44 thoughts on “UHI in California”

1. I will note with some amusement that the NCDC adjustment actually reduce California trends over the full period.
   .
   They also get rid of most urban-rural differences post-1960, but don’t do a particularly good job over the whole period for GRUMP and ISA proxies.

2. Very nice — data and presentation both.

   [Edit: Do the error bars represent the 95% confidence interval?]

   Zeke, of the 54 USHCN stations, the number that qualify as rural are 18, 36, 25, and 16, by the criteria listed.

   How many qualify as rural by all four criteria? Is the n high enough to enable statistically-significant calculations? I suspect it is — I’ll guess that there might be a dozen or half-a-gross among the 54.

   If so: what would the means be for the trends of 1895-2012 and 1960-2012, for these uber-rural stations?

   Maybe, before you ran the numbers (if you wanted to), Lucia could set up a betting pool. The smell of quatloos in the air…

3. AMac,

   The error bars are the 95% CIs of the trend (regressing anomaly against date using an AR1 method to deal with autocorrelation).

   I’ll check all four criteria. Would you want a comparison of those rural by all four criteria to those not rural by all four criteria, or to those urban by all four criteria?

4. Very nice presentation.

   Might I suggest one tiny improvement for your consideration: a one line English statement above about the 3 acronym-defined methods GRUMP, Nightlights, ISA so that they can be better understood without having to click through on the links?

   For example:
   GRUMP (Urban or Rural binary station designations by the Global Rural Urban Mapping Project – Columbia University)
   Nightlights (Percentage of Km^2 gridded areas lighted artificially; NOAA)
   ISA (Percentage of Km^2 gridded areas covered by Constructed Impervious Surfaces, pavement, buildings, concrete, etc.; NOAA)

5. Mickey Reno,

   Reasonable suggestion.

6. TMAX or TAVG?

   Interesting look at California cities via NCDC

   Los Angeles – Cooling from 1980 at -0.38 degF / Decade

   Fresno is flat from 1985 – .03F / decade

   How about trends from 1980/85 which tends to be the pivot point into cooling on the west coast?

7. Zeke,

   Thank you. A couple of thoughts:

   Could you post some version of raw data or subset of the analysis? txt file, csv or something? I realize it’s all publicly available, but would be cool to see nonetheless.

   California is interesting – big place, long coastline, mountains and desert, with population (and presumably temp stations) somewhat skewed to where the people are. A map would be awesome. I’d make one – i’m sure there are others on here who would do it better but I’d be happy to.

   I’m going with the ISA as most accurately reflecting what I believe to be UHI. And from a pure confirmation bias standpoint, it shows about the effect I would expect, based on a WAG.

8. Zeke: sorry for being obtuse, but is there a punchline in these graphs? I mean, Anthony would probably complain that adjustments warm recent rural stations trends, but other than that I’m not sure I can discern much.

   BillC:

   with population (…) somewhat skewed to where the people are.

   You bet! 🙂

9. Toto,

   There are pretty large and systemic difference between urban and rural trends across all urbanity proxies. Homogenization reduces differences over the last 50 years, but less so over the full century.

10. I couldn’t help but laugh when I saw the last figure. For the first two classification schemes, rural trends were lower before adjustments but were higher after.

    No matter what reasons you might be able to offer for it, that’s jarring.

11. “California is interesting – big place, long coastline, mountains and desert, with population (and presumably temp stations) somewhat skewed to where the people are.” –BillC

    The population is also somewhat skewed to where the ocean is. I guess Californians would rather have a view of the sea than of, say, Aguanga or Barstow.

12. Brandon,
    “it’s jarring”
    Yes, very much so. Since TOB has been removed in Figure 2, what we are left with to explain the substantial adjustments of the rural data seems to be either ‘homogenization’ of rural stations with more rapidly warming urban stations (absolutely indefensible, IMO), or station specific meta-data which shows why rural stations should be adjusted upward. It is the kind of thing that would seem to demand a very careful inspection. If rapidly warming urban stations are forcing rural stations upward via ‘homogenization’, then Anthony has a valid point, and the whole enterprise is questionable. If there are valid (substantive, non-homogenization, meta-data based, non-TOB) reasons for upward adjustments of the rural stations, then the meta data for each and every adjustment ought to be available for all to look at and consider.
    .
    Zeke,
    This really looks fishy to an old… er, somewhat senior scientist. 😉

13. Zeke, try 1980 on. What happens when the NOAA thinks the data is flat?

    http://i49.tinypic.com/9zr401.gif

14. SteveF,

    Well, it could be one of two things:

    1) There is a cooling bias over the last 50 years (say, the MMTS transition) affecting urban and rural stations similarly which is being corrected, warming both groups. At the same time, urban stations are being cooled via homogenization to correct for UHI.

