The recent post at WUWT covered a new analysis by Goddard & Tett (hereafter GT) showed how UHI has biased measurements in the UK. The paper concludes:
For an urban fraction of 1.0, the daily minimum 2‐m temperature was estimated to increase by 1.90 ± 0.88 K while the daily maximum temperature was not significantly affected by urbanisation. This result was then applied to the whole United Kingdom with a maximum T min urban heat island intensity (UHII) of about 1.7K in London and with many UK cities having T min UHIIs above one degree.
This paper finds through the method of observation minus reanalysis that urbanisation has significantly increased the daily minimum 2‐m temperature in the United Kingdom by up to 1.70 K.
The paper represents a trend in UHI studies toward using urban area or urban fraction to define areas as urban and to parameterize the effect: to express UHI as a function of urban area: This is in contrast to the early studies, for example, Oke (73) that tended to use population to parameterize UHI
Since Oke there has been considerable progress in understanding the complex phenomena of UHI and the science has moved beyond the simple approach of looking at population as a parameter that uniquely determines UHI. If everyone leaves a city, it will still have UHI.
Recently, at WUWT the following claim was made
“The present situation is one of large, continuing lack of research attention. There is not even a detailed description of how large the UHI effect is, using a representative set of city examples, let alone its uncertainty.”
This is actually not the case. This is a tiny fraction of the types of studies done.
And there are a growing number of papers (here, here, here ,here, ) that detail urban cool parks that may explain why UHI is so difficult to find the global record. Sites located in cities are not necessarily warmer than those in rural setting.
One of the most important advances has come in the area of quantifying the definitions of urban and rural. Oke and Stewart have transformed the field with their concept of the LCZ or local climate zone. Anyone who took pictures of temperature stations for Anthony’s surface station program will enjoy watching the entire video below and especially the parts after 23 minutes where microsite bias is discussed.
And now with the power of satellite imagery researchers can quantifiably categorize various type of urban/rural areas. This can be done automatically or manually: http://www.wudapt.org/lcz/ Stewart was motivated to do this categorization in part because a large number of urban/rural studies never objectively defined the difference between urban and rural and because they assumed that “urban” was a discrete category rather than a continuum.
GT found that the UHI effect in the UK was limited to biasing Tmin upwards, a result consistent with other findings. Wang (2017) looked at 750+ stations in China and also found a bias in Tmin of up to 1.7C at 100% urban cover. A figure that matches the result of GT.
Trends in urban fraction around meteorological station were used to quantify the relationship between urban growth and local urban warming rate in temperature records in China. Urban warming rates were estimated by comparing observed temperature trends with those derived from ERA-Interim reanalysis data. With urban expansion surrounding observing stations, daily minimum temperatures were enhanced, and daily maximum temperatures were slightly reduced. On average, a change in urban fraction from 0% to 100% induces additional warming in daily minimum temperature of +1.7 +- 0.3°C; daily maximum temperature changes due to urbanization are -0.4 +-0.2°C. Based on this, the regional area-weighted average trend of urban-related warming in daily minimum (mean) temperature in eastern China was estimated to be +0.042 +- 0.007 (+0.017 +- 0.003)°C decade1 , representing about 9% (4%) of overall warming trend and reducing the diurnal temperature range by 0.05°C decade . No significant relationship was found between background temperature anomalies and the strength of urban warming.
To many readers the maximum bias figure of 1.7C in Tmin at 100% urbanity may seem low, especially when you consider the figure at the top from Oke which shows a UHI of up to 8C. The difference lies in the methodology. Much of the early work done on UHI focuses on UHI max for any given day. They select conditions that show the largest values of UHI that can occur. Oke’s chart, for example, represents the maximum value of UHI observed on a given day. For example, he would select summer days with no clouds, and no wind and measure the max difference between a rural point of reference and a city point of reference. In the studies that show high UHI values they typically do not calculate the effect of UHI on monthly Tavg over the course of many years, as GT and Wang did. Since cloud free wind free days do not occur 365 days a year for years on end, the overall bias of UHI is thus lower for monthly records, annual records, and climate records. In one study the number of ideal days in a year for seeing a difference between urban and rural was 7 days of the year. A 40 year study of London nocturnal UHI, found that the average UHI was ~1.8C, and only 10% of the days experienced UHI over 4C. In short, Average monthly UHI is less than the maximum daily UHI observed at optimum conditions for UHI formation.
