Monsters Lurking in the Deep - or Some Oddities in a few Homogenised Gridbox Averages.

The Issue

HadISDH.landT has been compared with other land air surface temperature monitoring products (CRUTEM4, GHCNM3, Berkeley, GISS). Agreement is generally very good although HadISDH.landT shows smaller warming trends overall. During this comparison a handful of gridboxes were highlighted that contained strong disagreement between products. On further investigation, it is clear that inhomogeneities remain in all three datasets (HadISDH, CRUTEM and GHCNM - others not compared for this assessment) at various points in time and space. Analysis of the HadISDH homogenisation at the gridbox level has revealed some oddities - mostly due to intermittent temporal sampling of stations such that the gridbox average suffers drop-in and drop-out of stations. This has resulted in spurious changes in variance and possibly means in some cases. While this does not affect conclusions drawn on large scale averages, it does show that users of data at the individual gridbox level should be careful. It also shows that Climate Data Record creation is far from a 'done deal'. There is so much more to explore and learn and this can greatly enhance our understanding of the climate system and our confidence (reduce uncertainty) in main conclusions.

Bad Gridboxes

For HadISDH vs CRUTEM there are 13 gridboxes where trends are in different directions (Figure 1 - pink and red boxes) and for HadISDH vs GHCNM there are 11 (Figure 2 - pink and red boxes). These are listed below with the number of stations from each dataset in each gridbox (H=HadISDH, C=CRUTEM, G=GHCNM). Where 'H/C' is labelled this means there is poor agreement between HadISDH and CRUTEM. Where 'H/G' is labelled this means there is poor agreement between HadISDH and GHCNM. The dataset with the lowest number of stations is highlighted in bold as this suggests greater uncertainty/sensitivity in that dataset. All gridboxes with each dataset only having one station are highlighted in purple.

Northern Hemisphere (20N to 70N)
-172.500W, 57.500N H/G, H=1, C=1, G=1 
-167.500W, 62.500N H/G, H=2, C=2, H=1 

Tropics (20S to 20N)
-157.500W, 17.500N, HG, H=1, C=2, G=1
 -57.500W, 7.500N, H/C, H=2, C=7, G=11
157.500E, 7.500N, H/C, H=1, C=3, G=1 
-82.500W, -7.500S H/C, H=1, C=1, G=1
177.500E, -12.500 H/G, H=1, C=2, G=2
 -67.500W, -17.500S H/C, H=2, C=14, G=5
147.500E, -17.500S H/G, H/C, H=4, C=14, G=10  

Southern Hemisphere (70S to 20S)
-132.500W, -22.500S H/C, H=1, C=1, G=1
147.500E, -22.500S H/C, H/G, H=1, C=13, G=13 
122.500E, -27.500S H/C, H=1, C=3, G=6
112.500E, -27.500S H/C, H/G, H=2, C=2, G=3
-62.500W, -42.500S H/C, H=1, C=3, G=2 
167.500E, -47.500S H/C, H/G, H=1, C=2, G=1
-67.500W, -52.500S, H/C,H/G, H=3, C=7, G=4
112.500E, -67.500S H/C, H/G, H=1, C=1, G=1
 37.500E, -67.500, H/G, H=1, C=1, G=1

Figure 1 Decadal Trend Ratio Map of HadISDH compared to CRUTEM4. Trends are estimated using the median of pairwise slopes. Pale blue boxes show where CRUTEM4 has larger (further from zero) trends than HadISDH and in the same direction. Dark blue boxes show where CRUTEM4 has smaller (closer to zero) trends than HadISDH and in the same direction. Yellow boxes show where trends are very very small - often resulting in very large trend ratios. Pale pink boxes are where trend directions disagree and CRUTEM4 trends are smaller in absolute magnitude. Red boxes are where trend directions disagree and CRUTEM trends are larger in absolute magnitude. Red box identify those boxes that have been assessed in more detail.

Figure2.

Figure 2 Decadal Trend Ratio Map of HadISDH compared to GHCNM. Trends are estimated using the median of pairwise slopes. Pale blue boxes show where GHCNM has larger (further from zero) trends than HadISDH and in the same direction. Dark blue boxes show where GHCNM has smaller (closer to zero) trends than HadISDH and in the same direction. Yellow boxes show where trends are very very small - often resulting in very large trend ratios. Pale pink boxes are where trend directions disagree and GHCNM trends are smaller in absolute magnitude. Red boxes are where trend directions disagree and GHCNM trends are larger in absolute magnitude. Red box identify those boxes that have been assessed in more detail.

