The other day we discussed missing data in the Northern Indian Ocean, where the main best tracks archive showed storm track data up to the mid-1970s but lacked wind speed estimates, with a sharp decline in storm track occurrences in the 1970s.
In his comment on Webster, Curry et al, William Gray observed that there were data problems in the Southern Indian Ocean and South Pacific as well. Webster, Curry et al acknowledged the problem in respect to the North Indian Ocean in their Reply to Gray, available on the internet but never published in a journal, but denied that this was material since North Indian Ocean storms were only a small fraction of total storm activity.
Today I want to look at the situation in the Southern Hemisphere, where the majority of storms occur in the South Indian Ocean.
When I collated SH hurricane data, I was surprised to find that, although all other data sets used by Webster, Curry et al matched the data sets at unisys, this was not the case for the SH. Here the Webster Curry data set was substantially different between 1970 (the start of their data as used) and 1983, when the two data sets began to substantially match. Figure 1 below shows that the best tracks data had a substantial number of track location measurements without wind speed estimates prior to 1983 (as we observed in the Northern Indian Ocean) , while the Georgia Tech data had virtually no such examples.
Figure 1. Number of measurements with storm track measurements but no wind speed estimates. Orange – SIO and SPac; green – NIO; red- Georgia Tech SIO-SPac combined, showing virtually no such examples.
In their website (but not in the body of the original article), Webster, Curry et al report the following:
The southern hemisphere track data of 1970 – 2002 are provided by Charlie Neumann and are made up of a reanalysis of available data from Australia, Fji and JTW.
I have been unable so far to locate any publication describing Neumann’s reanalysis. However, whatever the merits of this particular reanalysis, one can reasonably say that Neumann wind speed estimates are not a homogeneous method to the post-1983 SH record. Now on to the SH information itself.
Figure 2 is an excerpt from Webster Curry Figure 3, showing the number of SH hurricanes and number of SH hurricane days in blue on the left. The coloring in the right panel is inconsistent as both unsmoothed SH and WPac are in black, but the SH annual measurements are lower and are tracked by the heavy blue smooth.
Figure 2. Excerpt from Webster Curry et al 2005 Figure 3.
Next, Figure 3 shows the number of actual SH measurements in the two archives divided by 4, bringing the series forward from 2004 to 2006-to-date (but Nov-Dec in SH are small months). Black shows the information archived at Unisys; red dashed shows the information used by Georgia Tech, demonstrating from a slightly different viewpoint the number of Neumann measurements included in the Georgia Tech data set.
It is also interesting to note the decline from 2004 to 2006, where the number of measurements is at rather low levels. If you squint, you can perhaps discern the pattern of the red dashed line of number of measurements (divided by 4) in the number of SH hurricane-days in the right panel of the Webster Curry figure. There is a slight difference in level owing presumably to some difference in cut-off point.
Webster Curry et al observed that there was no increase in number of hurricane-days, but perhaps could have drawn more attention to the decline in hurricane-days from the peak in the 1990s to 2004 (which has continued to 2006)
Figure 3. Number of archived SH wind speed measurements. Black – Unisys archive; red dashed- Georgia Tech version; blue – update to 2006 collated from Unisys data. I’ve collated individual storms for 2003-2006, interpolated them to 6-hour intervals and used this information to update the Best Tracks database, using this collation from 2003-2006 for both data sets.
The Interesting Part
Obviously the Neumann estimates are constructed differently than post-1983 measurements, but how can one evaluate the amount of potential non-homogeneity in the Neumann estimates.
As an experiment, I calculated the total number of wind speed measurements (as illustrated in Figure 3) plus the number of track measurements without wind speeds (as illustrated in Figure 1) – hypothesizing that the track locations without wind speed estimates would nearly all have had winds meeting storm benchmarks – a hypothesis which seems quite plausible to me based on admittedly limited experience with this data.
Figure 4 shows that there is now a very big difference between the two data sets. In the one case, there is a strong decline from levels in the 1970s with new lower levels in the 1980s; in the other (Georgia Tech) case, the levels remain fairly steady. In addition to the combined storm track information, I’ve also shown plots of measured hurricane-level measurements (everything divided by 4 to match "days") and in purple the number of Cat4/5 measurements ( divided by 4).
Figure 4. Count of measurements plus track location without wind speed measurement. Number of measurements meeting hurricane levels (blue) and Cat4/5 levels (purple). Total number of measurements divided by 4 to match number of "days". Left – Unisys archive; right – Georgia Tech.
There is either a real decline in number of storm-days or an inhomogeneity in the number of measured storm track locations. To establish that the Neumann estimates are valid would require a demonstration that there has been a non-homogeneity in decisions to record storm track locations. This is quite possible.
However, the onus is on Webster, Curry et al to demonstrate the existence of this non-homogeneity. The Neumann estimates are undoubtedly meritorious, but Webster Curry et al provided no discusison of how these estimates were made. And whatever else one may think of Gray, he has expressed caveats on how much reliance can be placed on pre-1985 estimates – caveats that were rather airily dismissed by Webster, Curry et al – but I felt that their Reply did not deal directly with some important issues raised by Gray.
As a closing illustration, here are, first, histograms of all wind speed measurements, separated at 1983 and below that corresponding histograms of all wind speed measurements meeting storm cut-offs. One sees that the distributions are somewhat different with meaurements in the 30-40 knot class less frequent after 1983, with increased proportions in bins with greater wind speed.
Figure 5. Histogram of All Wind Speed Measurements(in knots). Left – up to 1983; right- 1984 on.
Figure 6. Same as Figure 5, but limited to wind speeds greater than or equal to 35 knots (18 m sec-1)
The statistical problem faced by Webster, Curry et al is to show that this (IMHO) fairly subtle change in distribution is (a) a real change in distribution rather than a stochastic effect; (b) not a product of the inhomogeneity between Neumann estimates to best tracks estimates (however these were done.) If you re-read Webster Curry et al, I think that it would be fair to say that their statistical analysis was unequal to either challenge.
I’ll try to re-visit this issue on some later occasion.