Information is now being written to another page.
Let say:
For each field we want to know
FIELD_QC | |||
Not done | Bad | Good | |
|FIELD_DIFF|>tolerance | DN (different-not done) | DB (different-bad) | DG (different-good) |
|FIELD_DIFF|≤tolerance | SN (same-not done) | SB (same-bad) | SG (same-good) |
The QC hierachy is
The profile data (looking at a file, 4901152.2010.nc, when Data Mode is "A" throughout - A is `real time with adjusted values').
I want to be sure that the DACs aren't simply throwing out data that is clearly wrong and not storing it in their files. If so, we're in danger of not giving them credit for throwing out bad data.
Year | Start in Julian days |
---|---|
2004 | 19,724 |
2005 | 20,089 |
2006 | 20,454 |
2007 | 20,819 |
2008 | 21,184 |
2009 | 21,550 |
2010 | 21,915 |
I'm only looking at data where all 5 Data Centres are defined and we have delayed mode QC that is 100% accepted or 100% rejected.
Excluding -999, the profiles take the following values
Data centres are
Before 2009, coriolis_tmp_qc is always 9.
Using all the profiles defined by both the delayed mode and the Data Centres. Only considering delayed mode profile where all the levels are either accepted or rejected.
Bad-reject | Bad-accept | Good-reject | Good-accept | Total | ETS | Bias | FoM | |
BMRC | 2103 (1.106%) | 2442 (1.284%) | 833 (0.438%) | 184823 (97.172%) | 190201 | 0.3830 | 1.55 | 0.649 |
FNMOC | 2243 (0.861%) | 31875 (12.229%) | 4732 (1.815%) | 221798 (85.095%) | 260648 | 0.0351 | 4.89 | |
MEDS | 1195 (0.551%) | 12792 (5.900%) | 4092 (1.887%) | 198743 (91.662%) | 216822 | 0.0481 | 2.65 | |
UKMO | 115 (0.128%) | 6288 (7.025%) | 256 (0.286%) | 82844 (92.560%) | 89503 | 0.0133 | 17.26 |
Using all the levels defined by both the delayed mode and the Data Centres
Bad-reject | Bad-accept | Good-reject | Good-accept | Total | ETS | Bias | FoM | |
BMRC | 32,767 (0.237%) | 156,742 (1.136%) | 49,444 (0.358%) | 13,560,006 (98.268%) | 13,798,959 | 0.1330 | 2.31 | 0.546 |
Coriolis | 285,257 (1.674%) | 1,659,197 (9.735%) | 88,162 (0.517%) | 15,011,658 (88.074%) | 17,044,274 | 0.1219 | 5.21 | |
FNMOC | 55,623 (0.309%) | 1,981,459 (10.997%) | 183,518 (1.018%) | 15,797,903 (87.676%) | 18,018,503 | 0.0130 | 8.52 | |
MEDS | 10,325 (0.069%) | 877,361 (5.883%) | 7,614 (0.051%) | 14,019,228 (93.997%) | 14,914,528 | 0.0104 | 49.48 | |
UKMO | 24,648 (0.330%) | 407,955 (5.466%) | 1,883,760 (25.240%) | 5,146,907 (68.963%) | 7,463,270 | -0.0390 | 0.23 |
For all levels
Bad-reject | Bad-accept | Good-reject | Good-accept | Total | ETS | Bias | |
bmrc | 4,356 (0.037%) | 89,960 (0.771%) | 23,796 (0.204%) | 11,548,575 (98.988%) | 11,666,687 | 0.0350 | 3.35 |
coriolis | 11029 (0.095%) | 83411 (0.715%) | 44980 (0.386%) | 11527018 (98.805%) | 11666438 | 0.0761 | 1.69 |
fnmoc | 8540 (0.073%) | 86086 (0.738%) | 150541 (1.290%) | 11423068 (97.899%) | 11668235 | 0.0297 | 0.59 |
meds | 3459 (0.030%) | 91167 (0.781%) | 5135 (0.044%) | 11568471 (99.145%) | 11668232 | 0.0340 | 11.01 |
Profiles
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
7 (accept) | 0.77 | 99.23 |
8 (accept) | 7.7 | 92.3 |
11 (accept) | 22 | 78 |
10 (accept, only 71 values) | 80 | 20 |
9 (reject, only 197 values) | 65 | 35 |
12 (reject, only 45 values) | 75 | 25 |
However, putting a QC of 10 into reject does improve the skill
QC=10 is accept | QC=10 is reject | |
---|---|---|
ETS | 0.383 | 0.430 |
Bias | 1.55 | 1.34 |
FoM | 0.649 | 0.680 |
Ranking for levels
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
2 (accept) | 0.89 | 99.11 |
5 (accept, 53,096 values) | 65.9 | 34.1 |
4 (reject, 82,211 values) | 39.9 | 60.1 |
QC=5 is accept | QC=5 is reject | |
---|---|---|
ETS | 0.133 | 0.258 |
Bias | 2.31 | 1.40 |
FoM | 0.546 | 0.608 |
For levels
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
1 (accept) | 10 | 90 |
3 (reject) | 75 | 25 |
4 (reject) | 79 | 21 |
For profiles
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
94 (accept) | 17.5 | 82.5 |
95 (accept) | 19.8 | 80.2 |
96 (reject) | 20.4 | 79.6 |
97 (reject) | 21.5 | 78.5 |
98 (reject) | 27.3 | 72.7 |
99 (reject) | 29.1 | 70.9 |
100 (reject) | 43.8 | 56.3 |
Trying to improve the skill scores
QC > 95 is rejected | QC > 94 is rejected | QC > 93 is rejected | QC > 91 is rejected | QC > 90 is rejected | QC > 89 is rejected | |
---|---|---|---|---|---|---|
ETS | 0.0351 | 0.0362 | 0.0366 | 0.0370 | 0.0372 | 0.0368 |
Bias | 4.89 | 4.29 | 3.82 | 3.13 | 2.85 | 2.61 |
FoM | 0.512 | 0.512 | 0.513 | 0.514 | 0.514 | 0.514 |
For levels
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
98 (accept) | 10.4 | 89.6 |
99 (accept) | 11.5 | 88.5 |
100 (reject) | 23.3 | 76.7 |
Despite the bias being very high (8.52, so accepting too much data), we can't improve this (ETS and FoM are decrease if we reject QC of 99).
Profiles
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
5 (accept) | 5.41 | 94.6 |
1 (accept) | 6.05 | 93.95 |
4 (reject) | 24.3 | 75.7 |
3 (reject) | 15.0 | 85.0 |
Levels
QC flag (Jim's class) | % of bad | % of good |
---|---|---|
1 (accept) | 5.89 | 94.1 |
3 (reject) | 45.7 | 54.3 |
4 (reject) | 59.2 | 40.8 |
For the four data centres where we have a lot of data, the bias values are all greater than 1 - indicating that they're accepting more data than there is good data. This suggests that we might be able to improve the skill scores for the observation centres if they rejected more data. The obvious data to reject would be those data where the QC flags suggests the data is only sufficiently OK to keep, and is considered doubtful.
This has been possbile for BMRC where the skill scores have been significantly increased by rejecting profiles with a QC of 10 and rejecting levels with a QC of 5. A slight improvement to the skill scores for the profiles for FNMOC can be made by increasing the amount of data that is rejected. But changing the accept/reject criteria can't improve the skill scores for Coriolis and MEDS.