Quality control of ε estimates (QA2): Difference between revisions
From Atomix
Yuengdjern (talk | contribs) No edit summary |
Yuengdjern (talk | contribs) No edit summary |
||
| Line 4: | Line 4: | ||
# Use the coefficient <math>a_0</math> (the intercept of the regression) to estimate the noise of the velocity observations and compare to the expected value based on the instrument settings. If noise is too high, <math> \epsilon </math> are rejected. | # Use the coefficient <math>a_0</math> (the intercept of the regression) to estimate the noise of the velocity observations and compare to the expected value based on the instrument settings. If noise is too high, <math> \epsilon </math> are rejected. | ||
# Data segments for which the regression coefficient a<sub>0</sub> (see [[Processing your ADCP data using structure function techniques | previous step]]) is negative (implying a negative noise floor) are likely to be invalid and are typically rejected | # Data segments for which the regression coefficient a<sub>0</sub> (see [[Processing your ADCP data using structure function techniques | previous step]]) is negative (implying a negative noise floor) are likely to be invalid and are typically rejected | ||
# In the case of <math> \epsilon </math> estimated using the modified regression method that accounts for oscillatory motion, reject data for invalid values of <math> | # In the case of <math> \epsilon </math> estimated using the modified regression method that accounts for oscillatory motion, reject data for invalid values of <math> a_3 </math>. | ||
# A better indication of the quality of the fit is usually provided by looking at the ratio of the estimated <math>\varepsilon</math> value to that based on the 95%-ile confidence interval estimate of the a<sub>1</sub> regression coefficient e.g. reject values where the ratio exceeds a specified threshold | # A better indication of the quality of the fit is usually provided by looking at the ratio of the estimated <math>\varepsilon</math> value to that based on the 95%-ile confidence interval estimate of the a<sub>1</sub> regression coefficient e.g. reject values where the ratio exceeds a specified threshold | ||
# The goodness of fit (R<sup>2</sup>) for the regression provides a basic indication of the quality of the fit, data with low R<sup>2</sup> are typically rejected. | # The goodness of fit (R<sup>2</sup>) for the regression provides a basic indication of the quality of the fit, data with low R<sup>2</sup> are typically rejected. | ||
Revision as of 20:51, 3 June 2022
Quality control measures for each beam:
- Data segments for which the regression coefficient a1 (see previous step) is negative yield an imaginary [math]\displaystyle{ \varepsilon }[/math] value, which should be rejected
- Ensure sufficient [math]\displaystyle{ D_{ll} }[/math] samples were used in the regression.
- Use the coefficient [math]\displaystyle{ a_0 }[/math] (the intercept of the regression) to estimate the noise of the velocity observations and compare to the expected value based on the instrument settings. If noise is too high, [math]\displaystyle{ \epsilon }[/math] are rejected.
- Data segments for which the regression coefficient a0 (see previous step) is negative (implying a negative noise floor) are likely to be invalid and are typically rejected
- In the case of [math]\displaystyle{ \epsilon }[/math] estimated using the modified regression method that accounts for oscillatory motion, reject data for invalid values of [math]\displaystyle{ a_3 }[/math].
- A better indication of the quality of the fit is usually provided by looking at the ratio of the estimated [math]\displaystyle{ \varepsilon }[/math] value to that based on the 95%-ile confidence interval estimate of the a1 regression coefficient e.g. reject values where the ratio exceeds a specified threshold
- The goodness of fit (R2) for the regression provides a basic indication of the quality of the fit, data with low R2 are typically rejected.
Other measures (not flagged):
- Examine the distribution of [math]\displaystyle{ \varepsilon }[/math] estimates - in most situations, this would be expected to be log-normal
- Comparison of observed values with nominal values based on established boundary-forced scalings may also be informative and help to identify observation or processing issues
Quality control measures for final [math]\displaystyle{ \epsilon }[/math] estimate:
- Examine the consistency of [math]\displaystyle{ \varepsilon }[/math] between bins (if evaluated) and between beams as an indication of estimate reliability - the geometric mean between beams is frequently used as the representative value
[In progress]
How ADCP structure function quality-control flags are applied
The Q (quality control) flags associated with shear-probe measurements are not compatible with the Ocean Sites Ocean Sites for quality control (QC) coding.
Every dissipation estimate from every probe must have Q flag. The numerical values of the Q flags are as follows:
| Flag Mask | Bit | Flag Meaning | Example threshold value | Ex: True =1 / False =0 | Ex: Q value |
|---|---|---|---|---|---|
| 1 | Bit 0 | if FOM > FOM_limit | 2 | 0 | 0 |
| 2 | Bit 1 | if despike_fraction > despike_fraction_limit | 40% | 0 | 0 |
| 4 | Bit 2 | log(e_max)-log(e_min)|> diss_ratio_limit X \sigma_{\ln\varepsilon} | N/A | 1 | 4 |
| 8 | Bit 3 | if despike_iterations > despike_iterations_limit | To be confirmed | 0 | 0 |
| 16 | Bit 4 | if variance resolved less than a threshold | 50% | 1 | 16 |
| 32 | Bit 5 | manual flag to be defined by user | N/A | 0 | 0 |
| 64 | Bit 6 | manual flag to be defined by user | N/A | 0 | 0 |
| 128 | Bit 7 | manual flag to be defined by user | N/A | 0 | 0 |
| Final Q = 20 |
The Q flags are combined by their addition. For example a Q value of 20 means that the dissipation estimated failed both dissipation ratio limit test and the resolved variance test. A value of 255 means that all tests failed. The reasons for a failure can be decoded by breaking the value of Q down to its powers of 2.
Return to ADCP Flow Chart front page
