Older records may be fragile but was diligence any better in recent decades?
Guest essay by Bob Fernley Jones
Out of over 20,000 Bureau of Meteorology (BoM) weather stations on record , 112 have had their data “corrected” under a process known as homogenisation, (the acronym for it in full is ACORN-SAT). Oddly though, eight of these high quality sites, are admitted to, quote:
“…have some urban influence during part or all of their record, hence are excluded from the annual temperature analyses”.
Thus only the resultant 104 ACORN site records are used to establish temperature trends and it is these that are partly reviewed herewith. Some ACORN stations are known under a single site name but are actually several different locations combined, (typically moving from in-town to the airport), and their homogenisation is partly achieved by including data trends from surrounding stations of lesser status. The change from Fahrenheit to centigrade units in 1972 is a further complication involving the reading and rounding of 5/9 ratio smaller Fahrenheit units. Also of difficulty are significant issues in time of day in the readings.
In probing the reliability of the resultant ACORN data; which were affected by varying resources etcetera during the 104 years of ACORN, there is an immediate controversy over imbalance in short-term station records being used ONLY in recent times. (Oh, and some are in hot places).
Whilst many of the 20,000 “lesser” stations are now closed, (as are many older ACORN composite sites), there must be adequate older data to redress that imbalance. Furthermore, since homogenization has greater importance with older records embracing a time known for reports of great heat and drought, it is arguable that any bias in time-series distribution should be the other way around, if there is to be a serious effort to determine centennial trends. In other words, it is desirable to increase the sample size of the older evidence if it is deemed to be weaker; as seems to be the implication in making the homogenizations. So why is it that in ACORN, the sample size in recent data which should be more reliable has been increased when it ought to be the other way around; if older data are suspect? More short record info is at § 6.
However, detailed consideration of this background would bring controversies into play, and the focus of this study is a test for reasonableness in only the distribution patterns of data; thus arguably avoiding any controversy.
 Unzip the BoM file from the BoM FTP site for details.
 But controversially they are used in homogenization of rural sites and enable sensational media items of record heat at these Urban Heat Island (UHI) sites, including the grossly affected Melbourne R.O. which had to be closed Jan/2015; replaced by Olympic Park).
 Since 1964 all high/low readings were recorded during the 24 hours ending at 9 am; thus maxima are actually from the previous day. Prior to 1964 there were many other methods.
———————— And now about that data ————————
One of the difficulties in analysing the BoM data is the vastness of the database, which runs to about 38,000 lines of daily data in any typical long time-series record. Each series might have a mix of parameter concerns, but they are not easy to detect in those formidably long lists of numbers. Fortunately, Microsoft EXCEL software used herein enables sorting and plotting of those daily data into a visually comprehensible form. Any counterintuitive thought that there is too much noise in the daily data to make a sensible chart is not an issue because of high compression of that data which visually smooths such noise. Many features become prominently visible, the most obvious being seasonal cycling, blocks of missing data, and step-changes. Furthermore, EXCEL can use Search and Logic Formulae to find, replace or delete specific types of information, thus enabling the plotting of particular features.
The charts that follow are unusual in displaying a great deal of colour coded information into one page width. It is recommended to click them in the webpage to open them into a wider window, and to pause sufficiently to study the detail. Several charting methods are used in order to give alternative visualizations of data concerns for 24 stations, with most emphasis on recent decades. Sections §4 & §5 show two alternative methodologies together with expanded details from the main plots.
The main analysis starts from fig 2) and covers recent decadal periods at six ACORN sites on this main page and continues with 17 additional stations in sections attached.
Ahead of that, Fig 1 for Merredin is firstly viewed as background information, in order to show typical relationships over a long record. Merredin is an example station for detection of anticipated recent improvements in data quality versus records going back over a century, (additional examples in §5).
 Supplementary information is available for a total of 46 stations (44 % of 104) all having various corruptions in their data.
§1 – Baseline comparison; Fig – 1) Merredin, a long record:
· (Part 1a): The Bureau’s homogenised data was sorted using Microsoft EXCEL software to show a chart of all integer values over the past century as single points. (EXCEL found 7,778 integers, of which not all are visible because of limitations in pixel definition and overlapping, but varying densities are visible). Over a long time-series like this, a ten-percent distribution of integers is nominally expected, assuming good data. Whilst there are potential variations that could cause deviations from that norm, it is evident, (more accurately in the actual EXCEL spreadsheet counts), that in Fig 1a, intervals labelled ①②⑤&⑥ are close to that expectation. However, there is significant deviation in intervals ③&④ where there should be a higher number of integers, (Perhaps more so because there could be intent or a tendency for ocular approximations what with the smaller Fahrenheit units before decimalization in 1972?).
Whatever, the recent periods of ⑦&⑧ are seriously corrupted. Also, intervals ⑨&⑩ are bad at 17.1% & 15.3% although their 1,200-day sample is more modest. Nevertheless, there are serious unexplained inconsistencies in the records.
What is very odd in terms of data distribution is that the minima and maxima are affected for conflicting durations before 1972 including a baffling big “hole-in-the-middle” especially in the minima, as also seen in similar long record charts in §5.
