Is homogenisation of Australian temperature data any good?

Part 2. Marble Bar, the warmest place in Australia

Dr Bill Johnston[1]

scientist@bomwatch.com.au

Maximum temperature data for Marble Bar exemplifies all that is wrong with the Australian Climate Observations Reference Network – Surface Air Temperature (ACORN-SAT) dataset used to depict Australia’s warming. Subjective adjustments based on faulty metadata and highly correlated comparator sites that likely embed parallel faults results in trends in homogenised data that are unrelated to the climate.        

Read on … 

Background

Located 150 km southeast of Port Hedland on the edge of the Great Sandy Desert, Marble Bar is reputed to be the warmest place in Australia. The dry season from April to November is extremely dry, while during the ‘wet’, potential evaporation exceeds rainfall even in the wettest month of February. Nevertheless, irregular cyclonic intrusions from the Timor Sea and eastern Indian Ocean can wreak havoc across the Pilbara including at Marble Bar.

Temperature and rainfall have been measured at several locations under a variety of conditions by post office staff and volunteers since 1901. However, due to its isolation the weather station was rarely inspected by the Bureau of Meteorology (BoM), consequently metadata (data about the data) is unreliable. What was it like ‘doing the met’ under blistering hot conditions and how did equipment including thermometers fare with furnace-like winds, desert dust, and at the other extreme, cold winter nights? Typical of an arid swath of northwestern Australia, Marble Bar is an Australian Climate Observations Reference Network – Surface Air Temperature (ACORN-SAT) dataset used to calculate Australia’s warming (Figure 1).

Figure 1. Marble Bar in relation to other ACORN-SAT sites in northwestern Australia (left), and below, staff outside the original post office in 1900, which was built of local stone (State Library WA).

ACORN-SAT metadata failed to mention that a Stevenson screen was originally supplied in 1911 and that the site moved to the new post office on the corner of Francis and Contest streets in 1948 (Figure 2). Also, that the post office was extended in 1961 to accommodate the telephone line exchange which probably required the Stevenson screen to move up-hill towards the shady watered yard behind the residence. It seems the post office closed in 1984 and its functions including weather observations transferred to the general store opposite the Ironclad Hotel (now licenced post office) 230m east. Later, in 1988 the site allegedly moved east, probably to the Roadhouse, from where it relocated to its current position 100m southwest in 1997. 

Figure 2. The side-view of the 1948 post office facing Francis Street (left) photographed in 2016 with the 1961 telephone line exchange annex distant from the camera, and the former residence facing Contest Street (realestate.com.au online sales brochure). Original plans show a path between the residence and the rear of the post office and several out-buildings behind including fuel store, generator/battery shed, and garage.

Objective detection of changes in data using robust statistical methods and post hoc attribution is more reliable than using ACORN-SAT and site summary metadata which is often inaccurate and incomplete. Although now restored, by 1945 the original government offices incorporating the post office had been damaged by cyclones and was riddled with termites. Planned for in 1944, the new post office shown in Figure 2 was not opened until August 1948.

A down-step in Tmax data in 1945, which ACORN-SAT interpreted as site move appeared to be due to observers consistently rounding observations down to the nearest whole^oF, which after converting to Celsius effectively reduced average rainfall-adjusted Tmax by about 0.5^oC. While rounding ceased in 1960, data remained cooler than previously until 1988, when it was said the site had become overgrown and was moved “1 km to a residence”. Within that time the Stevenson screen moved clear of the 1961 telephone line exchange, which affected the frequency of high vs low extremes; observations were metricated from 1 September 1972, which affected precision; and in 1984 the site moved to the general store (the current licenced post office), where whole and ½^oC were over-reported from 1995 to 1997. It was only after the site relocated in 1998 that data were reasonably consistently observed to the nearest decimal place.

Lack of precision (low R²_adj) in the relationship between Tmax, rainfall and site changes, indicates the quality of Marble Bar data is relatively poor. While site changes caused Tmax data to warm 0.9^oC since 1901, no warming is attributable to the climate.

Homogenisation

The Bureau’s homogenisation methods are plagued by faulty metadata and the biased selection of neighbouring sites used to make adjustments.       

Read on …   

The most obvious problems with the Bureau’s temperature homogenisation methods are:

• Metadata for weather stations including Marble Bar is either misleading, scant or non-existent.
• As no weather stations have stayed the same, it is highly likely within regions that many had undergone concurrent changes/upgrades since they were first established. Examples include post-WWII changes at aerodromes in the late 1940s, installing telephone line exchanges at post offices in the 1950s, metrication in 1972, installation of the major trunk network microwave towers in post office yards by Telecom by 1975, the staged introduction of 60-litre Stevenson screens, automatic weather stations etc.     
• As many weather stations are affected by synchronous changes and all exhibit similar seasonal cycles, selection of comparator datasets from the pool of possible contenders on the basis of linear correlation of first differences is bound to rank those having parallel faults as candidates for making ACORN-SAT adjustments.

