Dr. Bill Johnston.
(Former NSW natural resources research scientist.)
Once a staging post on the track to Oberon and Bathurst, the delightful little village of Taralga is 44 km (27 miles) north of Goulburn New South Wales and 135 km (84 miles) from Canberra. The post office (Figure 1) is in the news recently because it is one of several weather stations where the Bureau of Meteorology were caught-out deleting low minimum temperature (Tmin) values for “quality control” reasons. Stories have been published in The Australian newspaper about deletions at Goulburn airport and Thredbo; and there may be others.
Figure 1. The Taralga post office (left) and Court House in 1910 (National Archives of Australia (NAA)).
Because it is used to homogenise a 1964 Tmin time of observation change at Canberra airport, Nowra RAN, Richmond RAAF and Sydney Observatory, which are ACORN sites used to calculate Australia’s warming (Australian Climate Observations Reference Network – Surface Air Temperature); its worth sleuthing Taralga’s data, especially since the site’s metadata is sparse and incomplete (the earliest site-plan is 1998 and there is no mention of when the current 0.06 m3 small screen replaced the former 0.23 m3 large one).
Faulty ACORN data won’t be properly adjusted using other faulty data.
Analysis of daily data available from 1957 shows average annual maxima (Tmax) stepped-up a hefty 0.96oC in 1965 and 0.95oC in 2004 (0.71oC and 0.91oC rainfall adjusted) (Figure 2). Two step-changes result in three data segments whose relationship with rainfall is linear and statistically parallel. Accounting for step-changes and rainfall explains 60.2% of Tmax variation and although residuals are variable in the time-domain (due to synoptic, site and missing-data effects) there is no additional hidden trend suggestive of climate warming. (Years having appreciable missing data (1970 to 1986) don’t make much difference overall so are treated as valid data.)
Local evaporation, which removes heat as latent heat (2.45 MJ/kg of water evaporated) can’t exceed the rainfall (mm=kg/m2); thus provided the yard is not watered, a dependent robust negative relationship is expected between Tmax and rainfall. This is confirmed statistically: rainfall reduces annual average Tmax by 0.21oC/100 mm of annual rainfall.
A Tmin step-change in 1973 (0.47oC) aligns with metrication and its magnitude may be affected by missing data (Figure 2). Perhaps the Fahrenheit thermometer was faulty; protocols were also tightened-up and many sites were visited and made more compliant (as there is not much point in changing the thermometer if sites are in a poor state, the screen may have been repaired, painted or moved to improve exposure).
Rainy years are cloudy, which reduces outgoing nighttime long-wave emissions causing Tmin to be warmer. However, as cloudy-days don’t always bring rain the effect is often not significant for particular sites. At Taralga it is. Cloudiness associated with rainfall causes average Tmin to increase 0.07oC/100 mm; the step-change and rainfall explains 20.0% of Tmin variation and there is no residual hidden trend.
 |
Aerial photographs in 1944, 1952 and 1989 at the National Library of Australia may throw some light on the problem but have not been accessed. However, as part of another study, a visit in April 2016 found a small screen, well exposed and maintained (Figure 3). The serial number shows it was made in 2002, thus probably installed in 2003 just before the Tmax up-step in 2004. So a link is established between a site change and the 2004 Tmax up-step.
Figure 2. A step-change in Tmax in 1965 indicates exposure of the Stevenson screen changed. The step-change in 2004 aligns with replacement of a large (0.23m3) Stevenson screen with a small one (0.6m3). The 1973 Tmin step-change aligns with metrication. Segmented regressions (right) are free-fit to show relationships between T and rainfall is robust and not coerced by the analysis. Despite variation due to missing data, lines are statistically parallel and median-rainfall adjusted differences are statistically significant. Accounting for step-changes and rainfall leaves no unexplained residual trend. (Dotted lines indicate average T and median rainfall.)
Mysteries remain. At NAA a 1942 post office plan possibly relates to replacing the verandah to accommodate a telephone exchange. In 1946 a timber-framed lavatory was erected in the yard, which isn’t there anymore. There are other notes up to 1985; and who knows when the yard was fenced-in on two sides with steel cladding?
Figure 3. The small Stevenson screen’s serial number indicates it was made in 2002 and probably installed in 2003. The previous large screen was likely to have been at the end of the concrete path on the right, which leads from the post office. Interestingly, there is a 5-inch (127 mm) raingauge but a standard 8-inch (203 mm) Rimco tipping-bucket gauge for rainfall intensity. (The small concrete pad was for a previous pluviometer.)
Changing the screen size affects the data.
Data are split into decade-long segments each side of the step-change (about 3600 data-days/tranche) and daily temperature distributions are compared. Sounds complicated … but it’s actually simple. We know a Tmax step-change happened and what caused it; distributions provide insights into the shape of the change.
 |
Thought-of as smoothed histograms, frequency is often visualised as probability density plots (Figure 4). But wait … we can also do exploratory statistical tests. The Kolmogorov-Smirnov test for equal distributions found pre and post 2004 Tmax distributions are different (Psame <0.05) while those for Tmin are the same; likewise the Mann-Whitney test for equal medians finds Tmax medians are different [17.5oC (pre) vs. 18.5oC (post)], while Tmin is not (6.0 vs. 5.8). However, as will be apparent, density plots and statistical tests don’t visualise how post-2004 data differ from those measured before 2004 in the large-screen.