Note: Prior to 1972, the data are shown in ⁰C but were found originally in ⁰F. The various controversies in rounding and conversion are not considered to be relevant here.
(Part 1b): Another aspect of the integers, is that they have occurred in both short and long clusters lasting from several days through to several or more years. Strangely, (but not uniquely), Fig 1b shows a high concentration of small clusters in a period similar to that of problems associated with AWS’. However, Merredin is archived as a manual station currently equipped with mercury thermometers according to this BoM source.
Additional comments on Fig 1 continue below the image. (Click to zoom into wider window):
|History from the ACORN site catalogue: The original site (10093) was at a research farm about 5 km west of town. It moved 150 m southeast in May 1957. The Research Station continued to make observations until 1985 and these observations were used to merge the two datasets. The current site  has been operating since 1966. It became overgrown for a time but was cleared in March 1986.|
· (Part 1a) Arrow markers: The four arrows relate to the ACORN site history and it can be seen that these points, show no correlation with the changing data patterns over the century-long period.
· (Part 1c): This unitless bar chart shows within the available pixel definition the distribution of days where ACORN values are coded 99999.9 = No data. Before the 1960’s the records are rather incomplete, but the middle period has good completeness despite the otherwise inconsistent patches ⑤⑥⑨&⑩. However, the last two decades are notable for their incompleteness, inaccuracy, and inconsistency.
Summary opinion on fig 1: No claim is made that Merredin is “typical”, although five comparable variations on it follow in §5. Neither can it simplistically be a criticism that it is “cherry-picking” to reveal its bad data record. It is one of the ACORN-SAT homogenised premier sites that are groomed out of many other surrounding sites, and claimed to be “Amongst World’s Best Practice”.
ÚÚÚÚ Now for those problematic modern decades ÚÚÚÚ
§2 – Figures 2–7) Six examples circa 2000:
Notes on reading the graphics:
•Click each image to open in wider window•
Red plots show maxima and blue the minima temperatures (0C). Darker shades show substandard values in whole ⁰C (integers) instead of to tenths. White space shows No Data. Colours plotted last may cover over part of earlier plots, so their sequencing is indicated. The data are all highly compressed daily values with limitations in pixel definition, but an annual seasonal cycle etcetera is obviously evident.
These two-part charts firstly survey integers for two decades around the time when AWS’s were introduced. The lower parts of the charts survey the full centennial record for distribution of missing ACORN data (99999.9 values). Caution; horizontal x axis scale differs greatly between the two parts.
Fig 2 – Larapuna (Eddystone Point):
From 1990 through to late 2003 the missing or integer values are excessive. (There are smaller “white spaces” in Fig 2a prior to 1997-2003). With introduction of the AWS there are no data in the maxima for nearly seven years, and, for the same period in the minima there are substandard whole-degree (integer) values. How did Head Office not notice or fail to correct these problems for some thirteen (~13) years? The simultaneous integer and blank data for ~7 years is intriguing; see also Fig 8 in §4).
Fig 3 – Cape Otway:
In this case the failure with the noted second AWS in both maxima and minima gave substandard integer values that remained uncorrected for over nine (>9) years!
The first AWS gave No Data in the minima for almost a whole year before it was replaced with new!
Fig 4 – Forrest:
This example is included primarily to show that substandard integer outcomes were apparent in 1993 with an early AWS immediately upon installation but remained uncorrected for over two years, and it was employed instead of the established and better functioning systems.
•AWS’ were mostly adopted circa late 1996•
Based on a limited search, various failure periods overlap within the BoM system until at least 2004, resulting in excessive integer counts spanning at least eleven (~11) years in total!
It is also of interest that in several cases, failure occurred only in the minima, an issue discussed further with figure 5a Bourke.
Fig – 5a Bourke:
A letter to me from the BoM of 3/June/2015 signed by acting Director of Meteorology and CEO, Dr Ray Canterford, responded to my follow-up questions from 2014 on bad data at Bourke:
…The occurrence of whole (integer) temperatures in the Bureau’s database is a well-known issue that affected early-generation Automatic Weather Stations (equipment), due to limitations in the coding software at the time. It is described on page four of the ACORN-SAT observation practices document on the Bureau’s ACORN-SAT website. [See §8 References] Where no adjustment is deemed necessary to the raw data, the rounded figures remain in the database…
But, (whilst not wishing to hypothesise in this study), unless the software differs between different hardware it seems unlikely that coding errors would erratically affect some stations but not others, and that the malfunctions are temporally sensitive in the software. In this graphic, maxima are affected earlier and for longer than in the minima. Minima values did not become corrupted until about three years after the hardware was installed. (This might suggest hardware instability rather than software “instability”?).
Fig 5b – Comments: In some cases as in the example of figure 2, it is seen that long periods may be void of data during AWS operation and it is apparent that together with the parallel substandard integer values that these two malfunctions are likely interconnected in some way. (Such concise parallels, also in fig. 8, are unlikely). Less obvious, in this example of Bourke, is that a more random distribution of increasing missing data can exist with an AWS. Limitations in pixel definition only give a broad impression of count but the EXCEL digitally determined values of day-counts of missing values associated with the AWS are for the minima 537 and for the maxima 415. Note that minima are plotted secondly so smother much of the maxima 99999.9 values underneath.