Using faulty metadata to subjectively determine times when something changed, allows changepoints to be manipulated to achieve pre-determined trends. It also discourages further investigation of their cause. For instance, the 1944/1945 down-step at Marble Bar was due to a precision problem not a site move. Other changes included that from 1966 to 1985 the frequency of daily observations greater than 95^th day-of-year percentiles was significantly depressed, probably due to watering or shade. Reconstructing what actually happened based on effects observed in data requires changepoints to be identified statistically using robust, statistical methods and post hoc investigation of causes.

Using Pearsons linear correlation to select up to 40 neighbouring sites for constructing comparative reference series is biased. As monthly averages cool from summer to winter and warm from winter to summer, residual cycles in first-differenced data inflate significances of correlations. Thus, from the pool of potential neighbours whose data are not homogeneous, linear correlation is likely to select those having parallel faults. Furthermore, increasing the number of comparators cannot overcome the high likelihood that station changes within regions are somewhat synchronised. 

Objective, replicable homogenisation would investigate time-related changes in properties of datasets using objective statistical tests (of which there are a number) and relate changes in the data to what is known about the site post hoc. For example, by comparing detected changepoints with BoM metadata, documents, maps, plans and aerial photographs held by the National Archives and the National Library of Australia, state authorities, museums, historical societies, newspaper reports etc. Even if supporting information is not available, statistical detection based on the probability of an abrupt sustained change against the NULL hypothesis of no change should be sufficient evidence that change occurred.

Historic data were not collected to be used decades in the future to detect trend and change. Due to inaccurate metadata, poor site control (screen and instrument deterioration, watering, shade), and prior to metrication on 1 September 1972, lack of precision by observers, Marble Bar data could not be regarded as high-quality. As the Bureau’s homogenisation methods are deeply flawed the ACORN-SAT project should be abandoned.

Two important links – find out more

First Link: The page you have just read is the basic cover story for the full paper. If you are stimulated to find out more, please link through to the full paper – a scientific Report in downloadable pdf format. This Report contains far more detail including photographs, diagrams, graphs and data and will make compelling reading for those truly interested in the issue.

Click here to download the full paper with photos graphs and data.

Second Link: This link will take you to a downloadable Excel spreadsheet containing a vast number of Marble Bar data points for the years from 1901 to 2020 and which was used in the analysis of the Marble Bar weather records to support the Full Report.

Click here to download the Excel spreadsheet with data points 1901 to 2020

Note: Line numbers are provided in the linked Report for the convenience of fact checkers and others wishing to provide comment. If these comments are of a highly technical nature, relating to precise Bomwatch protocols and statistical procedures, it is requested that you email Dr Bill Johnston directly at scientist@bomwatch.com.au referring to the line number relevant to your comment.   

---

[1] Former NSW Department of Natural Resources research scientist and weather observer.

Is homogenisation of Australian temperature data any good?

Part 1. Methods case study: Parafield, South Australia

Dr Bill Johnston

scientist@bomwatch.com.au

Imagine arriving at a crime scene with a toolkit of technical knowhow, where the only evidence of what happened is some maximum temperature data and incomplete notes about where they were observed and conditions that may have affected them.

Time starts NOW …..

Background

Methods used by Bureau of Meteorology (BoM) senior research scientist Dr Blair Trewin, lead author of Working Group I of the IPCC’s 2021 Sixth Assessment report, and a member of the World Meteorological Organisation Expert Team on Climate Monitoring and Assessment, and before him Neville Nicholls, Simon Torok, Paul Della-Marta (and others) and Karl Braganza, are in urgent need of review. 

All have published papers in peer-reviewed scientific journals relating to data homogenisation and based on those papers, which are in the public domain, claims have been made repeatedly by The Conversation, the ABC, The Guardian and Fairfax media and even The Australian, and for Braganza, to the 2019/20 Bushfires Royal Commission, that temperature measured at BoM weather stations has increased in recent decades due to anthropogenic emissions of CO₂.

Maximum temperature (Tmax) varies up and down depending on how much net energy is convected away as latent heat via evaporation from plants, soil and surfaces (collectively referred to as evapotranspiration). As evapotranspiration at dryland sites cannot exceed the rainfall, latent heat loss equals 2.26 MJ/kg of water evaporated, equivalent to is 2.26 MJ/mm of rainfall received.

Evapotranspiration of average rainfall at Parafield (443 mm) for example, would remove 1001.2 MJ of energy (443mm times 2.26 MJ), which is 15.7% of average total solar exposure of 6345 MJ/yr. The balance, not accounted for by evapotranspiration heats the air in contact with the ground, and it is that portion of the local heat balance that is measured during the heat of the day by Tmax thermometers held 1.2m above the ground in Stevenson screens (Figure 1). It follows that for any weather station, dry years are warm and the drier it is the warmer it gets.