Gridbox -57.5W, 7.5N Case Study

The time series for the gridbox centred on -57.5W, 7.5N is shown in Figure 3. It is clear from this figure that at the gridbox level there can be considerable disagreement between datasets - CRUTEM shows no trend but GHCNM and HadISDH show significant warming trends. We may expect GHCNM and HadISDH to be quite similar as both use the Pairwise Homogenisation Algorithm (PHA) to make adjustments for inhomogeneities. HadISDH applies extra adjustments through indirect PHA utilising detected changepoints from simultaneous dewpoint depression records. CRUTEM does not apply any homogenisation but many of the stations ingested are pre-homogenised at the National Meteorological Service level. It is quite possible that inhomogeneities remain in all datasets and it looks like CRUTEM contains a large inhomogeneity beginning around 2008. However, it is not clear that the lack of any such inhomogeneity in HadISDH and GHCNM is due to homogenisation because the difference series in the lower panel does not show an obvious negative adjustment at that time. However, the HadISDH time series has generally undergone negative adjustments throughout the record.

Figure 3 Monthly mean surface temperature anomaly time series for gridbox -57.5W, 7.5N. Data are shown from HadISDH.landT, CRUTEM4 and GHCNM. Decadal trends (and uncertainty ranges) are shown in the top left. Correlations are shown in the bottom right. The lower panel shows the difference series of HadISDH.landT (adjusted) minus HadISDH.landT (raw). This has been normalised to have a mean of zero of the last five years because no adjustments are applied in the last two years of data. The orange line is a lowess smoothed fit using 2 years of data to apply smoothing. This shows the aggregated impact of station homogenisation at the gridbox level.

A more detailed look at the HadISDH.landT time series from both the adjusted version and the raw version shows some interesting things. Figures 4 and 5 show the raw and adjusted time series respectively. This reveals that homogenisation has actually increased the variance over two periods in particular relative to the raw version. This seems odd because any adjustments applied during homogenisation are flat (non-varying seasonally) and can only be applied with a maximum frequency of 6 months. I think what we're seeing here is temporal intermittency in the underlying station records having an affect on the gridbox average - where a gridbox average is only ingesting three or fewer stations it is very sensitive to temporal drop-in/drop-out.

 

Figure 4 Grainy ncview dump of HadISDH.landT (raw) for gridbox -57.5W, 7.5N. These are monthly mean anomalies of surface temperature.

 

Figure 5 Grainy ncview dump of HadISDH.landT (adjusted) for gridbox -57.5W, 7.5N. These are monthly mean anomalies of surface temperature.

HadISDH only has two stations in this gridbox where as CRUTEM has 11 and GHCNM has 7. This means that the HadISDH gridbox average is very sensitive to temporal drop-in/drop-out of any one station. Indeed, a look at the number of stations contributing to that gridbox over time shows exactly that (very intermittent temporal sampling), especially over the periods of high variance in the difference series (e.g., 1973-1982, 1999-2005) - see Figure 6. Obviously this isn't ideal as far as climate monitoring is concerned - we require long-term stability. For future versions of HadISDH we may wish to have some intermittency check - removing isolated months of data.

Figure 6 A rather grainy dumped image from ncview showing the contributing number of observations for each month to gridbox -57.5W, 7.5N for HadISDH.landT.

We can look at the individual stations contributing to this gridbox and see the adjustments applied (sadly only as annual values because these are quickly grabbed plots). Figures 7 and 8 show the two HadISDH stations (812020, -57.033W, 5.95N, Nickerie; 812250, -55.2W, 5.45N, Zanderij) before (red) and after (blue) homogenisation, along with all other unhomogenised stations in the neighbour network for that station (black). Note that these two stations are also in CRUTEM and GHCNM. Station 812020 has platform style adjustments where the early period appears to be homogeneous with the most recent period. This can happen if something changes (e.g. the station is moved) and then returned to the original status (e.g. the station is moved back to where it was previously). Station 812250 has adjustments accumulating back in time. Interestingly, all adjustments for this station are cumulatively in the same direction - the raw data have been adjusted downwards, increasing the long-term trend.

