1. Produce a gridded version using only stations that report throughout the entire record because this should prevent any inhomogeneity related to station density.
How many stations/gridboxes can be used for this?
Compare trends and climate features between HadISDH.full and HadISDH.long?
Are there any regions that are no-long sampled?
2. Run PHA on the gridded values to see if these are homogeneous. If they are not this is either due to: changes in station density, remaining inhomogeneities or climate variability.
What is the frequency of detected changepoints?
Is it the same for HadISDH.full verses HadISDH.long?
Can these changepoints be attributed to known climate events (e.g., ENSO) or changes in station density?
3. Apply locations of changepoints found in DPD to T indirectly because DPD has a better signal-to-noise ratio than T. Changes found in DPD thought to originate from wet bulb temperature/RH sensor/dewcell in many cases would be changes that affect T also (station move/environment change/observing practice change/shelter change/simultaneous instrument change/automation). Changes found in DPD that also occur in T simultaneously (or within 12 months) do not need to be applied so need to check.
What is the changepoint frequency?
Do the homogenised series look sensible?
How do the trends/climate features differ between HadISDH.landT.PHAdirect verses HadISDH.landT.PHAindirect?
4. Investigate the higher frequency of changespoints at the beginning/end of the series. This is in part due to PHA often finding changpoints here but being unable to robustly assign an adjustment and so dissassigning the changepoint but PHAindirect still assigns the changepoints and adjusts.
In all cases PHAindirect applied adjustments are either due to PHAdirect logged adjustments (real changepoint) or data removals but to failed adjustment estimation (so a real changepoint). This means that there must be a higher frequency of detected changepoints at series endpoints, however, many of these cannot be robustly allocated an adjustment value and so they are not counted as changepoints in PHAdirect. It may be that more are found because of the restriction on detecting changepoints in the first and last two years. Any changepoints within these periods will then be detected in the later/earlier data.
5. Look at the seasonal variation in the gridbox trends - especially for RH?
Is it driven by any one season in particular?
Do seasons have opposing trends?
6. Look at SST trends around the North Atlantic and Indian Ocean - does this help to explain some of the features in the gridbox trends?
7. Look at land wind speeds (declining) for help explaining the decline in RH although wind speed decline is linear where as RH decline begins post-2000.
8. Look at land skin Temperature post 1995 to see if there is any correlation. This is closely related to soil moisture/vegetation patterns?
9. Compare the homogenised grids with CRUTS3.1, 20CR etc.
10. Uncertainty over regions where data are missing could be much larger than estimated by comparison with ERA-Interim coverage. This is very difficult to quantify. However, we can say things about quasi-global/over the observed regions?
11. Estimate measurement uncertainty for T and the humidity variable and combine.
How much are these two things correlated?
May need to add a covariance component into the quadrature combination.
12. Estimate uncertainty for the wet bulb temperature and also the RH (RH sensor). Use wet bulb temperature uncertainty up until 1980s and then swap to RH sensor uncertainty progressively for more and more stations (at random) or see which is the largest source of uncertainty.
Wet bulb uncertainty is likely to be largest at low humidity where as RH sensor uncertainty is likely to be largest at high humidities.
Histeresis affects RH sensors more than wet bulb sensors - slow to wet up and slower to dry down - resulting in a wet bias over all.
How many stations/gridboxes can be used for this?
Compare trends and climate features between HadISDH.full and HadISDH.long?
Are there any regions that are no-long sampled?
2. Run PHA on the gridded values to see if these are homogeneous. If they are not this is either due to: changes in station density, remaining inhomogeneities or climate variability.
What is the frequency of detected changepoints?
Is it the same for HadISDH.full verses HadISDH.long?
Can these changepoints be attributed to known climate events (e.g., ENSO) or changes in station density?
3. Apply locations of changepoints found in DPD to T indirectly because DPD has a better signal-to-noise ratio than T. Changes found in DPD thought to originate from wet bulb temperature/RH sensor/dewcell in many cases would be changes that affect T also (station move/environment change/observing practice change/shelter change/simultaneous instrument change/automation). Changes found in DPD that also occur in T simultaneously (or within 12 months) do not need to be applied so need to check.
What is the changepoint frequency?
Do the homogenised series look sensible?
How do the trends/climate features differ between HadISDH.landT.PHAdirect verses HadISDH.landT.PHAindirect?
4. Investigate the higher frequency of changespoints at the beginning/end of the series. This is in part due to PHA often finding changpoints here but being unable to robustly assign an adjustment and so dissassigning the changepoint but PHAindirect still assigns the changepoints and adjusts.
In all cases PHAindirect applied adjustments are either due to PHAdirect logged adjustments (real changepoint) or data removals but to failed adjustment estimation (so a real changepoint). This means that there must be a higher frequency of detected changepoints at series endpoints, however, many of these cannot be robustly allocated an adjustment value and so they are not counted as changepoints in PHAdirect. It may be that more are found because of the restriction on detecting changepoints in the first and last two years. Any changepoints within these periods will then be detected in the later/earlier data.
5. Look at the seasonal variation in the gridbox trends - especially for RH?
Is it driven by any one season in particular?
Do seasons have opposing trends?
6. Look at SST trends around the North Atlantic and Indian Ocean - does this help to explain some of the features in the gridbox trends?
7. Look at land wind speeds (declining) for help explaining the decline in RH although wind speed decline is linear where as RH decline begins post-2000.
8. Look at land skin Temperature post 1995 to see if there is any correlation. This is closely related to soil moisture/vegetation patterns?
9. Compare the homogenised grids with CRUTS3.1, 20CR etc.
10. Uncertainty over regions where data are missing could be much larger than estimated by comparison with ERA-Interim coverage. This is very difficult to quantify. However, we can say things about quasi-global/over the observed regions?
11. Estimate measurement uncertainty for T and the humidity variable and combine.
How much are these two things correlated?
May need to add a covariance component into the quadrature combination.
12. Estimate uncertainty for the wet bulb temperature and also the RH (RH sensor). Use wet bulb temperature uncertainty up until 1980s and then swap to RH sensor uncertainty progressively for more and more stations (at random) or see which is the largest source of uncertainty.
Wet bulb uncertainty is likely to be largest at low humidity where as RH sensor uncertainty is likely to be largest at high humidities.
Histeresis affects RH sensors more than wet bulb sensors - slow to wet up and slower to dry down - resulting in a wet bias over all.
No comments:
Post a Comment
Note: only a member of this blog may post a comment.