Early thoughts on 2014 min extent, area, volume

[AndrewP]

Feel free to post CT_area #s and PIOMAS #s as well.

Also feel free to share methodology.

Jaxav2

2013: 4.81
2012: 3.18
2011: 4.27

[Sourabh]

Could you also provide previous years' minimum for reference? That would be helpful in judging.

[Jim Hunt]

It's far to early to start speculating?

Methodology - Observation of various flavours of those three metrics over the last couple of years.

[TerryM]

Jim
It's certainly far to early for anything but speculation, and will still be on Feb. 10th.


Perhaps by the middle of June I may have some inkling of where things will end up but I can't imagine that anyone has drawn any conclusions before the freezing season has ended.


The only indicator I'm aware of at this time is that Nares Strait has formed it's ice arch a little late in the season, but that's not much to draw from.


Terry

[ChrisReynolds]

Can someone post a list of the recent minima in JAXA?

Qualitatively - I say no new record in 2014, it may be that I'm unwilling to go further than that. My CT Area prediction will not come before 20 June 2014, and I don't intend to make a prediction using another metric.

[AndrewP]

Well there is a moderate correlation between winter volume and summer extent min. There's also Chris Reynolds' bins to look at. And the spatial structure of volume. Long-term weather/climate variables to consider. And a general concept of the current and future state of arctic sea ice.

Even in June the standard error for a prediction is close to a million sq km. An accurate prediction (within half a million sq km) cannot consistently be made until August.

I'm thinking 4.5 Jaxa, 3.2 CT area, and 4000 for PIOMAS.

[ChrisReynolds]

Andrew,

By June it should be clear if the 'summer pattern', that typifies summers from 2007 to 2012 and failed to appear last year, will make an appearance in 2014, and by 20 June the CT Area June cliff will have happened so uncertainty due to that is reduced.

Anyone...

My CT Area prediction merely used the observation of no trend in losses from late June to the minimum - is there a similar behaviour in any of the other indices? If there is someone could use that to make a prediction, hint, hint.

[AndrewP]

Andrew,

By June it should be clear if the 'summer pattern', that typifies summers from 2007 to 2012 and failed to appear last year, will make an appearance in 2014, and by 20 June the CT Area June cliff will have happened so uncertainty due to that is reduced.

Anyone...

My CT Area prediction merely used the observation of no trend in losses from late June to the minimum - is there a similar behaviour in any of the other indices? If there is someone could use that to make a prediction, hint, hint.

Yes, as we get nearer to September predictability rises but June isn't all that key. In April the r-square for a regression of area and volume to minimum area is .65. By the end of June that value has only risen to .7.

I didn't run the numbers but I imagine that the r-square in December or January using volume is still .6-.65.

The reason you find the period before June 20 so key is you are ignoring volume. Even on June 20th volume is still a better predictor of the area minimum than area is. Area doesn't become a stronger predictor of the minimum until mid-July, and even then the use of both variables is a stronger predictor than the use of either individually.

[ChrisReynolds]

Andrew,

Correlation for 1979 to 2012 with CT Area at dates stated vs day 250 (7 Sept), detrended using an interannual difference series for correlation.

30-Apr   0.272038679
10-May   0.006226453
20-May   0.274000362
30-May   0.292624413
09-Jun   0.295481939
19-Jun   0.522089928 Jump to stat sig (95%) here.
29-Jun   0.513130283
09-Jul   0.594507103
19-Jul   0.61986281
29-Jul   0.666348591
08-Aug   0.747536103
18-Aug   0.787401904
28-Aug   0.909748509
07-Sep   1

The method I will be using has a range between lower and upper bounds of 0.527M km^2 to make every post 2007 prediction a 'hit', that's just 17% of the average minimum extent for 2007 to 2013, not a large window.

I've found PIOMAS volume to be a far poorer predictor, even using gridded data to establish a prediction method. Crandles has had a go but I'm not sure what the outcome was. Volume loss (in grid cells >2m thick at April) accounts for the long term trend in CT Area loss, but PIOMAS predicts CT Area and NSIDC Extent poorly on a year to year basis.

But as this is CT Area and that's not what this thread is about, I'll leave it at that. Oops - didn't catch the bit about discussing any metric.

[AndrewP]

Andrew,

Correlation for 1979 to 2012 with CT Area at dates stated vs day 250 (7 Sept), detrended using an interannual difference series for correlation.

30-Apr   0.272038679
10-May   0.006226453
20-May   0.274000362
30-May   0.292624413
09-Jun   0.295481939
19-Jun   0.522089928 Jump to stat sig (95%) here.
29-Jun   0.513130283
09-Jul   0.594507103
19-Jul   0.61986281
29-Jul   0.666348591
08-Aug   0.747536103
18-Aug   0.787401904
28-Aug   0.909748509
07-Sep   1

The method I will be using has a range between lower and upper bounds of 0.527M km^2 to make every post 2007 prediction a 'hit', that's just 17% of the average minimum extent for 2007 to 2013, not a large window.

I've found PIOMAS volume to be a far poorer predictor, even using gridded data to establish a prediction method. Crandles has had a go but I'm not sure what the outcome was. Volume loss (in grid cells >2m thick at April) accounts for the long term trend in CT Area loss, but PIOMAS predicts CT Area and NSIDC Extent poorly on a year to year basis.

But as this is CT Area and that's not what this thread is about, I'll leave it at that. Oops - didn't catch the bit about discussing any metric.

My analyses have shown volume is a better predictor until nearly mid-July (The r-squared value is higher for volume). The best method uses both volume and area. This makes sense both logically, and is demonstrable statistically.

Just look at our R-squared values. My method achieves .74 by June 10th, compared to just .29 for yours. For June 20th, my method has .68 compared to .52 for yours. My R-squared values in early June are the same as yours in early August.

A method that includes volume is absolutely necessary for any prediction made before mid-July. Even after mid-July it is beneficial.

The other objection to your method is that the relationship between early season area and minimum area was almost assuredly different in 1979 than it is today. You are essentially forcing two different relationships into one regression.

[ChrisReynolds]

Sorry but you didn't explain what you'd done enough for me to follow. You did detrend your timeseries prior to correlating I presume, if you don't do that then all the correlation is telling you is that you have two timeseries trending in the same direction. Can you post hindcasts?


For what it's worth, here's my prediction method.

Actually, as we can discuss metrics other than JAXA, I'll explain the method I'll be using this year when CT Area data is in for 20 June.

I noticed that the CT Area losses from late June to the minimum have no trend, call this the 'late summer loss'. This is the core of the prediction method. I simply subtract the average late summer loss from 1979 to 2013 from the CT Area on 20 June to give the 'central estimate'.

Before 2007 I'd have had a two sided bound around this central estimate. However as shown in my most recent blog post:
http://dosbat.blogspot.co.uk/2014/01/post-2007-summer-melts-ice-or-atmosphere.html
Melts since 2007 in terms of CT Area and NSIDC Extent have been unusually aggressive in July and August for most years. So for hindcast predictions in the post 2007 era I use a scaled standard deviation (sigma) subtracted from the central estimate to give a lower estimate. In other words I make the projection one sided in favour of more loss than the 1979 to 2013 average. I've found that a scale factor of 1.6 applied to the standard deviation encloses all the actual minimum for years from 2007 to 2013 between the central and lower bounds predicted for each year. EDIT - edit removed - they are all in - I'm just tired.
So once I know the figure for 20 June 2014 (CT Area) I'll subtract 5.02437 to give my upper bound (central estimate) and subtract 0.52729 from that to give my lower bound.

Oddly, if I use the post 2007 statistics the performance is marginally worse, so I've preferred statistics for the whole satellite period.

Here are the hindcasts for the post 2007 period.

Year    Upper   Lower   Actual
2007   3.211919466   2.684628276   2.9194391
2008   3.354493766   2.827202576   3.0035558
2009   3.948287566   3.420996376   3.4245975
2010   3.185710566   2.658419376   3.0721295
2011   3.100165966   2.572874776   2.9047396
2012   2.746367566   2.219076376   2.2340095
2013   3.572611466   3.045320276   3.5543971

[AndrewP]

Here's my hindcasts vs actual

2007: 2.87 vs 2.91
2008: 2.94 vs 3.00
2009: 3.65 vs 3.42
2010: 2.89 vs 3.07
2011: 2.94 vs 2.90
2012: 2.44 vs 2.23
2013: 3.39 vs 3.55

I ran a regression for 2007-2013 for minimum area with X variables of June 20th area and June 20th volume. The strongest correlating variable was by far area.

The regression I gave previously was for 2005-2012. For this time period, area and volume both correlate. This is because it is impossible to predict the 2006 minimum without the inclusion of volume. I think you will find that your method is in error by close to 1,000,000 sq km for 2006, as is the above post-2007 regression. The regression I did previously (2005-2012), however, is only off by 480k for 2006. So it is more accurate for 2006, but is more in error for subsequent years.

June 20th area has been a remarkable predictor 2007-2013 but it failed badly in 2006. I think there may be something key about ice conditioning in this period, especially in the post 2007 era of thin ice. You could thus argue that area is a better predictor in the post 2007 era after June 20th. However, the area-only method fails for the year 2006. Inclusion of volume better predicts year 2006. Another 2006 is bound to come along at some point.

I think I may end up using a formula such as the following this year for June 20th:

.8*(area anomaly) + 100*(volume anomaly) + X

(this is basically a compromise between the 2005-2012 and 2007-2013 area+volume regressions. The former has coefficients of .35 and 150, while the latter has coefficients of 1.08 and 0, respectively.)


Either way, mid-June or earlier volume remains a reasonable predictor.

Hindcast for volume-only regression on April 30th using a 2005-2012 regression:

hindcast vs actual

2007: 3.17 vs 2.91
2008: 3.50 vs 3.00
2009: 3.54 vs 3.42
2010: 2.92 vs 3.07
2011: 2.49 vs 2.90
2012: 2.55 vs 2.23

Again, not bad for April 30th. Never off by more than 500k, and on average only off by 200-300k.

[jdallen]

Oh, I agree it's too early to speculate with any sort of reliability, but I like letting my intuition out for a walk occasionally, to test whether or not it is noticing something conscious logic doesn't. Sometimes it points me in a useful direction. Intuition says new record by a modest margin, but no melt out.

[crandles]

My impression was that area is a much better guide than extent and on occasion extent is correlated in the wrong direction. Then someone pointed out use of 3 factors

1. (extent-area)
2. extent
3. snow cover

ignoring the snow cover for the moment

-f1* (extent-area) + f2 * extent
=f1* area +(f2-f1)* extent

as factor f1 is larger than f2, this explains why the extent is correlated in the wrong direction.

I have been meaning to see what I end up with using this, but haven't done so yet. It looked like it should be better than using just area or just extent.

I tend to think that looking at short sequences of years like 2007-2013 is bound to cause trouble when you do a genuine forecast, so I prefer relationships using data for all the years from 1979 onwards which sometimes have to be non-linear. Noise about a trend in 7 data points is quite likely to have a spurious trend that can be quite large. It doesn't fully solve the problem of there being large scope for error - use data to 2006 to predict 2007, or data to 2007 to predict 2008 or data to 2012 to predict 2013 and it is clear that it is easy to go badly wrong.

I suspect that snow cover may help projections from early in the melt season but will rapidly become unimportant as the melt season progresses. I haven't tested this but snow cover gets smaller and further from the sea ice edge.

[AndrewP]

Oh, I agree it's too early to speculate with any sort of reliability, but I like letting my intuition out for a walk occasionally, to test whether or not it is noticing something conscious logic doesn't. Sometimes it points me in a useful direction. Intuition says new record by a modest margin, but no melt out.

Well I don't know if you would call this reliable or not, but each of the last 7 years could be predicted within 500k using winter volume. That's better than most July predictions that I've seen.

[jdallen]

Oh, I agree it's too early to speculate with any sort of reliability, but I like letting my intuition out for a walk occasionally, to test whether or not it is noticing something conscious logic doesn't. Sometimes it points me in a useful direction. Intuition says new record by a modest margin, but no melt out.

Well I don't know if you would call this reliable or not, but each of the last 7 years could be predicted within 500k using winter volume. That's better than most July predictions that I've seen.

I'll have to evaluate that assertion, but as is true with many correlations ( such as predictions of stock market behavior )  it is dangerous to depend interpretation of an effect to draw a conclusion, without understanding of underlying causes. I'd add, that the similarities in those numbers exist only via the most tenuous of cursory examination.  Each seasons ice has been radically different from the others, when evaluated in detail.  As with the stock market, at this stage my intuition would be as reliable numbers derived by simply looking at unrelated scalar values.

[AndrewP]

I'll have to evaluate that assertion, but as is true with many correlations ( such as predictions of stock market behavior )  it is dangerous to depend interpretation of an effect to draw a conclusion, without understanding of underlying causes. I'd add, that the similarities in those numbers exist only via the most tenuous of cursory examination.  Each seasons ice has been radically different from the others, when evaluated in detail.  As with the stock market, at this stage my intuition would be as reliable numbers derived by simply looking at unrelated scalar values.

Well given the high heat of fusion of ice, it shouldn't be too hard to guess why the volume of ice at the start of any given year might correlate well to area 6-9 months later that same year.

Also, the accuracy of the hindcast is not 'tenuous and cursory.' The error is less than 500k for all years 2005-2013 over which span the difference between the largest and smallest year was 1400k.

The P-value is also .01 indicating 99% confidence and only a 1% chance the correlation occurs by chance.

Your post fits into the general tendency of this forum to mystify the melt-season processes, when in fact area and volume losses over the course of any given season have been remarkably consistent from year to year.

[werther]

AndrewP,

In the 10 months that you’ve been active on this Forum, you have been taking on almost all of our regular posters here, debating their concerns.
You did that, in diplomatic terms, elegantly. You expressed your own concerns to be credible, you argued on the basis of facts and logic where you dealt with the informed, bright posters found in this community. Some have been extremely patient, not in the least with regards to the valuable content of the Forum.

You are fully entitled to express your opinions. I’d even defend your presence on this Forum. Even though I completely disagree with the general tendency of your posts.

That tendency is to minimize concern on all aspects of global warming.

Out of my interest in this Forum (and Neven’s blog) and my own tendency to a sort of masochist pleasure in annoying my nerves I’ll probably keep reading your posts as well as most other content.
Nevertheless I’d humbly ask Neven to moderate you when you try to propell your assumptions by extrapolating them to be generally and scientifically valid. The same when you presumptuously rake all our posters on one heap on whatever subject you like.

[AndrewP]

I performed and posted the results of several linear regressions. I made no claim of scientific certainty and I submitted my results for consideration by others. I revised one or more of my assertions in response to valid criticism and defended them against invalid criticism. I don't see the issue. I also don't see how any of my posts in any way attempt to minimize AGW.

I would suggest attempts to sidetrack this thread from what was otherwise substantive discussion be moderated.

As I see it my points have been,

1. A regression that includes both volume and area is superior at predicting the minimum to one that includes only area until at least late July.
2. Volume is a better predictor than area until at least June 15th.
3. Volume is a useful predictor up to 9 months in advance at making reasonably accurate minimum predictions. Most of the incorrect 'new record' predictions made by this forum could have been statistically ruled out as reasonable guesses with very high confidence as early as February last year.

I don't see how any of these points 'minimize AGW.' And I don't see how any of these points make a claim to scientific certainty that is not justifiable or is beyond the ability of other posters to draw their own conclusions.

[ChrisReynolds]

I've dug out my spreadsheet in which I decided volume wasn't as good as the later area method I outlined for CT Area, I've updated it. The linear best fit to a scatter plot of Volume at daily maximum, and CT Area at daily minimum gives an equation:

MinArea = 0.2387*MaxVolume -2.3484.
R2 = 0.8003.

In reality a power function is more likely, but a linear fit probably works OK for this year and as Andrew is working with linear (ax +b) form equations I'll follow suit.

I've used the above equation to calculate a projected area for each year based on the actual volume at maximum. I've then worked out the difference between the two figures for each year - the error. The largest errors are +0.653M km^2 and -0.989M km^2. Which is why I concluded that this method wasn't as good as using CT Area on June 20th.

Andrew's method is different from this, so the stats and performance will be different.

Andrew,

2006 is an outlier, at -2.3 sigma, but any practical prediction system will always encounter outliers. In any case the method I use is adapted for the post 2007 period, which exhibits greater summer losses, as I outlined in my recent blog post I suspect this is due to ice state causing greater melt.

The idea of substantial predictability based on winter volume excites me because I have been saying for some time that extent and area are sideshows with volume being the key issue. Such predictability would fit into that framework very neatly.

Have you considered that in using the total volume for your regression you're introducing noise due to ice volume outside the  Arctic Ocean which plays no part in the state at minimum? You might want to try using the volume within the Arctic Ocean at maximum, I've got regional volume, using Cryosphere Today's regions, worked out here:
http://dosbat.blogspot.co.uk/2013/12/regional-piomas-volume-data.html
See under 'The Data', that data is derived from PIOMAS gridded data, which is monthly, however it was pointed out to me last year that using daily values might be a further source of noise, so using monthly averages might be of use.

I've just redone the method I outlined above using Arctic Ocean, Greenland Sea and CAA total April Volume, the std dev drops from 0.371 to 0.358, but there are still large outliers, the range is from -0.92 to 0.565M km^2. Not a great gain but your method may benefit.

[crandles]

using Arctic Ocean, Greenland Sea and CAA total April Volume

Why include CAA and Greenland? They seem fairly irrelevant to me.

I would have thought CAB+ Laptev+East Siberian+Chuchi+Beaufort. CAB being where ice is left while Laptev and ESS in particular are important as to whether you get a quick start in the melt season (The pre-conditioning bits per The Thinning of Arctic Sea Ice, 1988–2003: Have We Passed a Tipping Point? Lindsay and Zhang)

[ChrisReynolds]

Well CAA, because it counts towards the minimum. But I conceded the point regards Greenland  because the ice there is in transit.

I've only just had dinner and will read your comment above and reply in a while.

[ChrisReynolds]

My impression was that area is a much better guide than extent and on occasion extent is correlated in the wrong direction. Then someone pointed out use of 3 factors

1. (extent-area)
2. extent
3. snow cover

ignoring the snow cover for the moment

-f1* (extent-area) + f2 * extent
=f1* area +(f2-f1)* extent

as factor f1 is larger than f2, this explains why the extent is correlated in the wrong direction.

I have been meaning to see what I end up with using this, but haven't done so yet. It looked like it should be better than using just area or just extent.

I tend to think that looking at short sequences of years like 2007-2013 is bound to cause trouble when you do a genuine forecast, so I prefer relationships using data for all the years from 1979 onwards which sometimes have to be non-linear. Noise about a trend in 7 data points is quite likely to have a spurious trend that can be quite large. It doesn't fully solve the problem of there being large scope for error - use data to 2006 to predict 2007, or data to 2007 to predict 2008 or data to 2012 to predict 2013 and it is clear that it is easy to go badly wrong.

I suspect that snow cover may help projections from early in the melt season but will rapidly become unimportant as the melt season progresses. I haven't tested this but snow cover gets smaller and further from the sea ice edge.

I'm struggling to see the utility of your extent/area/snow idea, partly because Wipneus has shown that his area calculations differ from CT Area during the summer, indicating that what will emerge is a rather hard to interpret combination of two unlike datasets. But also I'm not convinced snow cover is of help when we have PIOMAS, I'm a bit unclear on this but it's my impression that PIOMAS uses NCEP/NCAR precipitation for snow cover. So in effect it's doing the factoring of snow impacts when calculating volume.

Ice state as shown by the June Cliff in CT Area - flat ice and melt ponding preconditioning the ice, seems to be a stronger and more immediate factor in the melt season, it also explains the fall off of correlations before this time.

I've not had the time to do a thorough comparison of using area/extent, I've not even had time to look for the same trendless late season losses in CT Area when using NSIDC Extent.

I agree with regards using short sequences, I should point out that at the core of my late June CT Area prediction method are the stats for the entire period 1979 to 2013. The concession to the post 2007 period is merely due to the observed positive anomalies of intermonth losses for June July and August, and that's coped with by adjusting the scaling of sigma (std dev'n) and using one sided bounds.

[ChrisReynolds]

Andrew,

Sorry to spam the thread with a load of replies, but I don't want to be tied up with this tomorrow night as I've a hard day at work tomorrow, and I've done the math now.

I am highly sceptical of your claims of high R2 over the winter - read that as a polite 'I think you're wrong'. Not on everything, or your claims of early prediction per se, just on this technical point. I asked you to clarify about detrending, but a lot has been discussed and you seem to have forgotten.

All data are derived from monthly averages, CT Area is area, PIOMAS volume is volume. I assume that where correlations are high (low) R2 will be high (low). All correlations are with the September at the end of each year (a year running from October to the following year's September).

A. Area.
Oct   0.78
Nov   0.80
Dec   0.83
Jan   0.81
Feb   0.81
Mar   0.71
April   0.60
May   0.63
June   0.88
July   0.94
Aug   0.99
Sept   1.00

B. Volume.
Oct   0.84
Nov   0.83
Dec   0.84
Jan   0.84
Feb   0.86
Mar   0.87
April   0.87
May   0.88
June   0.91
July   0.93
Aug   0.94
Sept   0.94

Note how in both cases the correlation of each stated month is high all through the winter, which supports what you are claiming.