    2) Homogenization is aliasing urban warming into rural stations, resulting in a warming bias.

    To test these two, one could run the homogenization algorithm using only rural stations, to avoid the possibility of aliasing in an urban signal during breakpoint correction. I’ve done this test myself, and will have more to report if the reviewers ever get back to me :-p

15. BillC,

    I’ll provide the raw data and station locations after work.

16. Re: Zeke (Aug 17 08:53),

    Arrgh, the Blackboard ate my comment.

    Shorter AMac — I think it would be fun to see Rural (x4) compared with non-Rural (x4), assuming the n’s are adequate. The difference between the trends would set an upper bound on the effect that urbanization/UHI could be having on the trends that have been reported.

    BTW, my bets would be:
    Rural x4 unadjusted, 1895-2012, 0.055 C/decade
    Rural x4 unadjusted, 1960-2012, 0.110 C/decade
    Rural x4 fully adjusted, 1895-2012, 0.050 C/decade
    Rural x4 fully adjusted, 1960-2012, 0.170 C/decade

    In other words, that any of the four criteria is about as discriminating as all four criteria applied together.

17. Zeke (Comment #101717),
    OK, if you have already done some work, why are you so shy to describe the results? The reviewers will never know what you tell us here. 🙂
    .
    But why not just look at the meta data? If there are adjustments for a change of station type, then it should be simple to verify that is the cause. I mean, why jump through your own…. OK, a better description… why make it any more difficult than it has to be? You can just look at the meta data a see if meta-data-based adjustments explain the rather large rural station adjustment. I don’t see any reason to bother with the homogenization algorithm to figure this out.

18. SteveF,

    That assumes that all major inhomogenities are documented in the metadata (sadly, they often are not, especially before 1980 or so). It also assumes that the impact of things like MMTS transitions is consistent. Unfortunately, when you combine an instrument location change with a new instrument you get a large variance in results: http://www.agu.org/pubs/crossref/2006/2006GL027069.shtml

19. Zeke, I like the graphs – they present a lot of information and clearly for old eyes to see.

    I did a quick and dirty analysis of US GHCN stations and used the GHCN classification for rural and urban and looked back as I recall to 1920. I modelled the temperature differences using all factors that might effect local temperature trends like elevation, latitude, and proximity to large bodies of water. I found a statistically significant difference between the adjusted temperatures trends with urban higher. I did both minimum and maximum.

    The hooker is that using the GHCN classification most of the stations are rural and thus the effect on the overall average was minimal.

    The paper under review – is it the study you were contemplating making to determine how well the GHCN algorithm homogenizes for UHI?

20. Kenneth,

    Pretty much (technically its looking at how well USHCN deals with UHI; GHCN is the next step though GHCN-M has too sparse a network to easily work with). You can see it in poster form here: http://www.geos.ed.ac.uk/research/earthtemp/themes/1_in_situ_satellite/Menne_EarthTemp_Edinburgh_2012_Poster.pdf

21. Zeke, it would appear from your graphs that the homogenization adjustment for rural/urban differences results in under adjustment. I had an email exchange pointing out to Menne that I can continue to find breakpoints in adjusted station temperature series through difference series with nearest neighbors and asking whether they had contemplated making a second iteration with their algorithm. He indicated that they were looking at it but at this point they did not want to over adjust.

    And, by the way did, you use altitude, proximity to water and latitude in your comparisons?