Figure 7 Annual mean surface temperature time series for station 812020 (Nickerie) for the raw (red) and homogenised (blue) data and all raw neighbours within the station network (black). Decadal trends and 90% uncertainty ranges are shown for the raw (red) and homogenised (blue) series.

Figure 8 Annual mean surface temperature time series for station 812250 (Zanderij) for the raw (red) and homogenised (blue) data and all raw neighbours within the station network (black). Decadal trends and 90% uncertainty ranges are shown for the raw (red) and homogenised (blue) series.

On analysis of the monthly mean series (not shown) these adjustments do not look too outlandish but it is also clear that they are not perfect. The breakdown is listed below for each station. b = both PHA and IDPHA implemented adjustments. i = IDPHA implemented adjustments. p = PHA implemented adjustments.

Station 812020:
Start Month, End Month, Actual adjustment, Cumulative adjustment, Type
464, 504, 0.00, 0.00, b
348, 463, 0.45, 0.45, i
305, 347, 0.05, 0.50, p
1, 304, -0.50, 0.00, b
 
Station 812250:
Start Month, End Month, Actual adjustment, Cumulative adjustment, Type
462, 504, 0.00, 0.00, p
385, 461, 0.25, 0.25, p
320, 384, 0.50, 0.75, p
258, 319, -0.46, 0.29, i
239, 257, -0.16, 0.13, p
218, 238, 0.37, 0.50, i
1, 217, 0.10, 0.60, b
 
So, in summary, the main reasons for differences between HadISDH, CRUTEM and GHCNM are station selection, number of contributing stations and their temporal intermittency and also homogenisation. The large inhomogeneity in CRUTEM in 2008 is odd given that there are 11 stations contributing which I would have thought would moderate this to some extent.  

 

Other Gridboxes of Interest

Out of the 39 gridboxes analysed in detail, either because the trend ratios were negative or the correlations were very low, a good number show interesting features. I have added the ones which I think show the most interesting things below:

Figure 9 - as Figure 1

The HadISDH gridbox time series in Figure 9 has been adjusted upwards relative to the raw time series resulting in a drastically different trend compared to CRUTEM. Similar magnitude jumps are not so apparent in CRUTEM except for the 2006 to 2011 period. This suggests CRUTEM has been adjusted or inhomogeneities were not present to begin with. Also, averaging over the 14 CRUTEM stations reduces sensitivity to inhomogeneities compared to having only 2 stations to average over in HadISDH.
Trends are very different.

Figure 10 - as for Figure 1.

There is a clear and large negative adjustment applied to HadISDH between 1973 to 1980 in Figure 10. For this gridbox all datasets only have one station and it appears to be the same station. Month-to-month variability is very similar suggesting these are the same station and version of the station. So, it ooks like GHCNM has also applied an adjustment at the beginning of the series - one that is larger than for HadiSDH. It looks like CRUTEM has not had any adjustments in this case.
Trends are very different.

Figure 11 - as for Figure 1.

The gridbox shown in Figure 11 is made up of a single station in all cases - this appears to be the same station for each dataset but there are subtle differences which could be due to a different version being used or something related to the method used to calculate monthly means (HadISDH is built from hourly data) or related to the data removals from HadISDH QC which would also affect the calculated monthly means. This difference is particularly noticeable in 1996. Interestingly, the upwards adjustment applied to HadISDH (and quite possibly GHCNM) does not go far enough to match up with CRUTEM. Given the good agreement for the rest of the time series this suggests that CRUTEM has had an adjustment applied but that it is much larger than for HadISDH (or GHCNM).
Trends are very different.

Figure 12 - as for Figure 1.

As for Figures 10 and 11, the gridbox in Figure 12 is made up of a single station for each dataset which appears to be the same station. The month-to-month variability is very closely matched. However, the large negative adjustment to HadISDH from 1973 to 1994 results in a zero trend unlike the significantly negative trends for CRUTEM and GHCNM. By eye this adjustment could well be correct - the earlier period does appear to have a higher mean than the period from 1996 onwards.

Figure 13 - as for Figure 1.

