BIO student cruise (BIO325) 2020 September
Data for calibration
To calibrate the oxygen data we collected water samples for winkler
titration at different depths and compared these with upcast oxygen
values from the CTD (recorded as the niskin bottle was fired).
Outlier removal procedure
Datapoints flagged as bad were removed prior to data analysis (bad
data included sampling errors, values below the threshold of 10 umol/kg,
negative winkler results and duplicate winkler samples with a difference
larger than 3 umol/kg.).
We also excluded datapoints shallower than 45 m.
After removal of flagged data and data above 45 m, we removed
additional outliers using an iterative outlier removal pocedure with
2.5 x RMSE as the outlier-threshold. If the point’s
residual value (value of the difference between the regression line and
the winkler value) was outside this limit, it was considered an outlier.
After removal of outliers, a new regression line and 2.5 x RMSE
threshold was calculated for the new dataset and the whole procedure was
repeated until no more datapoints were considered outliers or
until the regression had RMSE < 2.
Calibration method
We plotted CTD niskin bottle oxygen against winkler samples and
calculated the linear regression line. Calibrated values were obtained
by inserting raw CTD umol/kg as x in the linear regression equation
(calibration equation).
Oxygen sample overview
Oxygen samples were collected by Martine R. Solås. Winkler
titrations were done by Kristin Jackson-Misje.
We collected at total of 7 winkler samples. Of
these, 1 was/were flagged as bad during the initial
collection or because uncalibrated values were below the threshold of 10
umol/kg.
0 samples were part of a duplicate or triplicate
sampling (i.e. a total of 0 niskin bottle(s) had been sampled
more than once). 0 these duplicates/triplicates (i.e. 0 individual
winkler samples) were furthermore flagged as bad because the
between-samples difference was more than 3 umol/kg. The remaining
duplicates/triplicates were averaged before analysis so that we only had
one winkler value per niskin bottle per station.
After averaging of duplicates/triplicates and removal of flagged
data, we were left with a total of 6 unique winkler
samples.
1 of these samples were collected shallower than 45
m and consequently removed before further analysis.
The plots below give an overview of the offset between CTD oxygen
and winkler and whether there are clear systematic errors when the
offset is plotted against dissolved oxygen concentration, niksin bottle,
station number or depth/pressure.
Before we created our calibration curve, we removed 0 datapoint(s)
(0 % of remaining points) through 1 round(s) of iterative
outlier removal using 2.5 x RMSE as the outlier threshold.
This lead to a total of 5 datapoints for the final linear
regression (calibration).
Calibration
To calibrate the raw CTD data we inserted raw umol/kg as x in the
linear regression equation:
y = -1.41 + 1.0513 x
We can check for strong systematic errors in the calibration by
plotting the calibration-winkler-offset across dissolved oxygen
concentrations, depth, stations and niskin bottle.
Plots with a blue frame are plots where the y axis limits have been
set to -10 to 10 for better inspection (and comparison with other
cruises) of the trend visualized with the linear regression line in
blue. Only data labelled “bad” have been excluded from this trend line
(meaning that data above 45 m and those points that were removed in the
iterative outlier removal process are included).
## Warning: Not enough data to perform fit for group 1; computing mean instead.
Profiles of raw and corrected data
Finally we plot the CTD stations with collected winkler data to
compare the corrected data with raw data and winkler results. Points
represent the niskin bottle values (uncalibrated and calibrated) and
corresponding winkler data while the graphs represent the downcast CTD
(uncalibrated and calibrated). Downcast values shown here have been
smoothed through a rolling mean filter with window size 3.
Conversion of raw ml/l to corrected ml/l
We also created a calibration equation for the ml/l unit directly by
converting corrected umol/kg to ml/l and creating a regression between
corrected ml/l and raw ml/l: