Bob Tisdale writes:
I’ve been holding off telling you about my most recent post in hopes that GISS would continue with their warmest-year nonsense. And they did.
Using correlation maps, animations, graphs and a youtube video, the post shows how leftover warm water from an El Nino gets spun up into the Kuroshio-Oyashio Extension (KOE) where it continues to release heat during the La Nina. The KOE correlates with the Northern Hemisphere warming during an La Nina, and one of the datasets used for the graphs and correlation maps is GISTEMP LOTI.
The ENSO-Related Variations In Kuroshio-Oyashio Extension (KOE) SST Anomalies And Their Impact On Northern Hemisphere Temperatures
Guest post by Bob Tisdale
OVERVIEW
This post provides brief background information about the Kuroshio-Oyashio Extension (KOE), and discusses the relationship between NINO3.4 SST anomalies and the SST anomalies of the KOE following major El Niño events. Using correlation maps the post also illustrates the possible impacts of the KOE Sea Surface Temperature (SST) anomalies on North Atlantic SST anomalies, Combined Land and Ocean Surface Temperature anomalies, and Lower Troposphere Temperature anomalies.
INTRODUCTION
The Kuroshio Current and Oyashio Current are located in the western North Pacific Ocean. The Kuroshio Current is the western boundary current of the North Pacific Subtropical Gyre. Its counterpart in the North Atlantic Ocean is the well-known Gulf Stream. The Kuroshio Current carries warm tropical waters northward from the North Equatorial Current to the east coast of Japan. The East Kamchatka Current and the Oyashio Current are the western boundary currents of the Western Subarctic Gyre. The East Kamchatka Current is renamed the Oyashio Current south of the Bussol Strait (which is located about half way between Hokkaido and the Kamchatka Peninsula). They carry cold subarctic waters south to the east coast of Japan. The Kuroshio and Oyashio currents meet and form the North Pacific Current that runs from west to east across the North Pacific at mid latitudes. The Qiu, (2001) paper Kuroshio and Oyashio Currents. In Encyclopedia of Ocean Sciences, (Academic Press, pp. 1413-1425) provides a detailed but easily readable description of the two currents. Figure 1, from Qiu (2001), illustrates the general locations and paths of the Kuroshio and Oyashio Currents.
http://i51.tinypic.com/15zs014.jpg
Figure 1
As noted above, the Kuroshio and Oyashio Currents collide East of Japan and form the western portion of the North Pacific Current. These waters are often referred to as the Kuroshio-Oyashio Extension or the KOE. For the purpose of this post, I’ve used the coordinates of 30N-45N, 150E-150W for the Kuroshio-Oyashio Extension, Figure 2.
http://i52.tinypic.com/14twvox.jpg
Figure 2
CORRELATION WITH NORTHERN HEMISPHERE TEMPERATURES
Sea Surface Temperature (SST) anomalies for much of the North Atlantic warm (cool) when the Kuroshio-Oyashio Extension SST anomalies warm (cool). This can be seen in the correlation map of annual (January to December) Kuroshio-Oyashio Extension SST anomalies and annual North Atlantic SST anomalies, Figure 3.
http://i52.tinypic.com/fjj23r.jpg
Figure 3
And, as shown in Figures 4 (RSS) and 5 (UAH), annual TLT anomalies for much of the Northern Hemisphere correlate with the annual SST anomalies of the Kuroshio-Oyashio Extension.
http://i52.tinypic.com/311s49i.jpg
Figure 4
##############
http://i53.tinypic.com/2qsx7j8.jpg
Figure 5
The same thing holds true for combined land plus sea surface temperature datasets such as the GISS Land-Ocean Temperature Index (LOTI) data for the Northern Hemisphere, Figure 6. Much of the Northern Hemisphere GISS LOTI data warms (cools) as KOE SST anomalies warm (cool). (Also note the differences in the North Atlantic correlations in Figures 3 and 6. They’re based on the same SST dataset, so why are there differences? GISS deletes SST data from areas with seasonal sea ice and extends land surface data out over the oceans with its 1200km radius smoothing. Refer to GISS Deletes Arctic And Southern Ocean Sea Surface Temperature Data.)
http://i54.tinypic.com/303llxg.jpg
Figure 6
WHEN DOES THE KOE WARM?
As we’ve seen in past posts, the East Indian and West Pacific Oceans warm in response to El Niño events and then during the subsequent La Nina events. As part of the East Indian-West Pacific subset, the Kuroshio-Oyashio Extension warms significantly during La Niña events. Animation 1 is taken from the videos in the post La Niña Is Not The Opposite Of El Niño – The Videos. It presents the 1997/98 El Niño followed by the 1998 through 2001 La Niña. Each map represents the average SST anomalies for a 12-month period and is followed by the next 12-month period in sequence. Using 12-month averages eliminates the seasonal and weather noise. The effect is similar to smoothing data in a time-series graph with a 12-month running-average filter. Note how the Kuroshio-Oyashio Extension warms significantly during the La Niña event and how the warming persists for the entire term of the La Niña.
http://i53.tinypic.com/etb58j.jpg
Animation 1
Note in Animation 1 that the SST anomalies of the Kuroshio-Oyashio Extension were cool during the 1997/98 El Niño. The KOE actually started with depressed SST anomalies, and they did not drop significantly during the 1997/98 El Niño. Refer to Figure 7. On the other hand, the KOE SST anomalies did rise significantly during the transition from the El Niño to the La Niña in 1998. The other major El Niño event that wasn’t impacted by the aerosols of an explosive volcanic eruption was the 1986/87/88 event. The SST anomalies of the Kuroshio-Oyashio Extension cooled during the 1986/87/88 El Niño, but also rose significantly during the 1988/89 La Nina. We’ll take a closer look at that event later in the post.
http://i53.tinypic.com/2qa1onl.jpg
Figure 7
This response of the Kuroshio-Oyashio Extension to El Niño and La Niña events is easier to see if the NINO3.4 SST anomalies are inverted, Figure 8. That is, the Kuroshio-Oyashio Extension warms much more during the 1998/99/00/01 La Niña event than it cools during the 1997/98 El Niño. But could the significant drop in the Kuroshio-Oyashio Extension during the 1986/87/88 El Niño impact the global response to that El Niño? Again, we’ll examine that later in the post.
http://i52.tinypic.com/wjvow.jpg
Figure 8
WHY DOES THE KOE WARM DURING LA NIÑA EVENTS?
Let’s start with the El Niño. During an El Niño event, a significant volume of warm water from the west Pacific Warm Pool travels east to the central and eastern equatorial Pacific, where it releases heat primarily through evaporation. And most of the warm water from the Pacific Warm Pool water comes from below the surface. There is “leftover” warm water when the La Niña forms, and a portion of this leftover warm water is returned to the western tropical Pacific at approximately 10 deg N latitude. Video 1 illustrates global Sea Level Residuals from January 1998 to June 2001. It captures the 1998/99/00/01 La Niña in its entirety. The video was taken from the JPL video “tpglobal.mpeg”. The phenomenon shown carrying warm waters from east to west in the tropical Pacific at approximately 10 deg N is called a slow-moving Rossby Wave.
Video 1
Link to Video 1:
http://www.youtube.com/watch?v=MF5vZErQ6HM
Unfortunately, the video “tpglobal.mpeg” is no longer available at the JPL VIDEOS web page, but for those who would like to watch the entire video, I uploaded it to YouTube as Sea Surface Height Animation 1992 to 2002 – JPL Video tpglobal.mpg.
In Video 1, the warm “leftover” warm water from the 1997/98 El Niño is clearly carried as far west as the Philippines. Shortly thereafter Kuroshio-Oyashio Extension sea level residuals rise and remain elevated for the duration of the La Niña.
In addition, there are other factors that add to and maintain the elevated SST anomalies in the Kuroshio-Oyashio Extension during the La Niña. As shown in Animation 1 (the gif animation, not the video), Sea Surface Temperature anomalies outside of the tropical Pacific rise in response to the El Niño. The changes occur first in the Atlantic, then Indian, and finally the west Pacific. Sea Surface Temperature anomalies rise as changes in atmospheric circulation caused by the El Niño make their way eastward around the globe to the western Pacific. Then, during the La Niña, the opposite occurs for much of the globe. But in the tropical Pacific, the trade winds strengthen and the North and South Equatorial Currents return warm “leftover” surface waters from the El Niño to the west. So the western Pacific is warmed cumulatively by the El Niño and then by the La Niña. In the northwest Pacific, the Kuroshio Current carries the leftover warm water up to the Kuroshio-Oyashio Extension.
Additionally, the increased strength of the trade winds during the La Niña also reduces cloud cover over the tropical Pacific, which increases the amount of Downward Shortwave Radiation (visible light) there. The increased Downward Shortwave Radiation warms the tropical Pacific. The warmed water is carried to the west by the Equatorial Currents and the North Pacific Gyre spins the warmed water up to the Kuroshio-Oyashio Extension.
WHY IS THIS IMPORTANT?
In the post “RSS MSU TLT Time-Latitude Plots…Show Climate Responses That Cannot Be Easily Illustrated With Time-Series Graphs Alone”, I illustrated that the RSS Lower Troposphere Temperature (TLT) anomalies of Southern Hemisphere and of the Tropics (70S-20N) followed the basic variations in NINO3.4 SST anomalies, Figure 9. This is how one would expect TLT anomalies to respond to El Niño and La Niña events. El Niño events cause the TLT anomalies to rise because they release more heat than normal to the atmosphere, and La Niña events cause TLT anomalies to fall because the tropical Pacific is releasing less heat than normal.
http://i54.tinypic.com/r9h0d5.jpg
Figure 9
But the TLT anomalies of the Northern Hemisphere north of 20N, Figure 10, appear to rise in a step after the 1997/98 El Niño. That is, there is very little response to the 1998 through 2001 La Niña. It appears as though a secondary source of heat is maintaining the Northern Hemisphere TLT anomalies at elevated levels.
http://i53.tinypic.com/11lsb6e.jpg
Figure 10
A similar upward step can be seen in the GISS Land-Ocean Temperature anomaly index (LOTI) for the latitudes of 20N-65N, Figure 11. (North of 65N the GISS data is biased by their deleting Sea Surface Temperature data and replacing it with land surface data with a higher trend. Again, refer to GISS Deletes Arctic And Southern Ocean Sea Surface Temperature Data.)
http://i53.tinypic.com/34qr5t2.jpg
Figure 11
And a similar upward step is visible in the North Atlantic SST anomaly data, Figure 12.
http://i56.tinypic.com/1zewmqq.jpg
Figure 12
The North Atlantic SST anomalies, the Lower Troposphere Temperature( TLT) anomalies of the Northern Hemisphere north of 20N, and the Northern Hemisphere Land-Ocean Temperature anomalies (20N-65N) all rise in response to the 1997/98 El Niño, but fail to respond fully to the 1998/99/00/01 La Niña. The similarity of the curves can be seen in Figure 13.
http://i54.tinypic.com/200v0j5.jpg
Figure 13
The correlation maps in Figures 3 through 6 show that a portion of the warming of the Northern Hemisphere north of 20N should be a response to the elevated Kuroshio-Oyashio SST anomalies during the 1998 through 2001 La Niña. To further illustrate this relationship, Figure 14 compares the KOE SST anomalies (not scaled) to the three datasets shown in Figure 13. I did not scale the Kuroshio-Oyashio SST anomalies because I wanted to illustrate the differences in the magnitudes of the variations. The variations in Kuroshio-Oyashio SST anomalies are clearly far greater than the variations of the other three datasets in Figure 14. In fact, the KOE SST anomaly variations are about 40% to 50% of the variations in NINO3.4 SST anomalies (refer back to Figures 7 and 8).
http://i56.tinypic.com/29e0pvp.jpg
Figure 14
Figure 15 presents the same datasets as Figure 14, but in Figure 15, the Kuroshio-Oyashio Extension SST anomalies have been scaled. Keep in mind that the three Northern Hemisphere temperature anomaly datasets rise first in response to the El Niño.
http://i54.tinypic.com/25hl2tz.jpg
Figure 15
It appears the warming of the Kuroshio-Oyashio Extension during the 1998/99/00/01 La Niña and its interaction with the other datasets could explain a portion of the trend in Northern Hemisphere SST anomalies, TLT anomalies, and Land-Ocean temperature anomalies since 1995. The warming of the Kuroshio-Oyashio Extension during that La Niña counteracts the normal cooling effects of the La Niña and prevents the temperature anomalies for the three datasets shown in Figures 13, 14, and 15 from responding fully to the La Niña.
THE 1986/87/88 EL NIÑO & 1988/89 LA NIÑA
There is a similar effect during the 1988/89 La Niña. That is, Northern Hemisphere temperature anomalies do not drop as one would expect during a La Niña. But the response during the 1986/87/88 El Niño may help to confirm the impact of the Kuroshio-Oyashio Extension on Northern Hemisphere temperatures.
Figure 16 compares scaled NINO3.4 SST anomalies for the period of 1985 through 1994 to the same datasets used in Figures 13: North Atlantic SST anomalies, the Lower Troposphere Temperature (TLT) anomalies of the Northern Hemisphere north of 20N, and the GISS Northern Hemisphere Land-Ocean Temperature anomalies (20N-65N). Once again, the Northern Hemisphere datasets rise in response to the El Niño event, but don’t drop in response to the La Niña. Note also that the North Atlantic SST anomalies lag the NINO3.4 SST by more than 6 months during the ramp-up phase, but the lag in the Northern Hemisphere TLT and Surface Temperature datasets is excessive, about 18 months. Why?
http://i53.tinypic.com/iqx3te.jpg
Figure 16
Could the dip in the Kuroshio-Oyashio Extension SST anomalies during the 1986/87/88 El Niño have counteracted their responses to the El Niño? Refer to Figure 17. It compares Kuroshio-Oyashio Extension SST anomalies (not scaled) to the North Atlantic and Northern Hemisphere datasets. The drop in KOE SST anomalies is significant in 1986/87/88.
http://i51.tinypic.com/2cwjs6c.jpg
Figure 17
And in Figure 18, the Kiroshio-Oyashio SST anomalies have been scaled. The North Atlantic SST anomalies rise in response to the 1986/87/88 El Niño as noted earlier. The timing of the rises in the KOE data and the GISS LOTI data are very similar. But the rise in the TLT anomalies north of 20N precedes the rise in the KOE data. If the dip in KOE SST anomalies were the only factor preventing the TLT anomalies from rising in response to the El Niño, shouldn’t we expect the TLT anomalies to lag the rise in the KOE data? Or are the TLT anomalies responding to the rise in North Atlantic SST anomalies?
http://i52.tinypic.com/2r5xdl3.jpg
Figure 18
If we replace the RSS TLT data with TLT data from UAH, Figure 19, the lag decreases between the North Atlantic SST anomalies and the TLT anomalies north of 20N.
http://i52.tinypic.com/2wqbui9.jpg
Figure 19
CLOSING
An El Niño event releases vast amounts of warm water from below the surface of the west Pacific Warm Pool. But the end of an El Niño event does not mean all of that warm water suddenly disappears. The warm water sloshes back to the western tropical Pacific during the La Niña. And some of that warm water is spun up into the Kuoshio-Oyashio Extension where it continues to release heat.
Kuroshio-Oyashio Extension SST anomalies rose significantly during the La Niña events of 1988/89 and 1998/99/00/01. These warmings appear to have counteracted the effects of those La Niña events on North Atlantic SST anomalies, and on Lower Troposphere Temperature anomalies north of 20N, and on combined Land-Ocean temperature anomalies of the Northern Hemisphere between the latitudes of 20N-65N. During the 1997/98 El Niño, the drop in Kuroshio-Oyashio Extension SST anomalies was very small and the KOE does not appear to have had a noticeable impact on the effects of that El Niño. On the other hand, the Kuroshio-Oyashio Extension SST anomalies did drop significantly during the 1986/87/88 El Niño and they appear to have suppressed the effects of that El Niño on Northern Hemisphere temperature anomalies. But why did the Kuroshio-Oyashio Extension SST anomalies drop significantly during the 1986/87/88 El Niño but not during the 1997/98 El Niño? Differences in Sea Level Pressure?
SOURCE
Data for graphs are available through, and the correlation and anomaly maps were downloaded from, the KNMI Climate Explorer:
http://climexp.knmi.nl/selectfield_obs.cgi?someone@somewhere
Posted by Bob Tisdale at 6:39 AM
Sure, the ocean heat content doesn’t make a mark on the solar activity level. I don’t know what term to use really. ‘Bellweather’ seems a bit quaint. I think that your point is a minor quibble rather than a fatal flaw in my hypothesis though. The main point is that sunspot area and number are a proxy for solar activity level and it’s effect on ocean heat content, which is amplified by the amount calculated by Nir Shaviv in his JGR paper ‘Using the oceans as a calorimeter’.
A jargon note for P. Solar:
A running mean has a growing window anchored at the left end of a time series whereas the window of a moving average has constant width and floats with the window center. Misuse of the term “running mean”, as if it were synonymous with “moving average”, is widespread, so often one has to infer from context what is meant.
P. Solar says: “Bob, if you want to see how it compares to the original data why don’t you just refer to the graph I posted. The one you link under it is clearly a different dataset.”
Because this graph…
http://i55.tinypic.com/ftqbf6.jpg
…is not raw data. And this graph…
http://i56.tinypic.com/2958yg1.png
…does not compare TLT data to NINO3.4 SST anomalies, which is the topic of discussion.
You wrote, “Again better referencing of your sources would be useful.”
The TLT data source (UAH) is shown on my graph. Is it shown on yours or mentioned in your discussion? The NINO3.4 SST anomalies are HADISST. What NINO3.4 SST anomaly dataset did you use? HADSST2? ONI? ERSST.v3b? Kaplan? It’s not Reynolds OI.v2 because it starts too early.
And here’s the same graph that I posted above, but I replaced the UAH data with RSS TLT anomalies. My same comment applies to it as well:
http://i55.tinypic.com/2mocxp1.jpg
The closing sentences of Bob’s article: “But why did the Kuroshio-Oyashio Extension SST anomalies drop significantly during the 1986/87/88 El Niño but not during the 1997/98 El Niño? Differences in Sea Level Pressure?”
Ulric Lyons has offered a suggestion (“Different winters.”)
Interannual NPI was phased a little differently relative to SOI & PWP for the 2 events (including winter), but I wonder if Ulric is thinking more of AO/NAO than NPI? Some clarification might help. I do note a phase reversal in interannual solar-terrestrial coupling (using either aa or SW) beginning about the time of the ’98 El Nino. Perhaps Ulric is hinting at that?
A simpler possibility?
http://i29.tinypic.com/rwp8ut.png
Further to my last comment, I think this is what was on Bob’s mind when he proposed SLP:
http://i47.tinypic.com/29uvpe9.png
Bob, in pondering the answer to your question, you might want to dig out Thompson et al.’s (2009) COWL (Cold Oceans Warm Land) article. There are some parallels. I wonder if they are aware of the wave & mode switching that goes on between AO/NAO, NPI, & AAO/SAM. If they are not, this could be substantial.
Just a quick question from a layman with a limited understanding of the mechanics of ENSO events.
The sub-surface in the western pacific is now showing signs of warming bringing with it the prospect of a return to neutral or possibly even El-Nino conditions next year.
This made me wonder whether the cooler SST’s making their way East to West from the current La-Nina could eventually downwell in the Western Pacific preventing a transition back to El-Nino and thus the potential for back-to-back La Nina events?
What follows from this is what impact back to back La-Ninas might have on the KOE.
Bob, this is a note to confirm that T_ENSO & COWL are not nonlinearly independent. COWL is a mash-up of AO, NAO, NPI, plus a bunch of other NAM stuff and so it fits the framework outlined by Schwing+ (2003). Perhaps Thompson+ (2009) only checked for linear orthogonality without so much as a glance at the complex plane. Piers Corbyn must be LHFAO. You could potentially write the blogging article of your blogging career if you look into this carefully, but you’ll have to switch to using repeat 1 year smoothing because wide boxcars simply cannot render the patterns discernible for a lay audience. Best Regards.
Bob Tisdale:
>>
http://i56.tinypic.com/2958yg1.png>
…does not compare TLT data to NINO3.4 SST anomalies, which is the topic of discussion.
>>
No, it shows what I said it shows: that when you compare the two filters to the raw data you see that your use of running mean incorrectly attenuates the 1983 peak, and not by a small amount.
You stated that the gaussian was not a better filter because it “failed to capture the El Chincon” effects.
You erroneously concluded the rm to be better because you “expected” to see the second peak smaller. You did not check to see if this was actually present in the data and now I point it out to you, you seem to steadfastly refuse to accept the point.
You are throwing sand in the air linking other graphs and arguing about datasets. The one graph proves the point. The rest is a distraction.
So having established that your data processing is corrupting the very peaks you are trying to compare, I suggest it may be better using a different filter.
You are free to ignore that suggestion but do not try to make out I am incorrect.
David W: Figure 7, above, compares KOE and NINO3.4 SST anomalies since 1995. The data have been smoothed in it, and because of the smoothing, we lose the fact that there was a relaxation of La Nina conditions after the 2007/08 La Nina. That is, if we look at the same data without the smoothing…
http://i52.tinypic.com/f9qck1.jpg
…we can see that NINO3.4 SST anomalies actually rose to “zero” in the first half of 2008 before taking another swing down. The drop in NINO3.4 SST anomalies during the second half of 2008 didn’t register as an “official” La Nina, but KOE SST anomalies remained elevated. In fact, they didn’t drop until the the switch to the 2009/10 El Nino.
Paul Vaughan: Thanks for your research and insights into COWL. It sounds like you were examining the Thompson et al data. Did you notice how their “ENSO fit” data was biased up in early years compared to raw HADSST2 CTI data?
http://i38.tinypic.com/10rikb4.png
Meaning when they subtract the “ENSO fit” data from their global temp data, it will add to the global trend. I showed a comparison in my post on Thompson et al but no one picked up on it. It’s a small bias but it exists.
http://bobtisdale.blogspot.com/2009/09/thompson-et-al-2009-high-tech-wiggle.html
Since Thompson is using the same process in a 2010 paper to show how well climate models reproduce the 20th century temperature record, I may have to revisit that data. That 2010 paper is not worth a separate post, but I’m sure I can wiggle it into another as a “Why is this important?” side note.
Regards
“The data have been smoothed in it, and because of the smoothing, we lose the fact that there was a relaxation of La Nina conditions after the 2007/08 La Nina. ”
Yes this is a known issue with some smoothing techniques but not all. nino3.4 2008 peak-a-boo
P. Solar says: “You stated that the gaussian was not a better filter because it ‘failed to capture the El Chincon’ effects,” and, “You erroneously concluded the rm to be better because you ‘expected’ to see the second peak smaller.”
Because it is smaller, which was why I presented the two graphs that compared “raw” scaled NINO3.4 SST anomalies to global TLT anomalies in these two graphs:
http://i55.tinypic.com/152m6mc.jpg
And:
http://i55.tinypic.com/2mocxp1.jpg
You added, “You are throwing sand in the air linking other graphs and arguing about datasets. The one graph proves the point. The rest is a distraction.”
The reason I asked you which NINO3.4 dataset you used was so that I could use exactly the same data and present to you that the El Chichon eruption had in fact suppressed the global TLT anomalies, unlike your presentation of it. And you still haven’t identified the NINO3.4 data you used.
P. Solar says: “Yes this is a known issue with some smoothing techniques but not all. nino3.4 2008 peak-a-boo”
That’s right. The gaussian filter leaves in the seasonal component, while the running mean minimzes it.
Regards
Bob, yes I noticed (by comparing integrals of T_ENSO & CTI — makes it crystal clear). They’ve subtracted a global summary from CTI. (According to their reasoning this eliminates a bias-step in the CTI record.)
There are much bigger issues with the work.
The concept “COWL” was a valuable contribution and the authors are clearly very intelligent, but I cannot turn a blind eye to their foolhardy promotion of absolutely untenable assumptions. That the coupling is nonlinear is undeniable (and yet their decompositions are linear). The authors should at least acknowledge that multiscale hypercomplex factor analysis is necessary in such a context.
–
I sincerely hope readers see the connection between KOE & COWL. [If not, please see my notes above in which I link to Carvalho+ (2007).]
A key in identifying relatively elusive spatiotemporal signals is to pin down spatial modes that are somewhat stationary due to the lay of topography relative to dominant flows. Such a focal point exists in the North American Pacific Northwest (Northeast Pacific Ocean). Anyone understanding Schwing+ (2003) will be better positioned to make sense of the mid-20th-century discontinuity in the following article:
Courtillot, V.; LeMouel, J.L.; Blanter, E.; & Shnirman, M. (2010). Evolution of seasonal temperature disturbances and solar forcing in the US North Pacific. Journal of Atmospheric and Solar-Terrestrial Physics 72, 83-89.
Cautionary Note: Don’t uncritically accept the single ~1958 date given by Schwing+ (2003) without pursuing more detailed data exploration. (There is multivariate switching on a number of spatiotemporal scales. Seasons & regions are not irrelevant due to factors such as north-south asymmetry.)
–
Thanks Bob, Anthony, & Moderators.
Bob, the nino3.4 is the same one you indicated to me in to other thread so that should match what you use. I even posted a snippet of numbers and you confirmed it was the right data, so I don’t think there should be any confusion there.
“That’s right. The gaussian filter leaves in the seasonal component, while the running mean minimzes it.”
If you run a 13 month filter you will not get a cut off at a frequency of one year whatever filter you choose. For memory have a look at the two profiles in the article I linked before:
http://homepages.inf.ed.ac.uk/rbf/HIPR2/gsmooth.htm
13 months is the window not the cut-off and in the case of running mean you will get a load of harmonics still getting through as well. (see summer 1999)
It is true that the 13m rm is reducing the amplitude more than 13m gaussian but that does not in itself mean it is better filtering the data. I have already noted drawn attention to the small quantity of monthly noise that can be seen in parts of the rm output. You will see that G does not have that even thought the cut-off is lower. It is a very smooth curve, no h.f. component visible. Also note the way some rm peaks are skewed left or right when surrounding data is higher one side than the other (summer 2002).
In the last comparative plot I posted you can see the winter trough of 2001 and 2004 actually ends up as a slight peak in the rm, so it is not that it is taking out this feature better it’s actually causing a spurious increase as a result of the peaks either side.
Again, referring to the article, you will see that G freq response cuts off earlier than rm so you may want to compare to a wider window if you need to eliminate the annual signal.
Play with the script I posted above keeping gw=3 sigma but increase the value of sigma. A value of three gives amplitudes close to your rm without the defects I noted.
If you want to eliminate the annual component you need a heavier filter. The following uses w=12 (ie 25m window) and sigma=8 in the script I posted.
Here it is even clearer that rm is letting stuff through, skewing peaks and even inverting some of the smaller peaks.
This plot actually shows your correlation quite well.
regards.
P. Solar, the filters aren’t “inverting” peaks. As I explained in an earlier thread, that’s not a “problem” with the filter, but rather with interpretation. To get around the centering issue, a 12 month boxcar (1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) can be adjusted to a 13 month kernel with weights 0.5, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0.5. It makes sense to design a kernel that is mindful of dominant temporal modes of variation — in this case the terrestrial year. Your gaussian kernel might be a great option in contexts where no dominant periods exist, but it complicates interpretation of climate data. In practice “anomalies” do not have the properties that they have in theory. Repeat application of a 1 year smoother works well for many climate time series, but for some, 6 month smoothers make sense (e.g. for some equatorial series). For a time series like SOI, try 3 month, then 6 month, then 12 month. (You will note that repeat 3 month smoothing has the peak-&-trough location-preserving properties that you seem to be advocating. Repetition has the effect of creating a triangle rather than a bell.)
Thanks for you enlightened comments Paul.
It was careless of me to say inverting peaks. What I meant was showing a trough/peak where the original data shows a peak/trough. The effect, as I am sure you are aware is dependent on the relation of surrounding peaks and the width of the kernel (window) used.
You will note that the adjusted rm you suggest was included in the script I posted and documented with a comment. Since that modified kernel puts less weight on the outlying points (the root of all these issues in rm ) it does “help”. It does not really get around the problem. It is just one very crude concession to the idea of using a weighted mean. Gaussian is one case of a thoughtfully designed weighted mean with a clean monotonic frequency response.
Repeating a narrower window is often a good option if, like here, the window width is interacting with prominent features of the dataset. It also helps reduce the h.f. components getting through.
>> For a time series like SOI, try 3 month, then 6 month, then 12 month.
Hmm, it may “work well” but it would be less obvious to say what you had actually done to the data. Back of envelop sketching suggests it may have a frequency response a bit like a wobbly bell shape. It would be interesting to see calculated plot of its response.
>> Your gaussian kernel might be a great option in contexts where no dominant periods exist, but it complicates interpretation of climate data.
I’m not promoting gaussian as the holy grail of filters , it seemed better than box car running mean in this case and helped show some of the distortions produces by rm.
ANY filter complicates interpretation if it is used without thought and a fair degree of understanding. That is what I was getting at. I don’t see that the results of gaussian are more complicated to interpret than a box-car or repeated running mean. In fact the simpler form of the frequency response should make it more predictable.
Having identified a correlation Bob is looking at time lags. In the example we have been discussing here, the lag is different before and after 1998. In investigating what this means it seems fairly important that the filter is not bending and shifting the primary peaks as the 13m box-car is doing.
