Urban Heat Islands and U.S. Temperature Trends

We use four datasets to classify stations as urban or rural, all with 1 km spatial resolution:

  • Satellite nightlights – how bright a specific location is at night as observed from space.
  • Impermeable surfaces (ISA) – what percent of the area is covered in concrete or similar materials.
  • GRUMP – an urban boundary database using administrative borders and other factors (including nightlights) produced by Columbia University.
  • Population growth – 1930 to 2000 population growth data interpolated to kilometer resolution using U.S. Census data.

We also used two different methods to compare urban and rural stations: a station pairing method, where we looked at all possible permutations of urban and rural stations within 100 miles (160 km) of each other for each urban proxy, and a spatial gridding method where we used a grid-based approach to calculate CONUS temperatures separately using only urban and rural stations and compared the results. For the station pairing method, we imposed additional restrictions that the pairs must both have the same instrument type, to avoid accidentally conflating bias due to urban-correlated differences in the frequency transition from liquid-in-glass thermometers to electric MMTS instruments with actual urban-related warming.

Finally, we examined six different versions of U.S. temperature data separately for both maximum and minimum temperatures:

  • Raw station data with no adjustments
  • Station data with only time-of-observation bias (TOBs) adjustments
  • Station data with both TOBs and full PHA homogenization
  • Station data with TOBs, full PHA homogenization, and GISTEMP satellite nightlight-based corrections
  • Station data with both TOBs and rural-only PHA homogenization.
  • Station data with both TOBs and urban-only PHA homogenization.

We created estimates of urban-rural differences for each of the four temperature proxies, two analysis methods, six temperature datasets, and maximum and minimum temperatures for a total of 96 different combinations.

As shown in Figure 2 from our paper, there are significant differences in the warming rate of urban and rural stations in the raw (and TOBs-adjusted) data that are largely eliminated by homogenization, even when that homogenization is limited to using only rural stations (to avoid the possibility of ‘spreading’ the urban signal).

This can also be seen in the figure below (from our paper’s supplementary information), which shows urban-rural differences over the 1895-2010 period using the spatial gridding method:

We conclude that homogenization does a good job at removing urban-correlated biases subsequent to 1930. Prior to that date, there are significantly fewer stations available in the network with which to detect breakpoints or localized trend biases, and homogenization does less well (though the newly released USHCN version 2.5 does substantially better than version 2.0). In general, there might be a need for additional urban-specific adjustments like those performed in NASA’s GISTEMP for areas and/or periods of time in which station density is sparse, but they are rather unnecessary for the post-1930s CONUS data. The simple take-away is that while UHI and other urban-correlated biases are real (and can have a big effect), current methods of detecting and correcting localized breakpoints are generally effective in removing that bias. Blog claims that UHI explains any substantial fraction of the recent warming in the US are just not supported by the data.

Page 2 of 3 | Previous page | Next page