22. Zeke (Comment #101722),
    Ugg. If there is no uniform adjustment that can be applied for MMTS, and only sketchy meta data available for everything else, then it once again depends on analysis of ‘discontinuities’ in the individual station records…. and perhaps too much ‘expert opinion’ could be involved. It is beginning to sound like there is just not sufficient meta-data available, and we may be doomed to argue about adjustments to the raw temperature record forever. How sad. How frustrating.
    .
    How much could a couple hundred representative (pristine, unambiguous, perfectly sited, beyond doubt) set of stations cost? Certainly less than Solyndra or a single failed aerosol satellite launch. Multiply it to include the whole of Earth’s land surface area, and it is still a pittance.

23. Kenneth,

    For the poster/paper, latitude wasn’t really much of a concern since the network is so oversampled in the U.S. You get good coverage for both urban and rural in most places.

    We didn’t strictly correct for distance from coast/water, but we did use a station pairing approach that creates discrete urban/rural pairs within 100 miles of each-other with the same instrument type (and a MMTS transition date no more than 5 years apart). There were few significant differences between the results of the station pairing approach and the spatial gridding approach.

24. SteveF,

    We have USCRN, and there are efforts underway to create similar networks in other parts of the world. Unfortunately, they are less than useful for historical trends.

    I wound’t give up on automated homogenization. It should be possible to test how biased results are by seeing how they deal with various introduced inhomogenities in synthetic data (e.g. the Williams, Menne, and Thorne paper). We submitted an abstract to the AGU this year to do a similar comparison using the Berkeley approach, Menne’s PHA, and a new Bayes Factor approach to pairwise breakpoint detection that they are working on.

25. Toto, Jorge and anyone else who had a laugh at my expense, that’s what happens when I write too fast (or think faster than I write).

    What I meant to say – population is skewed to the coast, and I’m expecting the temp stations are as well.

26. Zeke,

    Thanks for the link to the poster. I will read and think about it.

27. Zeke (Comment #101728),

    The problem is that the urban/rural pairs in Caleefornia are almost certainly urban (coastal) rural (inland). There must be a better approach.

28. SteveF,

    I wouldn’t be too sure, since a good portion of the coast here in CA is undeveloped (state park), especially in the northern part of the state. Note that the graphs in this post are only using spatial gridding; the discussion of station pairing is more germain to the poster I linked earlier.

29. BillC/SteveF

    I uploaded the metadata (see update in the OP) if you want to play around with it. I’ll stick up the station records as well shortly.

30. Have you controlled for latitude or altitude?

    Nice to be able to comment here again! 🙂

31. Andrew,

    Kenneth already asked about that (well, latitude). For this particular project, I just did a fast and dirty spatial gridding analysis.

    If anyone is interested in the unprocessed USHCN data, it is available here: ftp://ftp.ncdc.noaa.gov/pub/data/ushcn/v2/monthly/

    The station_ids in there are the same as those in my metadata file, if you want to only keep the CA stations.

32. Zeke (Comment #101729)

    “I wound’t give up on automated homogenization. It should be possible to test how biased results are by seeing how they deal with various introduced inhomogenities in synthetic data (e.g. the Williams, Menne, and Thorne paper). We submitted an abstract to the AGU this year to do a similar comparison using the Berkeley approach, Menne’s PHA, and a new Bayes Factor approach to pairwise breakpoint detection that they are working on.”

    I agree. If you can come up with simulated data where the truth is known and those data are representative of what we know about the factors that can affect temperatures you would have a foolproof testing system. Obviously it is those factors that we do not know about or imperfectly know about that can be a problem. I would like to see some of these benchmark tests devoted to what if propositions where a possible problem, but not established as a problem, or maybe even not possible currently to establish as a problem, is introduced into the simulation. If the result is that the effect is minimal we move on, while if it is large, we might want to investigate it further.

    What makes an objective/automated system attractive is that it would be easier to test with simulated data. Meta data in a simulated test would have to be assumed to be correct or sufficiently in place – and that is a big assumption.

    I agree with SteveF about looking at some of these problems by building stations where we can change the micro conditions or at least control them. I think scientists in general tend not to tackle problems that way but would rather use existing data. Building those stations would be more along the lines of an engineering approach – although the scientist who needs to generate her own data can become very interested and adept in building equipment to that end.

33. Very interesting comparison Zeke. I had not seen this before so thank you for posting.

    You are a lot closer to this than I am and it may take me a while to absorb it all. Homogenizing between urban and rural temps is not an intuitive method for addressing UHI in my view. Do you think that the NOAA homogenization process being used here effectively deals with the UHI problem or does this introduce a broader warming bias to all stations.