In the early days of my curiosity of matters BoM, I asked their enquiry desk why it was that Bourke had a very complete record prior to introduction of an AWS but then it deteriorated very badly. Here follows part of their response from mid last year:
Case # E3IH2A1856:
“…There are many different reasons for missing data. We refer you to our previous correspondence on the issue of missing AWS data. In general, one should not assume that an Automatic Weather Station (AWS) would necessarily provide more complete data than manual observations, especially in the instance that the AWS replaces a site with a very good history of data completeness. Bourke’s long-term data completeness over its whole record is 98.2%, very close to the median. The completeness of the Bourke PO data was exceptionally high. The completeness of the AWS has not been as good, but has returned to above 98% in the last 3 years…”.
Whatever, the strikingly bad performance of the AWS was allowed to exist for over a decade and has only improved over the last three or so years.
Fig 6 – Rutherglen:
The variation in this example is less visible, so the colour contrast is increased as an aid. Careful examination reveals a different pattern of small clusters only within the larger ellipse, (which spans some 1,500 days).
Shortly after installation of the 1998 AWS the first of a series of some 17 short clusters of integers occurred. (They are more see-able in the source EXCEL spreadsheet). Minima were not necessarily in phase with maxima, but again, it seems less likely that a software problem could be so erratically and temporally sensitive. (A conditional is that some of the “thinner” intervals are only three or four integers wide and might be unusual clusters within possibility, but less unlikely so when max&min are paired, as are some of them). However, the several pairs of “broader” clusters cannot possibly be flukes. An issue here is that there are clear malfunctions which seem to point towards inherent hardware instability rather than software coding issues as claimed by the Bureau. (Occurring in either short or very long clusters sometimes years after trouble free installation and suggestive in retrospect of; We don’t really know how or why…..when they later became attentive to it?).
§3 – Discussion on figures 2 through 6:
These five graphics give samples of the main concerns. Additional examples and other concerns can be accessed by clicking the links under the following sections; §4 – Fig’s 7 -16) & §5 – Fig’s 17 – 22). §6 Fig’s 23 & 24
This station search was primarily on remote coastal and inland regions under the consideration that remote sites are those that may be arguably in need of AWS’ and that perhaps transmission/logging of data might be potential factors. (Also UHI effects are less likely and station names comprising only one location more common)
Any suggestion that these remoter sites are unrepresentative is countered by them all being part of the acclaimed high quality ACORN-SAT homogenized system.
§4 – Figures 7 – 16) Nine more examples of problem sites circa 2000:
CLICK: section4.pdf to open in new window
•Figures 10 through 16 show much the same sort of information but offer different visual perspectives•
§5 – Figures 17 – 22) Centigrade versus Fahrenheit, including half-degrees:
CLICK: section5.pdf to open in new window
•These five examples are long records similar to that of introductory Fig 1 but with some visually expanded details•
§6 – Nineteen short record sites tabulation and Figures 23 & 24
This is further to the opening Background discussion, CLICK: section6.pdf to open in new window
§7 – Basic Conclusions:
· It is nonsensical when there are thousands of older stations on record, that older records that are argued to require ACORN corrections should have a smaller sample than the “more reliable” largely uncorrected data of recent times.
· In the 24 stations reviewed herewith, corrupted data continued decadaly as if no one was particularly interested back then, at a time before alarming climate change theories became more popular.
· These bad data were especially associated with AWS’s and differed from other bad data in earlier times.
· Extracting from §5: …following 1972 decimalization there were significant increases in integer value counts. This is counterintuitive given that decimals are easier to read with 9/5 larger centigrade units. Thus, conversion from the original Fahrenheit temperatures has somehow gone painfully awry…
· The net outcome is that the ACORN data are not credible for reliable trend determinations.
· The penultimate conclusion is heightened (although it was not discussed herein) by another fact of the dismissal by the Bureau of strong data and comparative reports of hotter times prior to their ACORN “Start of Time” of 1/Jan/1910. Accounts of great heat in the past were paralleled by prolonged and devastating droughts. Tellingly, expert hydrologists advise that droughts result in higher air temperatures.
 E.g. Watkin Tench’s Book on Port Jackson settlement chapter 17, describing myriad bats and birds dropping dead from the trees etcetera. Etcetera.
 E.g. Professor Stewart Franks and histories such as the so-called “Federation Drought” from the late 1880’s.
§8 – References:
ACORN Station Catalogue. (Including history of sites involved)
ACORN-SAT Website main page This includes other menu items including claims of peer review “World Quality”
Dorothea Mackellar (During visit to England 1908, before the beginning of ACORN time)
§9 – Disclosures:
I’m a retired mechanical engineer with no past or present funding for this research from anyone or any conflicting vested interests or motivations other than to see fair play in science-driven policy making.
Compiled by Bob Fernley-Jones Melbourne Oct/2015.