The results are not hugely different but removing some of the obvious defects should help a more precise analysis of the lag between the various cycles.
The bottom line, as you are well aware, is that this is not a trivial subject and awareness of all these issues is important in interpreting “features” we perceive in filtered data to be sure that they are real and not artifacts of the filtering process or being corrupted and distorted by it.
regards.
P. Solar, I can agree that nonsensical notions about smoothing & lags abound in all forums devoted to the discussion of climate variations.
–
P. Solar wrote, “Repeating a narrower window is often a good option if, like here, the window width is interacting with prominent features of the dataset.”
Good to see this enlightened comment. Certainly the mainstream has gotten caught up in the notion that anomalies have some ideal properties which they absolutely do not have in practice.
–
My attention to this thread has expired. Thanks to all who have participated. I look forward to Bob’s future posts.
Just for the record , in case anyone is interested, here is the frequency response of the combined 3 month, 6 month, 12 month running mean filter. (unscaled x-axis)
triple-sync
So my intuitive guess was correct, it is quite similar to the form a gaussian plus a small amount of ripple in the stop band.
Running this filter in three passes probably allows use of a narrower total window for a result similar to the gaussian. This may be an advantage where one wishes to run as near to the end of the data as possible (eg. as will global mean temps perhaps) a the cost of having run three filters instead of one.
Man, my gut feeling was a lot close than I realised.
Here is the gaussian overlayed with the tripled up running mean Paul suggested above. It is uncannly close to the gaussian apart from the ripple.
triple-sync vs gaussian
Paul Vaughan says:
>>It makes sense to design a kernel that is mindful of dominant temporal modes of variation — in this case the terrestrial year. Your gaussian kernel might be a great option in contexts where no dominant periods exist, but it complicates interpretation of climate data.
>>
So in conclusion to this discussion on filters, I agree, it makes sense to design a kernel. That seems to be a point that is almost universally ignored. Not just by amateurs but by major climate data organisations and university professors !
If you want to remove a 12 month signal you don’t start with a 12 month (or 13 month) window and a crappy filter which lets through significant amounts of what you imagine you have filtered out and moves peaks around and replaces some smaller peaks with troughs.
Designing a filter means at least considering what the frequency response looks like and deciding if it is suitable. This you do _before_ looking at the data that comes out the other end, deciding it looks like what you’d expect and not going any further in determining what you are doing to the data.
It’s interesting that you suggest my gaussian is not suitable for climate data yet propose a triple running mean that has a very similar response except for retaining some of the defects of a simple running mean.
You don’t explain how the gaussian is not suitable or how it “complicates” interpretation.
The only down side I see to G is that it requires a wider window, so data cannot be filtered quite so close to the start and end . If this is a problem the triple rm may be a good second best since the defects in the filter are not huge.
Interestingly enough the KNMI service where Bob sources much of his data used to use gaussian until they noticed it was exaggerating the post 2000 cooling. (Never got noticed while it was exaggerating the 1990s warming !) . In fact this was not due to the gaussian filter itself but was due to the fact that they were using a partially filled window at the end of the data.
Clearly this has no mathematical validity at all and is such a stupid thing for a professional body to do it merely underlines how little thought or understanding goes into much of what is presented a climate science even from such august bodies.
The other problem with the proposed 3,6,12 month running mean is that only the first can be done without a time shift. Introducing two shifts at different stages of the multipass filter will have some odd effects on how it distorts the data. The result will not be as clean or simple as the comparison I posted.
Also the start and end sections that don’t have a complete window will be additive in multiple runs so the only advantage it has over gaussian is largely lost.
If Paul Vaughan is using that sort of technique I can see why he has doubts (expressed elsewhere) about the use of looking at phase lags in climate data.
In conclusion, the idea that people have that running mean is easy to understand is only based on their lack of understanding. Though it is simple to implement, understanding the resulting effects is far from simple and not good when you do understand.
Sticking with a bad filter because other, less well informed people think they understand it, seems like a pretty bad way to go.
RUNNING MEAN MUST DIE.
Hadcru and NCDC are in. NCDC and GISS are in good agreement. It’s obvious that Hadcru is the “odd man out” this november and this year, not GISS: http://img684.imageshack.us/img684/1617/gissncdchadcruthadcrut3.jpg