Building a 'Quicklook' Homogenised RH Product - version 1

What and why?                                                                                      

RH (a measure of the level of saturation of the atmosphere, not the amount of water vapour) over land appears to be declining over recent years according to ERA-Interim reanalysis. The most recent observational product ends in 2007 and is in good agreement with ERA-Interim until that time. The HadISDH land specific humidity product is now available and shows that land specific humidity has remained relatively flat over recent years. Previous thinking suggested that RH should remain relatively constant over time over large scales.                        

                                                                                                                                  

The question remains: is the surface atmosphere becoming less saturated? We cannot address this over ocean in the near future but a 'quicklook' at RH over land is possible.

A number of steps need to be taken to produce 'HadISDH.landRH.quicklook' from the HadISD base that was used for HadISDH.landq:                                      

• calculate monthly mean RH
• run the monthly means through the PHA homogenisation code to remove gross inhomogeneity
• average the monthly means and monthly mean anomalies over 5by5 degree gridboxes
• calculate cosine weighted area averages for the globe, hemispheres and tropics.

Main conclusions:                                                                                 

The homogenisation of the land RH makes some considerable changes to the data at the local level (Figure 1a) - far more so than for specific humidity. However, ~68% of gridboxes have the same direction of trend both before and after homogenisation (Figure 1b). 

For the global average (Figure 1d) both the raw and homogenised data show a decrease over time, largely driven by post-2000 years. For the Northern Hemisphere (Figure 1e) the homogenised data show drying out over land (becoming less saturated as opposed to having less water vapour) whereas for the raw data this trend is not significant. For both the tropics and the Southern Hemisphere (Figures 1f,g) the homogenised data do not show a clear trend in either direction whereas the raw data show a decrease.

Gridbox decadal trends are shown in figure 2. This indicates widespread drying but strong regional signals. The moistening over India, the Carribean, northeastern North America and northeastern Asia is notable.

This is still a draft product and further work needs to be done to assess the uncertainty in these values. However, comparing the raw and homogenised data goes some way to assessing the uncertainty in the data. 

I think it is reasonable to say that the drying over land is likely to be a robust signal. RH has steadily declined since 2000 except for over the tropics. 2012 looks like the driest year since records began in 1973. However, this drying is driven largely by the mid-latitudes and data-coverage is poor in the Southern Hemisphere and Tropics so large uncertainty remains here.

Figure 1: Difference between trends (1973–2012) in HadISDH (land RH) before and after the pairwise homogenisation process. (a) Ratio of decadal trends from the raw HadISDH (land RH) compared to homogenised HadISDH. Note non-linear colour bars. (b) Scatter relationship between homogenised and raw decadal trends for HadISDH (land RH). The percentage of grid boxes present in each quadrant is shown. (c) Distribution of grid-box trends for the homogenised and raw data. (d–g) Large-scale area average annual anomaly time series
and trends for homogenised HadISDH and the raw data relative to the 1976–2005 climatology period.

Figure 2: Decadal trends in relative humidity for HadISDH over 1973–2012. Trends are fitted using the median of pairwise slopes and high confidence assigned with a black dot where the trend is very likely to be significantly different from zero. Note non-linear colour bars.

Remaining issues:                                                                                 

Are the stations removed the same as for land q and if not do we know why?

Why are there no issues of sub-zero RH?

Is it clear which may be the best humidity variable to homogenise if we cannot homogenise them all independently (which may lead to physical inconsistencies between them)?

How to we build an uncertainty model for RH?

Working:                                                                                               
I started with the same subset of 3694 HadISD stations as for HadISDH.landq. These contain QC'd sub-daily temperature and dewpoint temperature. I converted these to RH at the sub-daily resolution and then averaged over each month where there were sufficient data to do so.

These monthly RH values were then run through the PHA homogenisation algorithm which takes networks of highly correlating stations and compares each in turn with all of its neighbours to detect changepoints and make appropriate adjustments. In most cases this appears to do a reasonable job (Figure 3) but there are occasions where issues remain (Figure 4). This is the first time PHA has been used on RH data and a more detailed validation needs to be undertaken. 

Here are the top 100 stations (ISD IDs) with the largest adjustments (in %) applied:
04231099999  26.3100
36982099999 -25.7900
44284099999 -22.9400
44277099999 -22.5800
72654899999  22.2500
44287099999 -20.6400
44213099999 -18.8500
36982099999  17.9600
38545099999 -17.9300
04260099999 -17.2500
60714099999 -17.0900
44218099999  16.9900
59948099999  16.8500
76577099999 -16.6000
56106099999 -16.5100
55664099999 -16.3800
85934099999 -16.1600
70321099999 -16.0400
62103099999  15.8600
58367099999 -15.6700
72467699999  15.6100
94346099999 -15.5300
44373099999  15.5000
04202099999 -15.3900
72446713930  15.3600
41240099999 -15.3100
72677624036 -15.3100
84735099999 -15.2000
72412799999  15.1200
44352099999  14.9500
38001099999  14.8200
72408614793 -14.6600
17330099999 -14.6600
08044099999  14.5600
78367011706  14.3800
22602099999  14.3000
72493023230 -14.2700
08044099999 -14.1300
72582524121 -14.0900
01160099999 -14.0700
47136099999  14.0100
70265099999  14.0000
71943099999 -13.9600
44213099999  13.8900
11032099999  13.8900
40272099999 -13.8400
17022099999  13.7800
72383023187  13.7700
16453099999  13.7300
82281099999 -13.6600
40310099999 -13.6600
55299099999 -13.5300
16667099999  13.5100
60035099999 -13.4100
74694113786  13.3700
72821099999  13.3300
40191099999  13.3300
44232099999  13.3200
72582524121  13.2700
16021099999  13.2400
44241099999  13.2000
44354099999 -13.1500
72257003933 -13.1100
42647099999  13.0800
16110099999 -12.9800
72486023154  12.9700
87509099999 -12.9100
55472099999 -12.8700
48456099999  12.7700
38262099999 -12.7300
44336099999  12.7300
40270099999 -12.7300
78016013601 -12.7000
56106099999  12.6400
16110099999  12.5600
01403099999  12.5500
72594624286  12.5300
01078099999  12.5000
35358099999 -12.4900
72408614793  12.4800
70454025704 -12.4000
44341099999  12.4000
47070099999  12.3800
01465099999  12.3800
44284099999  12.3600
08429099999 -12.3400
38353099999  12.3100
08429099999  12.3000
01055099999 -12.3000
85201099999  12.2800
83768099999  12.2800
72290693112 -12.2800
85201099999 -12.2600
83612099999 -12.2600
89611099999  12.2400
15410099999 -12.2300
78388099999 -12.2000
16153099999  12.2000
01205099999 -12.2000
55279099999 -12.1600