I've detrended the data by calculating each month's area or volume as relative to the September of the previous melt season, so these show changes throughout the winter and following summer with no multi-year trend.

C. Each month's area vs area in September.
Oct   0.46
Nov   0.54
Dec   0.53
Jan   0.52
Feb   0.55
Mar   0.48
April   0.46
May   0.47
June   0.63 Jump!
July   0.85
Aug   0.96
Sept   1.00

D. Each month's volume vs area in September.
Oct   -0.00
Nov   0.18
Dec   0.29
Jan   0.38
Feb   0.48
Mar   0.51
April   0.56
May   0.66 Jump!
June   0.72
July   0.73
Aug   0.79
Sept   0.80

E. Each month's volume vs volume in September.
Oct   0.19
Nov   0.29
Dec   0.32
Jan   0.43
Feb   0.51
Mar   0.53
April   0.58
May   0.75 Jump!
June   0.84
July   0.90
Aug   0.98
Sept   1

So what is happening here? (I should note - I've previously been justly told off by Peter Ellis for screwing up by not detrending before correlating)

In A and B what is being detected is that all the months have downward trends, so the correlations are telling us nothing about potential limits to predictability and the strengths of linkages between various months and the end of the freeze/melt season cycle, they are just telling us all the months have downward trends. This is the same for the R2 of a regression, fail to detrend and you get the message that all months have downward trends.

In C, D and E, we are getting information about the actual linkages between various months in the seasonal cycle and the strengths of those linkages. In all of these cases correlations rise from low over the winter to high from May to June onwards. This doesn't mean that predictability is impossible over the winter, or certain over the summer, but it does mean that one should expect poorer performance from a winter/spring prediction than for a prediction made later in summer. Poorer performance implies wider uncertainty bounds to the prediction.

Of course it may simply be that the winter technique is inately better than the summer one, that is a different matter. But claiming greater correlation or R2 over winter than for a detrended series tells us nothing about the method because it is merely detecting the trend of losses throughout the years.

[jdallen]

I'll have to evaluate that assertion, ...

Well given the high heat of fusion of ice, it shouldn't be too hard to guess why the volume of ice at the start of any given year might correlate well to area 6-9 months later that same year.

Also, the accuracy of the hindcast is not 'tenuous and cursory.' The error is less than 500k for all years 2005-2013 over which span the difference between the largest and smallest year was 1400k.

The P-value is also .01 indicating 99% confidence and only a 1% chance the correlation occurs by chance.

Andrew;

My choice of language was poor and was not intended to characterize you.

My disagreement with you has to do with the logic behind the method.  I think you touched on a key aspect of that with the word "Hindcast".  Your methodology derives its substance from the premise that the ice itself provides an effective predictor.  I disagree with that premise, and rather submit that it is estimations of *heat flow* which are the best predictors.  The challenge is in the chaotic system we have, establishing that.  Ice in and of itself provides a useful correlation only because conditions governing heat flow in the past were enough similar to produce a reasonable model of behavior for that history.

Similar to the stock market, when attempting to develop a prediction of price based on past price behavior (something with which I actually have some reasonably broad experience), you can *absolutely* derive an equation or mathematical operation which will describe the relationship between a sequence of values over time. It will absolutely have the appearance of being predictive, but in fact is completely dependent on the driving conditions behind the value of the stock, rather than any particular state of the price at a given time.

Apply that to the behavior of ice.  Your presumption of predictability relies heavily on the environmental conditions around the refreeze and melt being reasonably consistent.  It *will* fail, as stock price prediction, when the underlying drivers supporting or undermining the arctic pack change.  I think we are watching exactly that unfold right now.

Hindcast and correlation are good tools - they describe past behavior.  They also can lead us to understanding what the actual mechanisms are driving the change.

Your post fits into the general tendency of this forum to mystify the melt-season processes, when in fact area and volume losses over the course of any given season have been remarkably consistent from year to year.

I'll take issue with you here.  "Mystify"??  Not hardly.  The science and physics is really pretty concrete.  What is confusing is the chaos of the many interactions that take place, and the sheer scale of the system, which results in highly discontinuous behavior.  An illustration of this can be found in the core statistic of AGW itself - CO2 in the atmosphere - and the difference between past paleoclimate estimates of temperature and what we see now.  The effects have yet to catch up with the change in state.  Another illustration is pretty clearly visible in our current weather; significant low temperature anomalies across much of central and eastern North America, while Alaska, much of Siberia, and portions of the high Arctic itself see temperatures skewed similarly in the opposite direction.