As for Figures 10, 11 and 12, the gridbox in Figure 13 is made up of a single station for each dataset which appears to be the same station in terms of month-to-month variability which is very closely matched. However, there are deviations compared to CRUTEM prior to 1977 where no matching adjustment has been applied to HadISDH that suggest either different versions of the station are being used or that CRUTEM has had a different adjustment applied. GHCNM and HadISDH match up until around 2008 which is the time allocated as a changepoint in HadISDH, all data prior to this point have been adjusted downwards. This suggests that either GHCNM has had no adjustments applied in this case, or that they are smaller than for HadISDH. Trends are very different.

Figure 14 - as for Figure 1.

Figure 14 shows a gridbox with both large differences in trends and moderately matched variability. The large positive adjustment applied to HadISDH between 1974 and 1977 looks a little odd as the proceeding data are much lower. In this case, HadISDH only has one station and CRUTEM and GHCNM only have two - so all are quite sensitive to temporal intermittency and quality of the contributing stations. GHCNM is in agreement with HadISDH at the beginning of the record in terms of above average anomalies to some extent but then these two datasets disagree over he 1977 to 1982 period.

Figure 15 - as for Figure 1.

Despite having 14 stations listed for the gridbox in Figure 15, GHCNM has no time series for this gridbox. The CRUTEM time series appears to be in three distinct chunks, each with a different mean. Moderately large adjustments have been applied to HadISDH. Although HadISDH also has a step up between the middle section and the most recent chunk it is difficult to feel confident about the information shown in the gridbox from either dataset.

Figure 16 - as for Figure 1.

 The gridbox shown in Figure 16 has 5 stations listed for GHCNM, and a trend and correlation is given. However, the time series is very temporally intermittent such that it does not show up when plotted with a line because there are no consecutive months. It is quite possible that inhomogeneities remain within both the HadISDH and CRUTEM series here - certainly the large negative anomalies in CRUTEM during 2013 are suspicious.

Figure 17 - as for Figure 1.

 Figure 17 shows a gridbox where HadISDH and CRUTEM trends are more similar but CRUTEM correlates far better with GHCNM. HadISDH has undergone a series of upward adjustments throughout the record which will have reduced the long-term trend. There are two periods of particular disagreement with CRUTEM: ~1982/83 and 1989 to 1995.

Figure 18 - as for Figure 1.

Figure 18 shows a gridbox where GHCNM is clearly strange. The variance is intermittently very large - this is unlikely to be correct.

Figure 19 - as Figure 1.

Figure 19 shows what looks like very good adjustments on the part of HadISDH. There is generally good match up between the datasets in terms of correlations (this station was chosen as a poor correlation - it was the 8th lowest between HadISDH and GHCNM - so in general correlations are good). The match up remains good across the two periods of large adjustments: 1973 to 1997 and 2007 to 2011. This suggests that GHCNM and CRUTEM have undergone similar adjustments or that 10 station sin CRUTEM or 2 stations in GHCNM did not contain such inhomogeneities in the first place.

Figure 20 - as for Figure 1.

Figure 20 shows a gridbox that although classed as a low correlation gridbox isn't that bad. It shows upwards adjustments throughout the whole of the HadISDH series that appear reasonable and for the most part are inline with CRUTEM apart from the 2011 to present day period. This suggests that CRUTEM didn't have an adjustment in 2011 where as HadISDH and GHCNM did. CRUTEM has three stations contributing where as GHCNM and HadISDH only have one. So, the anomalously high most recent period could be due in part to drop-in/drop-out of one of those stations.

Figure 21 - as for Figure 1.

The gridbox in Figure 21 looks like the inhomogeneity ending in 2006 has been successfully adjusted in HadISDH and GHCNM but still remains in CRUTEM. Of course the homogeneity should not be identical in all datasets - while HadISDH has only one station contributing CRUTEM and GHCNM have seven.

Conclusions

• The vast majority of gridboxes agree very well.There are fewer than 10 gridboxes with correlations lower than 0.6 for each dataset pair.

• For a handful of gridboxes large uncertainties remain at the gridbox level between products.

• The main reasons for these uncertainties are: 
• different stations/versions of stations/quantity of stations making up the gridbox average 
• temporal intermittency in stations and sensitivity to this in gridboxes with very few contributing stations 
• homogenisation methods 

• Its is likely that the homogenisation does not always do a good job but it does appear improve the majority of time series. 

• Inhomogeneities are apparent in CRUTEM and GHCNM