34. ivp0,

    On the subject of how NOAA’s homogenization deals with UHI, and the extent to which it might be “spreading” the urban signal, see this: http://www.geos.ed.ac.uk/research/earthtemp/themes/1_in_situ_satellite/Menne_EarthTemp_Edinburgh_2012_Poster.pdf

    There is also a more extensive version of this currently under review, which I’m not really at liberty to discuss at the moment.

35. Zeke,

    Got it. Thanks!

36. Zeke,

    I have played around with the data you sent. I hope you’re still reading this page as I want to ask some questions.

    First, I tried to re-create the trends you showed in Fig. 2 above using the data files you posted. The results are here. The blue diamonds in my figure overlay directly onto your averages for all but the GRUMP urban stations. Not sure why this is, but I also had to switch ISA for Nightlights with respect to the column headings in your data file, in order to recreate your graph.

    I added the red squares to my graph to recalculate the trends per your data from 1960-2007, since this appears to be what is available on the
    USHCN ftp site.

    Now, I tried to create a parallel analysis with a different method. I wanted stations that had some sort of continuity over the 1960-2007 period. I used the NOAA’s annual averages from the tob data, and calculated 5-year average for each station. If a station had at least 1 annual average report in each interval (e.g. 1960-1964, etc.), it got a 5-year average. Of the 54 stations, 37 met this criterion. I used OLS to come up with a 1960-2007 trend in K/decade for each station separately. Of course, this somewhat over-weights 2005-2007, treating it as 5 years.

    I then tried to re-create your analysis using averages of these trends. I used simple averages which of course doesn’t take into account spatial coverage. I did this for both tob and fully adjusted data. Doing this resulted in this graph. Most categories’ trends were lowered by doing this, which is interesting, though the fully adjusted data didn’t change much.

    Station counts when doing the above:
    GRUMP: 27 urban, 10 rural
    Population: 27 urban, 10 rural (with a different distribution)
    ISA: 20 urban, 17 rural
    Nightlights: 12 urban, 25 rural

    Obviously for the above, I don’t know exactly how much of the difference is made by the choice of stations, and how much is made by the lack of spatial weighting. Also, for the fully adjusted data, all the annual averages are infilled by whatever method NOAA uses, whereas there are blanks in the tobs dataset.

    I can think of lots of ways to play with this. Since there is some discussion above about 1) combining urbanity proxies and 2) looking at instrumentation changes, I started with a combination of both. the results are interesting. From your metadata and my continuity analysis I got the following stats:

    Of the 37 stations that satisfied my continuity criteria, 16 were CRS station type with no instrumentation change listed (which makes sense). The MMTS stations usually had a year listed. I would guess that there should generally not have been any station moves among this group, but maybe you know.

    From the 16 continuous CRS records, here is the breakdown by sum of urbanity proxy type for that station (i.e. 0 is rural in all 4 categories, 4 is urban in all 4 categories)

    Type; # stations; tob mean; adj mean
    CRS,U=0; 4; 0.10; 0.16
    CRS,U=1; 2; 0.04; 0.16
    CRS,U=2; 2; 0.25; 0.25
    CRS,U=3; 3; 0.11; 0.14
    CRS,U=4; 5; 0.26; 0.17
    CRS, All; 16; 0.16; 0.17

    This seems to need explaining to me: Of stations with some decent continuity during the years 1960-2007, adjustments other than tob, to CRS type stations, in CA, warmed rural stations to the mean, cooled urban stations to the mean, and raised the mean a tiny bit.

    Since this is such a small sample, I would love to see what the result of this for the whole CONUS is. Is the metadata you posted, available for the whole country in that same format somewhere? If it’s on NOAA, I didn’t see it, but I will look.

37. BillC.

    you can get all that metadata by using the package
    Metadata which is hosted on CRAN. you will have to
    pull CRS versus MMTS from NOAA.

    or zeke , may post his whole datasheet.

    Of all these proxies GRUMP is the least reliable and has the worst producer accuracy. It is essentially administrative boundaries.

38. Steve Mosher,

    Thanks. Depending on what Zeke says I will go grab the metadata for the whole US, as now that I have the analysis set up it won’t be hard to repeat it. Depending on what that turns up and what other comments I get, it might be something worth taking further.