To wit; the temperature now is not a good predictor of that in the future, if CO2 in the atmosphere stays the same. 

In view of this, I would say rather, your post is illustrative of the perception of many people, who tend to oversimplify a dynamic system and overlook its complexity.  I consider it incomplete.

[crandles]

I'm struggling to see the utility of your extent/area/snow idea, partly because Wipneus has shown that his area calculations differ from CT Area during the summer, indicating that what will emerge is a rather hard to interpret combination of two unlike datasets. But also I'm not convinced snow cover is of help when we have PIOMAS, I'm a bit unclear on this but it's my impression that PIOMAS uses NCEP/NCAR precipitation for snow cover. So in effect it's doing the factoring of snow impacts when calculating volume.

The idea re extent and area is that outside the extent boundary there is ocean/land with low albedo which gathers heat. By ocean current or winds that heat can be carried towards ice or away from ice with around 50% probability for each.

When it comes to polynia, the extra heat absorbed will be moved by currents or wind but whichever direction it goes it comes into contact with ice with near 100% probability. So not all areas without ice cover are the same - those within the pack are more important.

Sorry I cannot remember who suggested this.


My comments on snow cover were about snow on land not about snow on ice hence my talk of it only being relevant early in the melt season.

[Chuck Yokota]

AndrewP: Forgive me if I am misunderstanding you, but did you say that you produced your hindcasts using data from the same years that you used to create your regression?  In that case, the hindcasts would not provide validation of your method, but simply show how precise your curve-fitting was.  There could be any number of spurious relationships that happen to work for that particular set of data, but could disappear when applied to fresh data from another year.

[AndrewP]

Yes Chris I am also having a hard time keeping up.

No I have not been detrending. I've been running various regressions using area, volume, or area and volume to predict the minimum. The hindcast I posted used a regression 2005-2012.

So yes, part of the high R2 is due to the long-term trend in both the X and Y variable (declining). Nevertheless I believe a lot of that relationship is causal. I don't care what the weather was in 1980, the year began with twice as much volume as recent years, and the minimum area would be way higher even if there was 2007 or 2012 weather.

I like your method, the only objection I have is that if the previous year had a low (high) area-volume ratio, them there would be a tendency using the volume correlation to underestimate the rebound (drop off).

Perhaps detrend by using the 5-yr trailing average?

I can agree that knowing this January's volume doesn't provide us with a lot more predictive power than what we had 3-6 months ago already. Running some daily correlations it doesn't really provide much power until late April. It remains a better predictor than area until July.

The key is knowing what your starting condition or background state is.

A method that uses both area and volume would be useful during the months of June and July. In May we can get some good predictive power from volume alone.

I also did a bunch of regressions testing out different Seas to include, and didn't find a whole lot of improvement over just using total volume, unfortunately.

[ChrisReynolds]

Crandles,

Thanks for clarifying, I agree that snow cover could prove a useful proxy for land warming later in the summer with consequent impact on melt. I still keep pondering whether there is a role for snow loss on the early season increased melts shown by PIOMAS post 2010, although I'm tending once again towards ice state causing that.

Coincidentally, since I got up this morning I've been hacking away at the issue of late summer PIOMAS losses having reduced. I need to think about the data I've just produced, but at present I think it supports the suggestion by both yourself and Dr Zhang that it's due to there being less volume of ice from which melt can occur.

Chuck,

This is why I've used the method I use based on a notable lack of trend in late summer losses. The prediction then becomes a useful test of whether the ice state is changing in respect of the late summer losses- last year I was told ice state meant my prediction would fail, it didn't, therefor ice state from 1979 to 2013 hadn't changed so as to affect late summer CT Area losses.

I'm sure you'd agree that while hindcasts don't validate a model they're a critical test of any proposed forecasting method.

Andrew,

There will be a causal element in that ice state over winter will impact the following summer. However due to weather impacts through spring and summer it is to be expected that the correlations in winter (with the minimum) will be less than those in summer, and that even in summer, the correlations with the minimum climb to '1'.

The only method I'm using is the 20 June forecast of CT Area in terms of CT Area, I merely produced the forecast of area in terms of winter PIOMAS volume to show why I'd concluded that method wasn't as satisfying. Area/Volume ratio is inverse of thickness, including thickess (Volume/Area) would be an improvement to that method.