The current best estimate by the IPCC is that no more than 10% of the century trend for Tavg is due to UHI and LULC. If we take the century trend in land temperatures to be 1.7C per century, for example, then the 10% maximum bias would be .17C on Tavg. The IPCC does not make an independent estimate for Tmin or Tmax, only Tavg, because the major analysis products only use Tavg.
In summary, it is indisputable that UHI and LULC are real influences on raw temperature measurements. At question is the extent to which they remain in the global products (as residual biases in broader regionally representative change estimates). Based primarily on the range of urban minus rural adjusted data set comparisons and the degree of agreement of these products with a broad range of reanalysis products, it is unlikely that any uncorrected urban heat-island effects and LULC change effects have raised the estimated centennial globally averaged LSAT trends by more than 10% of the reported trend (high confidence, based on robust evidence and high agreement). This is an average value; in some regions with rapid development, UHI and LULC change impacts on regional trends may be substantially larger.
Both GT and Wang look at the urban fraction over a 10km buffer surrounding the station. This is probably at the radius limits of the LCZ. There is no “typical” range for LCZ analysis, but in general analysts consider the zones 1 to 10km in size. In LCZ analysis the fraction of imperious surface is one of the quantifiable features that determine the LCZ type. In general, urban fraction divides LCZ thusly:
A) Areas with less than 10% impervious surface are “unbuilt”
B) Areas with 10-20% impervious surface are sparsely built
C) Areas with 20+ % built are what we would typically call urban
There are some notable exceptions to this, in particular some heavy industry areas may have small urban fractions less than 10%. From field testing we know that different LCZ zones have different temperatures. See table 2 here for a study of LCZ in Berlin over the course of a year.
Armed with this metric we can begin to classify temperature stations by the percentage of urban fraction in their local climate zone. In theory we don’t have to make a bright line distinction between rural and urban, but rather we have a metric for the relative urbanity of a site that goes from 0% impervious surface in the LCZ to 100%.
In Berkeley Earths study of UHI we broke some ground by being the first study to use satellite data for urban surface to classify the urban and the non urban. We used a MODIS data set with a 500m resolution. However, two things concerned me about that dataset: 1) the imagery was taken during northern hemisphere winter and could falsely classify snow covered urban as rural. 2) the true resolution was more like 1km as a pixel wasn’t defined as urban unless 2 adjacent 500m pixels were urban. 1kmsq is not a small area. To accommodate for this and to accommodate for location errors we looked at 10km radius around each site and a site was classified as Non rural if it had 1 urban pixel. Our results found no difference in trend between urban and non urban. Still, the 1 km sq resolution bothered me. We can now address that issue with higher resolution data.
Available satellite imagery has expanded since the publication of that paper and much more accurate data is now available. GT used 250m data, for example and “paywalled” data is available below 30meter resolution. For my study of GHCN version 4 metadata I considered two different sources:
A) ESA 300 meter data
B) 30 meter data made available here http://www.globallandcover.com/GLC30Download/index.aspx.
Each dataset has pro’s and cons. The 30 meter data is quite voluminous and comes in tiles complicating the process of determining urban fraction. The 300 meter data is easier to work with but doesn’t really work very well if you want to know what the surface is like within 100 meters of the station. It cannot work well for microsite analysis. Also, neither dataset is perfect. Every land classification system has errors: natural pixels (typically bare earth) that are classified as urban, and urban pixels that are misclassified as natural. It’s helpful, thus, to compare the 30meter data with the 300 meter data and to cross check both with other signs of urbanity such as population and night lights.
GHCN v4 will be the next land dataset published by NOAA for use in global average temperature studies. It is currently in beta and going through a validation and verification process. NASA GISS will adopt adjusted GHCN v4 as its primary data source for global land temperatures. And then they will apply their UHI correction which in practice does not reduce the trends in any substantial way. The number of stations in GHCN V4 has increased over V3 to more than 27,000 total stations. The dataset will come in two variants: Uncorrected by NOAA; and debiased by NOAA’s PHA algorithm.