Figure 4. Probability density plots of Tmax and Tmin each calculated over identical temperature ranges for the decade before 2004 (black line) and the decade after the small screen was installed (red dashed line). The test statistic (the Kolmogorov-Smirnov test for equal distributions) indicates density distributions for Tmax are not the same; Tmin distributions are not different at the P05 level of significance.
What are percentiles?
Percentile temperatures are daily values ranked by 1%-frequency increments. So 1% of observations are less than the 1st percentile; 2% are less than the 2nd percentile and so on. Each end of the data-range upper and lower extremes are calculated usually as values greater than the 95th and less than the 5th percentile. There are other convenient breakpoints: 25% of daily temperatures are less than the lower quartile; 25% also exceed the upper quartile (75th percentile); the mid-point temperature (which may be different to the mean) is the median (50th percentile).
Percentile differences throw light on the nature of the disturbance: did data step-up uniformly across the percentile range; randomly; or did some sections of the data distribution change systematically?
Percentiles calculated for the decade before 2004 (which is the reference) are deducted from those for the decade after (Figure 5). The expectation is that Tmax differences will be random around the up-step value of 0.95oC; and as there was no up-step in Tmin, random each side of zero. Differences are appraised graphically (Figure 5).
Small-screen Tmax is biased systematically – bias increases with the temperature being measured up to the median (17.5oC), levels-off at the 60th percentile (19.8oC) with an asymptote 1.2oC warmer than equivalent pre-2004 percentiles. So the up-step is caused by the combination of higher temperatures being measured generally, combined with upper-range bias.
 |
Tmin percentile differences contradict equivalence of Tmin probability density plots and the Kolmogorov-Smirnov test for equal distributions. Although median Tmin (6.0oC) lies close to zero percentile-difference (hence the medians of respective distributions are the same) upper-quartile small-screen Tmin is skewed persistently higher by 0.2oC; and between the median and lower quartile, lower by around -0.3oC to as much as -0.6oC (Figure 5). Relative to pre-2004 percentiles the range of the difference is about 0.5oC. Even though there is no apparent change in average (median) Tmin, tails of the distribution are different.
Figure 5. Percentile differences [percentiles calculated for the decade 1 January 2004 to 31 December 2013, minus percentiles for the decade before 1994 (from 1 January 1994)] The LOWESS curve provides a visual reference. The behaviour of extremes [data >95th percentile (30.2oC (Tmax) and 15.0oC (Tmin) and less than the 5th percentile (8.5oC and –2.5oC)] are highlighted.
An additional point illustrated by Figure 5 is that Tmax and Tmin temperatures less than respective 5th percentiles (8.5oC and -2.5oC) depart from the general trajectory abruptly, systematically, adding weight to the likelihood they are adjusted up; between January 2004 and December 2013 some 140 individual values may be affected. At the warm-end of the spectrum, except for the highest (100th percentile) values, which could be random outliers, Tmax and Tmin greater than the 95th percentile (about 180 individual values) are clustered and randomly dispersed around the trajectory indicated by the LOWESS curve.
Conclusions.
- Taralga is one of hundreds of cases from across Australia (many of them ACORN sites) where the change from large Stevenson screens, in use since the late 1800s to small screens caused Tmax (and in some cases Tmin) to abruptly step-up. In a similar way as shown in Figure 5, analysis of many individual sites shows bias increases with the temperature being measured. At sites where the mean or median each side of a screen-change are statistically the same, small screens cause distributions to be skewed; spuriously implying that daily extremes have increased due to the climate.
- There is evidence that temperatures at Taralga less than the 5th percentile are adjusted-up and that for the decade since 2004, some 140 individual data may be affected.
- No Tmin step-change is detected at Canberra airport, Nowra RAN, Richmond RAAF or Sydney Observatory attributable to a time of observation change in 1964. Adjusting imaginary faults in ACORN data using other faulty data is unscientific and disingenuous. An open public inquiry into Australia’s Bureau of Meteorology is urgently needed to clear the air.
[pdate 8/25/17 8:30 pm PDT. Changed second conclusion bullet from 95th percentile to 5th percentile per author’s instructions~ctm]
When there’s a step change from whatever cause, how is it determined which record is “correct”? Since there’s no independent, objective “true” measure of the temperature at the site, isn’t it just as possible that the new equipment reads high as the old reads low, or vice versa?
James,
If you measure T on say the southern side of your house, then after a few decades shift the thermometer to the western side; or build a shed, or concrete the yard or put in a garden and keep it watered; a pool. Would you have changed the weather?
Measured T is an anomaly relative to the long-term mean, which is the site benchmark (the balance of heat sources and sinks). Shifts in the mean track changes in that balance. If the place T is measured is chaotic (not a consistent background heat balance), although there will still be a mean (daily mean, monthly mean, annual mean); data will be chaotic and useless for describing the weather. So reliable temperature measurements are only achieved if the conditions under which they are observed are consistent. Hence the value of statistical tests (and visualisations) for determining if data are fit-for-purpose.
(There are many examples of where a site moved somewhere cooler for a few decades; then warmer again. There are also many examples of where the data obviously changed, but no one remembers why and it was not written down, or the metadata is lost or ignored.) I don’t know what happened in 1965 at Taralga for instance; and I don’t have much interest in investing the time and effort into finding out. However, something did change that affected exposure of the screen.
Thanks for your interest,
Cheers,
Bill