Figure 1. Maximum and minimum thermometers held horizontally; and dry and wet-bulb thermometers held vertically in the well-maintained 60-litre Stevenson screen at Cranbourne Botanic Gardens, Melbourne.

The First Law Theorem, that available energy is expended via latent and sensible heat pathways, provides a physical reference frame for evaluating the quality of weather station data based on the relationship between Tmax and rainfall, which should be linear and negative. The proportion variation explained by rainfall, known as the coefficient of determination (R² or more exactly R² adjusted for the number of terms in the model (R²_adj)) provides a measure of the quality of Tmax data (Figure 2). Watering, deterioration of equipment and lackadaisical practices reduce the goodness of fit (R²_adj <0.05 or 50% of variation explained), while for an excellent site it may be as high or higher than 0.7 (>70%).

The first tool in the toolbox, naïve linear regression of the form average Tmax ~ annual rainfall, investigates the fitness of Tmax data using rainfall as the comparator or control variable. The coefficient (Tmax/100mm of rainfall), its sign (+ or ‑), significance (P level) and variation explained (R²_adj) assist diagnosis of the dataset.

Read on …   

Figure 2. The naïve relationship between average Tmax and annual rainfall at Parafield. Although the slope is negative and highly significant (P <0.001), only 16% (R²_adj = 0.16) of variation is explained.

The line in Figure 2 only explains the rainfall portion of the Tmax signal and so while rainfall reduces Tmax ‑0.04^oC/100mm and the coefficient is highly significant (the P level is less than 0.001), only 16% of variation is explained, thus, the relationship is very imprecise. So, what of the unexplained variation – the residual non-rainfall portion of the signal?

Calculated as the difference between the line (the equation), and each Tmax data-point, most statistical packages provide a table of residuals, which may be positive or negative. However, programs like Excel require them to be calculated manually (there are examples on the internet).

Independent statistical methods are used to examines Tmax ~ rainfall residuals – that part of the signal not related to rainfall, which if all variation is explained, should be random and free of systematic effects.

Read on …   

Residuals from naïve regression are usually small, zero-centred numbers. Adding the dataset grand-mean to each value reconstitutes the original scale without changing their relativity or other properties. Except that the points in Figure 3 appear random, the Figure provides little additional information. However, while the data seem sound, low precision suggests a variable may be missing.

Figure 3. Rescaled residuals from the Tmax ~ rainfall relationship in Figure 2.

Linear regression partitions the original dataset into the portion of the Tmax signal attributable to rainfall, which is the physically deterministic covariable described by the linear fit in Figure 2, and variation that is independent of rainfall, which are the re-scaled residuals in Figure 3.

Next, the rescaled residuals in Figure 3 are analysed in the order they were observed and checked for changes that may indicate background, non-rainfall effects (Figure 4).

Figure 4. Step-change analysis of re-scaled residuals identified step-changes in 1973 which was caused by a site move before 1968, and another in 1999, probably caused by replacing a former 230-litre Stevenson screen with a more sensitive 60-litre one. (Data from 1955 to 1973 exist, but have not been digitised)       

Step-changes Tmax residuals in Figure 4 were highly significant, meaning there was a very low probability they were due to chance. Analysis was undertaken using the Excel 2007 version of Sergi Rodionov’s Sequential Three-step Analysis of Regime Shifts (STARS) v6.3 (ICES Journal of Marine Science, 62: 328e332 (2005) doi:10.1016/j.icesjms.2005.01.013), which is in the public domain (https://sites.google.com/view/regime-shift-test/downloads). While other step-change tools are available (see https:// www.marinedatascience.co/blog/2019/09/28/comparison-of-change-point-detection-methods/), STARS is convenient, reliable and comparable with other methods.   

Closing the case. Metadata, archived aerial photographs maps, plans, satellite images etc. are used where possible to confirm step-change scenarios and to find the combination that best fits the data. There is only one optimal outcome, which is that segmented responses to rainfall are the same (segmented regressions are parallel) and that rainfall-adjusted group means are different (segmented regressions are not coincident). While the data stepped up, no evidence was found that the climate had changed or warmed.

The BomWatch approach is objective, robust and replicable and does not involve data manipulation. It is also amenable to batch-processing of paired Tmax-rainfall datasets in the Bureau of Meteorology’s database, which would put to rest the hype about climate change and warming once-and-for-all. 