I think that using April volume is really the earliest feasible, although as the volume increase rate decreases approaching that time perhaps March could be used.

I suspected any gain from using volume of seas in the Arctic Ocean without ice volume elsewhere would be small, but it still might be worth revisiting if a method using April volume has been established in order to see if it gives a marginal improvement. By far most of the volume is within the Arctic Ocean in April, so the gain in fidelity will be marginal.

[ktonine]

While any of the statistics we have during winter may not be accurate or reliable predictors of the fall minimum (because summer weather).  It's still interesting to look at year-to-year comparisons.

Taking an idea from wayne's blog, I submit a comparison of surface OLR for DEC 2012 and DEC 2013





Make of it what you will, but I see this as foretelling a very interesting melt season.

[Andreas T]

How are these values of outgoing longwave radiation arrived at? I take these as mainly representing surface temperatures.
What makes the coming season interesting for me is the thick ice in the beaufort sea according to this:http://www7320.nrlssc.navy.mil/hycomARC/navo/arcticictn/nowcast/ictn2014013118_2014020100_038_arcticictn.001.gif
Either this gets wiped out in a warm melt season or withstands a cool melt season. Looks like the arctic gambles with high stakes.

[jdallen]

2012 and 2013 at about the same date for comparison.
http://www7320.nrlssc.navy.mil/hycomARC/navo/arcticict/nowcast/ict2013020118_2013013000_035_arcticict.001.gif

http://www7320.nrlssc.navy.mil/hycomARC/navo/arcticict/nowcast/ict2012020118_2012020100_035_arcticict.001.gif

I really don't like the look of all that thick ice in the Beaufort.  I suspect it is doomed unless it gets pushed to the northeast.

It seems like there are gross similarities between 2012 and 2014 in the distribution of ice between 1.5 and 2.5M.  Echoing others elsewhere, I think what happens in May and June will be most crucial. Considering how cool the summer was last year,  I suspect the *best* we can hope for is a repeat of 2013.

[ChrisReynolds]

Andreas,

A primary cause of the 2010 volume loss event was the export of a large amount of MYI into Beaufort and thence to Chukchi. This ice then melted during the 2010 melt season. This caused a large volume drop but not a large area/extent drop as the exported thick ice seemed to retard area/extent losses.
http://dosbat.blogspot.co.uk/2013/03/what-caused-volume-loss-in-2010-part-2.html

It is possible that this large export indicates a muted melt of area/extent next year (particularly with persistent ice in Chukchi and the ESS) but with a large volume loss. This could wipe out the gain of 2013 remarkably quickly.

Referring to my blog post linked to above, there's an SLP plot in it showing the role a dipole had in the export event. There's a similar dipole in the January average SLP for 2014 so far. However winter 2010 had remarkably strong high over the Arctic (2010 being an extreme WACC year), that's not really the case this year.

[ChrisReynolds]

Spurred by Ktonine's comments about earlier melt in PIOMAS on the sea ice current conditions thread...

The linear trend has a slope of +0.12 days per year, later melt, over the full 35 years that's an advance of 4 days, i.e. peak volume gets later. It's very noisy but even visually the trend is clear. I need to look at gridded regional data to get a grip on why this is happening. I suspect that thinner ice in the central arctic leads to more growth.

Anyway...

Some years show a decline in volume between 1 April and 1 May volumes, suggesting volume loss started in April, rather than May as is more typical (where the decline starts from May 1 to June 1, with gain before then). Specifically these are:

1988
1990
1995
2007
2010
2013

Now 2010 and 2007 are large volume loss years...

1988 -0.323 (k km^3)
1990 -0.970
1995 -2.426
2007 -2.535
2010 -2.257
2013 +1.719

Of these years only 1988 is below average loss (-0.337), the rest are above average, apart from 2013.

1991 lost -3.55 on the previous year, 1998 lost -2.630 on the previous year, 1993 lost -2.63, 2002 lost -1.387. So lack of an early freeze does not prevent large losses.

But looking back at the list above, after 1995 we're in the period of rapid volume loss, and apart from 2013 the other three years showed very large losses of volume after an early start. In 2013 the initial gain of volume (w.r.t. 2012 ) was in May - without that would we have seen another large loss of volume.

I suspect this might be a pattern worth keeping an eye on when making predictions, which is why I posted this here.