Guest Post by John Goetz
Adjustments to temperature data continue to receive attention in the mainstream media and science blogs. Zeke Hausfather wrote an instructive post on the Climate Etc. blog last month explaining the rationale behind the Time of Observation (TOBS) adjustment. Mr. Hausfather pointed to the U.S. Climate Reference Network (CRN) as a source of fairly pristine data that can be used to analyze TOBS. In examining the CRN data, there is no doubt that the time of observation affects the minimum, maximum, and average temperature recorded on a given day. Also, changing the TOBS one or more times during a station’s history can affect the station’s temperature trend.
Temperature adjustments have bothered me not because they are made, but because there is a broad assumption that they skillfully fix a problem. Somehow, climate scientists are capable of adjusting oranges into apples. However, when adjustments are made to temperature data – whether to correct for TOBS, missing data entries, incorrect data logging, etc. – we are no longer left with data. We are left instead with a model of the original data. As with all models, there is a question of how accurately that model reflects reality.
After reading Mr. Hausfather’s post, I wondered how well the TOBS adjustments corrected the presumably flawed raw temperature data. In the process of searching for an answer, I came to the (preliminary) conclusion that TOBS and other adjustments are doing nothing to bring temperature data into clearer focus so that global temperature trends can be calculated with the certainty needed to round the results to the nearest hundredth of a degree C.
The CRN station in Kingston, RI is a good place to examine the efficacy of the TOBS adjustment. This is because it is one of several CRN pairs around the country. Kingston 1 NW and Kingston 1 W are CRN stations located in Rhode Island and separated by just under 1400 meters. Also, a USHCN station that NOAA adjusts for TOBS and later homogenizes is located about 50 meters from Kingston 1 NW. The locations of the stations can be seen on the following Google Earth image. Photos of the two CRN sites follow – Kingston 1 W on top and Kingston 1 NW on the bottom (both courtesy NCDC).
The following images are of the Kingston USHCN site from the Surface Stations Project. The project assigned the station a class 2 rating for the time period in question, 2003 – 2014. Stations with a class 1 or class 2 rating are regarded as producing reliable data (see the Climate Reference Network Rating Guide – adopted from NCDC Climate Reference Network Handbook, 2002, specifications for siting (section 2.2.1) of NOAA’s new Climate Reference Network). Only 11% of the stations surveyed by the project received a class 1 or 2 rating, so the Kingston USHCN site is one of the few regarded as producing reliable data. Ground level images by Gary Boden, aerial images captured by Evan Jones.
CRN data can be downloaded here. Download is cumbersome, because each year of data is stored in a separate directory, and each file represents a different station. Fortunately the file names are descriptive, showing the state and station name, so locating the two stations used in this analysis is straightforward. After downloading each year’s worth of data for a given station, they must be concatenated into a single file for analysis.
USHCN data can be downloaded here. The raw, TOBs, and homogenized (52i) files must be downloaded and unzipped into their directories. All data for a station is found in a single file in the unzipped directories. The Kingston USHCN data has a file name that begins with USH00374266.
Comparison of Kingston 1 NW and Kingston 1 W Temperatures
Both Kingston CRN stations began recording data in December, 2001. However, the records that month were incomplete (more than 20% of possible data missing). In 2002, Kingston 1 NW reported incomplete information for May, October, and November while Kingston 1 W had incomplete information for July. Because of this, CRN data from 2001 and 2002 are not included in the analysis.
The following chart shows the difference in temperature measurements between Kingston 1 NW and Kingston 1 W. The temperatures were determined by taking the average of the prior 24-hour minimum and maximum temperatures recorded at midnight. The y-axis is shown in degrees C times 100. The gray range shown centered at 0 degrees C is 1 degree F tall (+/- 0.5 degrees F). I put this range in all of the charts because it is a familiar measure to US readers and helps put the magnitude of differences in perspective.
Given the tight proximity of the two stations, I expected their records to track closely. I found it somewhat surprising that 22 of the months – or 15% – differed by the equivalent of half a degree F or more. This makes me wonder how meaningful (not to say accurate) homogenization algorithms are, particularly ones that make adjustments using stations up to 1200 Km. With this kind of variability occurring less than a mile apart, does it make sense to homogenize a station 50 or 100 miles away?