Read on …   

Bias in the BoM’s homogenisation methods

Comparative homogenisation methods used by the Bureau are highly questionable. Problems include:

1. Use of faulty and unreliable metadata for specifying/verifying changepoints in target site data (data to be homogenised);
2. First-differenced linear correlation with the target-site data, selects comparators that likely embed parallel faults;
3. Overly complicated methods lack transparent checks and balances that provide confidence that their methods are trustworthy.

Data for Parafield airport (Bureau ID 23013) commence on 1 April 1939 but site-summary metadata does not provide coordinates for the site. Google Earth Pro shows the current Almos automatic weather station (AWS) is near the centre of the airport about 1 km south of the control tower at Latitude ‑4.7977, Longitude 138.6281. While data show a naïve trend of 0.31^oC/decade (3.1^oC/century) a gap in available data from 1955 to 1973 is unexplained. Due to step-changes before 1973 and in 1999, data are not homogeneous, nevertheless, together with other sites, Parafield was used to homogenise ACORN-SAT sites at Oodnadatta, Tarcoola, Woomera, Ceduna, Kyancutta, Port Lincoln, Nuriootpa, Adelaide (Kent Town) and Robe in 2011 (Figure 5).

Figure 5. The first (2011) iteration of ACORN-SAT used data for Parafield Airport, 15 km west of Adelaide, with other cross-correlated sites to homogenise ACORN-SAT sites at Oodnadatta (859 km distant), Tarcoola, Woomera, Ceduna (546 km), Kyancutta, Port Lincoln, Nuriootpa, Adelaide (Kent Town) and Robe. ACORN-SAT sites are indicated by red squares. Weather stations in South Australia having >20 years of data are indicated by open circles, some of which are named.

In contrast to Acorn-Sat v.1, the 2021 iteration (v2.2) used Parafield with other cross-correlated sites to only adjust Adelaide (Kent Town), Cape Borda, Port Lincoln, Robe, Kyancutta, Nhill and Ceduna (Figure 6). So, something in the method caused Parafield to be used for some sites in 2011, but not the same sites in 2021. Illustrating the complexity of the Bureau’s methods, radar graphs show location of the 24 sites used to homogenise Ceduna in 2021 and the 30 sites used for Oodnadatta. (The great circle distance from Oodnadatta to Deniliquin airports is 1263 km and to Cobar 1096 km.)

Figure 6. ACORN-SAT v.2 used Parafield airport data with other cross-correlated sites, to only homogenise ACORN-SAT sites at Adelaide (Kent Town), Cape Borda, Port Lincoln, Robe, Kyancutta, Nhill and Ceduna, which is a geographically less-dispersed suite compared to v.1. The radar plots locate the of 24 sites in total, used to homogenise ACORN-SAT data for Ceduna and the 30 sites used for Oodnadatta.

Homogenisation of Australia’s temperature record underpins the Australian government’s decarbonisation agenda including upheavals in energy supply networks, rapid increases in energy costs that will force energy-intensive industries offshore, and the destruction of habitats and despoiling of landscapes by wind turbines, solar panels and woody-weeds.

While true costs have not been revealed, the impact on future generations will be at a scale never envisaged by those advocating for change. While elites will accumulate wealth and power, those agitating on their behalf for climate justice and other paradoxical causes are laying the foundations for their own highly indebted, personally limiting Orwellian future. Without abundant, reliable, affordable energy and exports of minerals, coal and agricultural commodities to support orderly market-based transitions, including to nuclear power, Australia risks becoming a voiceless, crippled, heavily manipulated society devoid of a productive base.

Conclusions

• The Bureau’s homogenisation methods lack rigor, replicability and quality control and are likely to result in trends that have nothing to do with the climate. Selection of comparator neighbouring datasets on the basis of first-differenced correlations axiomatically identifies those with parallel faults.
• Use of reference series composed of data that are not homogeneous to detect and adjust changes in ACORN-SAT data has no scientific or statistical merit and should be abandoned.     

An important link – find out more

The page you have just read is the basic cover story for the full paper. If you are stimulated to find out more, please link through to the full paper – a scientific Report in downloadable pdf format. This Report contains far more detail including photographs, diagrams, graphs and data and will make compelling reading for those truly interested in the issue.

Click here to download the full paper including, photos and tables of data

Note: Line numbers are provided in the linked Report for the convenience of fact checkers and others wishing to provide comment. If these comments are of a highly technical nature, relating to precise Bomwatch protocols and statistical procedures, it is requested that you email Dr Bill Johnston directly at scientist@bomwatch.com.au referring to the line number relevant to your comment.   

Dr Bill Johnston

25 November 2022

Note: Line numbers are provided in the linked Report for the convenience of fact checkers and others wishing to provide comment. If these comments are of a highly technical nature, relating to precise Bomwatch protocols and statistical procedures, it is requested that you email Dr Bill Johnston directly at scientist@bomwatch.com.au referring to the line number relevant to your comment.