Here are some examples of how the PHA has detected changepoints against the neighbour network and applied adjustments. Some segments are missing in the homogenised station data because adjustments could not be resolved.

Figure 3: An example of PHA making what appear to be sensible adjustments to homogenise the data. Homogenised (blue) verses raw (red) station monthly mean RH for 607140 (Bizerte, Tunisia, 37.25N, 9.80E, 3.0m) against its network of raw neighbours (black). Decadal trends in RH (%) are shown with 5th-95th percentiles.

Figure 4: An example of PHA making what appear to be illogical adjustments to homogenise the data. Homogenised (blue) verses raw (red) station monthly mean RH for 042600 (Paamiut, Greenland, 62.00N, -49.667W, 15.0m) against its network of raw neighbours (black). Decadal trends in RH (%) are shown with 5th-95th percentiles.

PHA code cannot be resolved for 11 stations:

68104099999 -22.88     14.43            0   NM WALVIS BAY (PELICAN          

68994099999 -46.88     37.87           21  ZA MARION ISLAND               
84782099999 -18.05    -70.27          458 PR TACNA                       
85469099999 -27.17   -109.43           69 CH ISLA DE PASCUA              
85470099999 -27.30    -70.42          290 CH COPIAPO                      

85488099999 -29.92    -71.20          146 CH LA SERENA                   
85930099999 -52.40    -75.10           52  CH FARO EVANGELISTAS             

89022099999 -75.50    -26.65           30  AY HALLEY                        

91610099999   1.35    172.92            4   KB TARAWA                      
91643080705  -8.53    179.22            2   TV FUNAFUTI NF          

91943099999 -14.48   -145.03            3  PF TAKAROA  

These are different from the three stations that could not be resolved for landq.

245 stations are now too short to calculate climatological means:
010470,011060,011150,011940,012120,
012280,020550,026220,026300,031540,032080,032620,033140,033470,034925,038740,038800,
039670,060960,062470,064960,074280,080480,082320,082720,083830,105140,111380,111820,
111900,112140,121250,122150,122300,122850,124550,125500,126700,131730,133480,133690,
134650,140240,161140,165220,166840,170700,170840,222820,222920,229390,270080,277190,
280440,284650,292530,294180,294560,294770,295760,295870,303560,303790,305260,306370,
306640,307120,307290,309150,312630,353020,376820,382220,400010,400090,400160,400390,
400610,402650,403400,407060,407450,408090,415300,415940,419780,420270,425590,433330,
442150,442390,442560,443140,443470,471620,471920,484770,485650,505480,547510,597920,
600600,600960,605060,606020,606110,607280,610170,612300,612700,612960,614010,614920,
614970,620560,621610,636410,637910,648600,677810,678430,678530,678610,679690,683280,
684610,702315,702320,702350,702610,702757,711430,712220,716100,718260,718420,718450,
722029,722038,722119,722345,722348,722429,722533,722598,722599,722673,722867,722897,
722903,722925,722955,723069,723119,723307,723345,723536,723555,723815,723825,723930,
724035,724067,724094,724097,724288,724466,724506,724721,724800,724936,725128,725245,
725314,725396,725524,725715,725955,726375,726515,726796,727344,727436,727458,727834,
727928,727934,744865,744900,745090,762555,762863,763503,763820,764050,764230,764915, 
765560,766540,767755,780620,784600,784820,785350,785830,787110,787170,787410,789580, 
802100,804070,804100,804190,804270,812000,844250,860110,860680,862330,862600,862970,
864300,865300,870160,870780,873200,875930,876400,876480,878520,889630,890560,890590, 
890660,895710,912210,915820,964810,969350,972600,983290,984280,984460,985480,987410  

0 stations suffer from issues of subzero RH (unlike for q)

26 stations suffer from issues of >100% (supersaturated) RH
061200,064760,110220,110280,111300,151080,200870,202920,219820,230320,319600,470610,
547760,577760,584370,586660,700260,700860,703080,718936,724837,725975,745160,857990,
870970,984400         

3438 stations continue through to the gridding process         

Stations have been averaged across 5 degree by 5 degree gridboxes.

Linear trends have been estimated using the median of pairwise slopes (Sen, 1968; Lanzante, 1996) method. Where intervals defined by the 90 % confidence limits on the median of the slopes (5th and 95th percentiles of the pairwise slopes) are both of the same sign as the median trend presented in the grid boxes, the trend is presumed to be significantly different from a zero trend. This is indicated by a black dot within the grid box (Figure 2). This means that there is higher confidence in the direction of the trend, but not necessarily the magnitude. The spread of the confidence interval provides the confidence in the magnitude