Comparison of Kingston 1 NW and Kingston 1 W Data Logging
A partial cause of the difference is interruption in data collection. Despite the high-tech equipment deployed at the two sites, interruptions occurred. Referring to the previous figure, the red dots indicate months when 24 or more data hours were not collected. The interruptions were not continuous, representing a few hours here and a few there of missing data. The two temperature outliers appear to be largely due to data missing from 79 and 68 hours, respectively. However, not all differences can be attributed to missing data.
In the period from 2003 through 2014, the two stations recorded temperatures during a minimum 89% of the monthly hours, and most months had more than 95% of the hours logged. The chart above shows that calculating a monthly average when missing 10-11% of the data can produce a result with questionable accuracy. However, NOAA will calculate a monthly average for GHCN stations missing up to nine days worth of data (see the DMFLAG description in ftp://ftp.ncdc.noaa.gov/pub/data/ushcn/v2.5/readme.txt). Depending on the month’s length, GHCN averages will be calculated despite missing up to a third of the data.
Comparison of Kingston USHCN and CRN Data
To test the skill of the TOBS adjustment NOAA applied to the Kingston USHCN site, a synthetic TOBS adjustment for the CRN site was calculated. The B91 forms for Kingston USHCN during 2003-2014 show a 4:30 PM observation time. Therefore, a synthetic CRN 4:30 PM observation was created by averaging the 4:00 PM and 5:00 PM observation data. The difference between the USHCN raw data and the synthetic CRN 4:30 observation is shown in the following figure. Despite a separation of approximately 50 meters, the two stations are producing very different results. Note that 2014 data is not included. This is because the 2014 USHCN data was incomplete at the time it was downloaded.
Although the data at the time of observation is very different, perhaps the adjustment from midnight (TOBS) is similar. The following figure represents the TOBS adjustment amount for the Kingston USHCN station minus the TOBS adjustment for the synthetic CRN 4:30 PM data. The USHCN TOBS adjustment amount was calculated by subtracting the USHCN raw data from the USHCN TOBS data. The CRN TOBS adjustment amount was calculated by subtracting the synthetic CRN 4:30 PM data from the CRN midnight observations. As can be seen in the following figure, TOBS adjustments to the USHCN data are very different than what would be warranted by the CRN data.
Adjustment Skill
The best test of adjustment skill is to take the homogenized data for the Kingston USHCN station and compare it to the midnight minimum / maximum temperature data collected from the Kingston CRN 1 NW station located approximately 50 meters away. This is shown in the following figure. Given the differences between the homogenized data from the USHCN station and the measured data from the nearby CRN station, it does not appear that the combined TOBS and homogenization adjustments produced a result that reflected real temperature data at this location.
Accuracy of Midnight TOBS
Whether the minimum and maximum temperatures are read at midnight or some other time, they represent just two samples used to calculate a daily average. The most accurate method of calculating the daily average temperature would be to sample continuously, and calculate an average over all samples collected during the day. The CRN stations sample temperature once every hour, so 24 samples are collected each day. Averaging the 24 samples collected during the day will give a more accurate measure of the day’s average temperature than simply looking at the minimum and maximum for the past 24 hours. This topic was covered in great detail by Lance Wallace in a guest post two and a half years ago. It is well worth another read.
The following chart shows the difference between using the CRN hourly temperatures to calculate the daily average, and the midnight minimum / maximum temperature. The chart tends to show that the hourly temperatures would produce a higher daily average at this station.
Discussion
Automated methods to adjust raw temperature data collected from USHCN stations (and by extension, GHCN stations) are intended to improve the accuracy of regional and global temperature calculations to, in part, better monitor trends in temperature change. However, such adjustments show questionable skill in correcting the presumed flaws in the raw data. When comparing the raw data and adjustments from a USHCN station to a nearby CRN station, no improvement is apparent. It could be argued that the adjustments degraded the results. Furthermore, additional uncertainty is introduced when monthly averages are computed from incomplete data. This uncertainty is propagated when adjustments are later made to the data.
A Note on Cherry-Picking
Some will undoubtedly claim that I cherry-picked the data to make a point, and they will be correct. I specifically looked for the closest-possible CRN and USHCN station pairs, with the USHCN station having a class 1 or 2 rating. My assumption was that their differences would be minimized. The fact that a second CRN station was located less than a mile away cemented the decision to analyze this location. If anyone is able to locate a CRN and USHCN pair closer than 50 meters, I will gladly analyze their differences.
================================
Edited to add clarification on y-axis units and meaning of the gray band to the description of each figure.
Edited to add links to the CRN and USHCN source data on the NCDC FTP site.
wpDiscuz