39. Bill C,

    Work is busy today so I won’t have time for a complete response later, but as far as the differences in our analysis go make sure you are using 2.5×3.5 lat/lon grid cells and an AR(1) autocorrelation method in the regression.

    The full USHCN metadata is here: http://rankexploits.com/musings/wp-content/uploads/2012/08/ushcn_updated_metadata.csv

    The code I use for gridding is:

    *Assign each station to a grid cell
    gen grid_size_lat = 2.5
    gen count = -90 + (grid_size_lat / 2)
    gen latgrid = .
    forvalues latcoords = -90(2.5)90 {
    replace latgrid = count if (lat > `latcoords’ | lat == `latcoords’) & lat < (`latcoords' + grid_size_lat) replace count = count + grid_size_lat } gen grid_size_lon = 3.5 replace count = -180 + (grid_size_lon / 2) gen longrid = . forvalues loncoords = -180(3.5)180 { replace longrid = count if (lon > `loncoords’ | lon == `loncoords’) & lon < (`loncoords' + grid_size_lon) replace count = count + grid_size_lon } *Create a unique gridbox identifier gen latseperator = " lat " gen lonseperator = " lon" egen gridbox = concat(latgrid latseperator longrid lonseperator) drop latseperator lonseperator *Create a grid weight based on latitude (formula courtesy of RomanM) gen grid_weight = sin((latgrid+grid_size_lat/2)*_pi/180) - sin((latgrid-grid_size_lat/2)*_pi/180)

40. Zeke, Steve Mosher,

    Based on the total USHCN metadata I re-calculated the individual station trends for the CRS shelters nationwide using the same method (5-year averages, trends if all 5-year periods had at least one year reported), here are the stats:

    total USHCN stations satisfying the criteria – 920
    cutoffs; GRUMP is binary, ISA 10, pop 10, nightlights 30

    type; # stations; raw mean; tob mean; adj mean
    CRS,U=0; 54; 0.21; 0.26; 0.27
    CRS,U=1; 21; 0.22; 0.28; 0.28
    CRS,U=2; 16; 0.21; 0.26; 0.27
    CRS,U=3; 17; 0.21; 0.24; 0.25
    CRS,U=4; 24; 0.27; 0.28; 0.27
    CRS, All ; 132; 0.22; 0.26; 0.27

    Based on the high similarity between the tob mean and adj means across all urbanity scores, which is unlike California, I’m not sure if that bears more looking into.

    Given that earlier it was said that tobs adjustments usually corresponded with an instrumentation change, it’s mildly interesting that the most urban stations received very little tobs adjustment compared to the other types.

    Zeke – I understand how you are creating the grid assignments and grid weights but what are you doing to average over the grid? Are you just calculating a weighted average of grid cells having at least one temperature station (after calculating internal grid cell averages) or are you doing some sort of spatial interpolation?

    BTW the nightlights seems to be the strictest criteria for considering something “urban”. by contrast GRUMP is the strictest for classifying something “rural”.

41. Er, the 920 above is all stations with the continuous trend as defined above, regardless of type. Only 132 are CRS per the table.

42. BillC,

    Tobs changes tend to predominantly occur at co-op stations. Urban stations are less likely to be volunteer-run, and more likely to have hourly measurements that do not require TOBs.

    As far as averaging over grids go, I convert the absolute temps into anomalies using a consistant baseline period, average all anomalies by gridcell for each month, and weight grid cells by their surface area in the national reconstruction (less important for CONUS, very important for global analysis).

43. AMac (Comment #101720)
    August 17th, 2012 at 1:48 pm

    Arrgh, the Blackboard ate my comment.

    Lazarus Add-On. Lose nothing. Ever. Save yerself scadoodles of time an’ aggervation. Promise.

44. Zeke,

    As far as averaging over grids go, I convert the absolute temps into anomalies using a consistant baseline period, average all anomalies by gridcell for each month, and weight grid cells by their surface area in the national reconstruction (less important for CONUS, very important for global analysis).

    Sure, but what do you do for grid cells that don’t have any stations? There are sure to be many, and so we have many different interpolation algorithms…

Comments are closed.