The coming zero snow depth season #3

One of a series; also see editions: #1, #2, #4.  Care, some of what follows isn't pretty.

What’s needed to get an answer here is to connect a robust prediction of future climate to our future snow depths. I don’t have to tell a snow weather audience that the best predictions come from the best, most sophisticated models; we see that every week from ECMWF and BOM ACCESS. The best predictions of future climate come from sophisticated global climate models. The challenge for predicting future snow depths in a tiny patch of alpine Australia is that the current crop don’t have anything like the required resolution or fidelity for that task.

The approach others have adopted is to build a strong statistical and mechanistic connection from the global model to the local effect.  The Hennessy study and its recent Victorian update use global climate model output to infer a regional climate prediction (by down-scaling) from which they interpolate local weather parameters to drive a high-resolution distributed snowpack model¹.  The original Hennessy report is pretty dated and provides insufficient detail for our purpose, but the 2012 update² (Bhend, Bathols & Hennessy 2012) is much more interesting.  Here’s its figure 16:

Bhend_Bathols_Hennessy_figure_16

 

That shows peak snow depths at Rocky Valley Dam near Falls Creek, Victoria, Australia since 1954, with prediction bands for 2020 and 2050 from the study’s snowpack model.  The three bands at each are for three IPCC SRES greenhouse gas emission scenarios: B1 (low), A1B (medium), and A1FI (high).  The SRES scenarios were superseded for the most recent IPCC Fifth Assessment Report (AR5 WG1, 2013) by “representative concentration pathways”, with RCP6.0 (medium) and RCP8.5 (high) looking pretty near the track we’re on:

Global_monthly_temps_rev7_extrapolated

 

I think it reasonable to focus on Bhend et al’s medium and high A1B and A1FI scenarios, which had similar intent to the medium and high RCPs.  The prediction bands shown by Bhend et al for 2020 and 2050 reflect the range of results from the 18 different GCMs they used to drive their snowpack model.  The lines connecting those add a bit to that range to represent “natural variability”, but we’re going to manage that separately.  To progress, we need to generate a continuous prediction function (which Bhend et al don’t provide).  I’m going to do that by fitting³ a second order polynomial (a quadratic, or parabola), which is a highly parsimonious choice for a dataset with obvious acceleration.  I also fill the data gaps in the Rocky Valley Dam record by correlation⁴ with the Spencers Creek record (100 km away, near Charlotte Pass, NSW, midway between Perisher Valley and Thredbo), and extend the moving average trace out a bit by progressively reducing the 20-year averaging interval down to 10 years at the end (the dashed extension)⁵.

Here’s what we get:

Rocky_Valley_peak_trend_extrapolated

 

It’s striking how well the polynomial trend fits the moving average trace, and also fits the Bhend et al prediction bands.  The trend hits zero in 2067, just 53 years from now.

 

So how soon can we expect to see the first “real” zero depth season at Rocky Valley Dam⁶?  That’s easy; just compute the residuals from the trend (differences between raw data and trend), fit a distribution⁷, and extrapolate the resulting band of variability assuming 50% residual proportionality⁸ as previously adopted for the Spencers Creek record:

Rocky_Valley_peak_trend_extrapolated_band

 

It remains to integrate (add up) the intersection of that variability band with zero.  That gives:

Rocky_Valley_probability_zero

 

That’s about 50/50 by 2040 (26 years away) and nearly certain by 2060.

Note that that’s a “best estimate”; it doesn’t attempt to incorporate estimation uncertainty (as opposed to natural variability), for example as suggested by the length of the Bhend et al estimation bars⁹. Also, much of the Falls Creek resort area is higher and further west (with more precipitation) than the Rocky Valley Dam measuring course, so zero natural snow across the resort would be expected to happen somewhat later.

Next I want to try a similar approach for the high altitude Spencers Creek snow depth record. Zero at Spencers Creek means virtually no snow anywhere in Australia (there is higher and better exposed terrain than Spencers Creek, but not much).

 

Notes:

1. Details are in my earlier posts and the reports linked there, including 2.:

2. Climate change impacts on snow in Victoria (2012), J Bhend, J Bathols and K Hennessy; Centre for Australian Weather and Climate Research (a CSIRO / Bureau of Meteorology partnership)

3. The fit is to the moving average trace (not the raw data), plus the mid points of both the A1B and the A1FI prediction bands at 2020 and also at 2050 (both are included to increase their weighting in the regression). This approach is dodgy of course, but actually appears to work rather well. [It’s closer to engineering than science. Engineering is about doing; you have to progress with a workable approximation … not wait for the definitive answer that never arrives.  Science is about understanding, which can sometimes be obscured by careless approximation.]

4. The correlation coefficient is 0.86.  The Rocky Valley Dam peak depth data is missing for 1973 (record annotated “insufficient snow for survey”; 55 cm by correlation), and I don’t currently have 2012 (127 cm by correlation) and 2013 (104 cm).

5. The moving average is plotted centred (not trailing, as finance types are want to do).

6. Based on the Spencers Creek correlation, the 1973 “zero” (“insufficient snow for survey”) looks to be more the result of incomplete dedication to measurement than genuine absence of snow.  Perhaps someone can recall 1973 at Falls?

7. A Pearson type III distribution, to allow for the slight positive skew in the residuals.

8. To allow for some heteroscedasticity.

9. The Bhend et al estimation bar lengths derive from the spread of the GCM ouputs used to drive their model. It’s questionable whether that fairly represents the overall uncertainty implicit in their analysis. (If anything the polynomial fit would tend to suggest that the 2020 bars probably overestimate uncertainty. The 2050 bars are too far out in the extrapolation to judge.)

7 comments to The coming zero snow depth season #3

  • […] The coming zero snow depth season #3    First snow […]

  • […] One of a series, also see editions: #2, #3 […]

  • Matthew

    How important is snow melt to filling up the dams. Would a zero snow season have a detrimental effect on water supply or electricity supply in NSW in the seasons that followed? Or would a problem only exist in earnest if there were many successive seasons of dry weather? Great blog.

    • Gerg

      Tks. Snowpack provides “free” short-term water storage for users of alpine runoff: in Australia the hydro schemes, irrigators and town water supply, not to mention environment along rivers. I did some numbers a while back on the significance of a biggish Kosciusko snow accumulation to a (then) nearly empty Eucumbene Dam (Snowy Hydro’s largest storage). I can’t remember the answer; I think it added some fair bit, but not enough to fill it. It’s obviously significant — that’s why both Snowy Hydro and AGL Hydro measure snowpack depth & density; money in the bank to them. But how significant? Don’t know.

  • Michael tee

    Hi Gerg, still stand by your max prediction of 170cm+-? Given the current warm records in May is there reason to be less optimistic?

    • Gerg

      I’ll do an update next month after various parameter sources update. ENSO is actually looking a bit less threatening, and SST is yet to rise to what I assumed, so could even improve slightly. There’s nothing in the data to indicate a wipeout season. (But there wasn’t in 2006 either…)

  • […] One of a series, also see editions: #1, #2, #3. […]