To create enhanced metadata for this new set of stations the procedure is fairly straightforward. You take the latitude and longitude of the station and then locate it in the appropriate GIS dataset. For 30meter data which exists in UTM tiles, you have to re-project and stitch 2 tiles together to handle cases where a station may be located near to a tile border, or 4 tiles together when a station is located near a tile corner.
For every station we can create “buffers” or collections of all the land class within various radii. For this post I’ll report on the 10km radius to be consistent with GT and Wang who also look at 10km buffers.
One important note. The purpose of this is not to assess the specific site micro characteristics: surface properties within 0- 500 meters that are within the viewshed of the sensor. Rather I will look at the LCZ, the local area climate zone out to 10km and answer the question: just how urban are the temperature stations used by climate scientists who study the global average? Do we actual draw our samples from heavily urban areas as defined by Oke’s and Stewart’s LCZ classification system.
The map from GT is instructive here
Are the stations that will be used by NASA GISS in red zones or in blue zones? What fraction are in red? And what fraction are in blue areas? And what shade of blue?
Some other things to note. The land classification data is taken at 2015 for 300 meter data and 2010 for the 30 meter data. Underlying this analysis is the assumption that site areas are not “unbuilt” over time. I assume a station that shows 0% built area in 2010 did not have any built area before that time. One other subtlety that people miss is that stations that register as heavily built in 2015 may have been rural during their recording time. For example, you can have a station that reports temperatures for 1850 to 1885, and then stops reporting. The urban fraction data refers to the urban cover of that site at 2015 or 2010. If you simply classify this site as urban, it may not be accurate as you are interested in the temperature data that was collected in the 1850 to 1885 time period. If the station was rural during that period, and you classify it as urban because of its urban cover today, then you can confound urban/rural studies.
Using the same criteria as GT and Wang (2017) we can see that the vast majority of stations are located in LCZ’s that have less than 10% urban cover (blue line below).
The using 30 meter data results in slightly fewer stations in the 0-10% ranking. This is to be expected as 300 meter data is not small enough to detect roads or airport runways while 30 meter data can in most cases. Using the regression approach of GT and Wang, we can also make a first order estimate of the size of the Tmin bias in a global record constructed from stations with this magnitude of urban cover: ~.13C. This would translate into a ~.06C bias in Tavg, within the estimate made by the IPCC. Note this is a simplistic estimate that does not take the spatial distribution of the stations into account, and it could be higher, or lower, but not substantially.
One thing to note is that we are able to check how robust the procedure of looking at 10km buffers around the site is by using the same procedure with CRN stations which have been selected to minimize their urban exposure: over 95% of CRN stations have less than 10% urban cover within a 10km radius of the site.
The big picture takeaway is this. UHI studies like GT and Wang focus on UHI over long periods of time: years instead of days. When you just focus on UHI max during selected days at selected cities, you will get high values for max UHI. However, when you look at dozens to hundreds and thousands of stations over months and years, the bias figures for UHI drop substantially. It’s these figures that matter for UHI bias in the global land record. Further when you look at all the stations in the inventories rather than the worst cases, you see that the vast majority of stations are located in areas of low urban cover:0-10%
This brings me to my last two points. While the fraction of urban cover within a 10 km radius does give you comparability with GT, it misses two things. These two things could be more important and I think they deserve some more attention. Those issues are: UHI in small area towns and microsite bias. The potential UHI issue in the global record is not a large city issue. The charts above should tell you that. Areas with large dense urban cover do not dominate the inventories of stations. They just don’t. The more plausible cause of UHI in the global record would come from small areas of urban cover. It’s unfortunate that most people focus on the photos of large cities and the papers about large cities, when actually, the problem may be smaller cities, at least as the global record is concerned. My suggestion is to aim at the right target with your analysis and critiques.
The second issue is the issue of microsite. Wang 2017 wrote
Changes associated with urbanization may impose influences on surface-level temperature observation stations both at the mesoscale (0.1–10 km) and the microscale (0.001–0.1 km). For a specific observing station, small local environmental changes may overwhelm any background urban warming signal at the mesoscale. Due to the lack of a high-quality data set of urban fraction at the microscale, we can hardly quantify the microscale urban influence on the observed temperatures.
In other words, the metadata that matters most is the metadata of the first kilometer. A good site in an urban setting can be better than a bad site in a rural setting. My bet is this: if you expect to find bias in the record, you should be looking at that first kilometer. Microsite is more important than UHI.