviously a small difference - hourly data is no holy grail for breakup prediction - but it seems that readers were correct to suggest that we can do better than daily average temperatures. A more thorough investigation would require a more careful modeling effort with daily data... perhaps I'll return to that next spring.