by Zeke Hausfather
Measuring temperatures in the U.S. no easy task. While we have mostly volunteer-run weather station data from across the country going back to the late 1800s, these weather stations were never set up to consistently monitor long-term changes to the climate.
Stations have moved to different locations over the past 150 years, most more than once. They have changed instruments from mercury thermometers to electronic sensors, and have changed the time they take temperature measurements from afternoon to morning. Cities have grown up around stations, and some weather stations are not ideally located.
All of these issues introduce errors into the temperature record. To detect and deal with these errors, NOAA uses a process called homogenization which compares each station to its neighbors, flags stations that show localized changes in longer-term temperatures not found in nearby stations, and removes these local breakpoints. While the impact of these adjustments on temperature records is relatively small globally, in the U.S. it has a much larger effect due to the frequent changes that have occurred at our volunteer-run Historical Climatological Network (USHCN) stations (specifically time of observation changes and instrument changes). Fixes to errors in temperature data have effectively doubled the amount of U.S. warming over the past century compared to the raw temperature records.
A picture of the three redundant temperature sensors at a U.S. Climate Reference Network station.
To help resolve uncertainties caused by reliance on the historical network, NOAA began setting up a U.S. Climate Reference Network (USCRN) starting in 2001. USCRN includes 114 stations spaced throughout the U.S. that are well sited and away from cities. They have three temperature sensors that measure every two seconds and automatically send in data via satellite uplink. The reference network is intended to give us a good sense of changes in temperatures going forward, largely free from the issues that plagued the historical network.
While the USCRN will provide excellent data on the changing U.S. climate in the future, in the past we are stuck with the historical network. What we can do, however, is use the USCRN to empirically assess how well our adjustments to the historical network are doing. Specifically, since we know that the USCRN is largely free of issues, we can see if adjustments to USHCN are making station records more similar to nearby USCRN stations. That is the focus of our study recently published in Geophysical Research Letters (non-paywalled version here).
Figure 1 from Hausfather et al 2016.
For overall contiguous U.S. temperatures, the record from raw USHCN, adjusted USHCN, and USCRN are quite similar during the period of overlap, as shown in Figure 1 from our paper, reproduced above. USCRN does have a noticeably higher maximum temperature trend than both raw and adjusted USHCN data, though the cause of this is still unclear (see our paper for more discussion of this divergence).
We do see large differences between raw and adjusted USHCN data from individual stations when we compare them to nearby USCRN stations. Here we looked at all possible pairs of USHCN and USCRN stations within 50, 100, and 150 miles of eachother. The 100-mile case is shown in Figure 2 from our paper below, but all distance cutoffs have fairly similar results (and are available in our supplementary materials).
Figure 2 from Hausfather et al 2016.
In the vast majority of cases adjustments served to make the USHCN trends much more similar to those of proximate USCRN stations, particularly for larger divergences. Since we know that the reference network is largely free of measurement problems, this increases our confidence that adjustments are effective at finding and removing problems in the historical network.
Figure 3 from Hausfather et al 2016. Trend differences in mean temperatures are shown.
The spatial structure of variation in adjusted USHCN data is also much more similar to that seen in the USCRN, as shown in Figures 3 and 4 in our paper. Figure 3, above, shows the distribution of trend differences between USHCN and USCRN stations for raw and adjusted data, and compares them to the distribution of trend differences between proximate USCRN stations. While the adjusted USHCN trends are still slightly biased low (due entirely to differences in maximum temperatures), the shape of the distribution of USHCN adjusted trend differences is much more similar to that found between homogenous USCRN stations.
The Climate Reference Network was only established in 2004, so we can only directly test the adjustments for the recent period. However, the algorithm used to detect and adjust for problems in the historical network is applied equally over the past decade and the past century, so the fact that it seems to work well during the past 10 years increases our confidence that it also effectively deals with problems in the past, though this conclusion is somewhat tempered by the potential changing nature of inhomogeneities over time. Other work using synthetic data by Williams et al (2012) and Venema et al (2012) as well as comparisons to reanalysis data by Vose et al (2012) also suggest that adjustments are effective in removing localized inhomogeneities in the temperature record without introducing detectable spurious trend biases.
JC note: As with all guest posts, keep your comments relevant and civil.