1. Latest updates on alternative, fully automated approaches to Optical Sensors Radiometric Calibration and data quality
IDEAS+ WP3520 Calibration and Data Quality Toolbox
Steve Mackin, Jeff Settle and James Warner

2. Contents
This presentation will be a summary presentation indicating which areas work well and which still require further investigation. Given the time constraints it will not be in depth, but will summarise some of the findings.
Principles
Areas being investigated
Conclusions

3. principles

4. Basic principles
Currently radiometric calibration and data quality relies largely on infrequently acquired on-board data (Sentinel-2 has a calibration cycle every month) or through vicarious calibration using specific sites that may have environmental limitations (Dome-C during its winter, cloud cover over Libya 4). So opportunities are limited.
OUR APPROACH
All images collected contain useful information for assessing changes in radiometry and data quality, this provides very high temporal sampling of parameters of interest.

5. Basic principles
The automation means that we can get very accurate estimates of key parameters without requiring human intervention which is a key advantage.
In the monthly cyclic reports from ACRI, they state they have not validated the diffuser estimates of the SNR because…
“The SNR assessment from EO data has not been applied to OLCI-B considering …that it requires a significant amount of human and machine resources that can be more efficiently used for other tasks”

6. Early OLCI SNR slide This doesn’t have to be the case.
Data from two days using the coarse algorithm. Each dot is an image. Upper envelope defines SNR
Totally automated in processing and analysis. More in the SNR section.

7. Types of images we use, essentially everything

8. Advantages of using normal images
By using normal images, we have a much higher sampling interval, every 41 seconds for Sentinel-2 rather than every month using on-board devices.
We can therefore monitor and update our results with a much higher frequency than many on-board devices and vicarious methods which use specific sites, that can only be accessed infrequently.
By using normal images, in theory we can update coefficients automatically as soon as an issue is encountered (detector non-uniformity, where a single detector responsivity changes dramatically in a short period of time)
We also avoid “dead” periods where a specific site cannot be used, such as the polar sites in Antarctica and Greenland that for precision work can only be used effectively for one to two months per year.

9. Areas being investigated

10. Areas being investigated (all in-orbit) using normal images
Absolute Calibration drift
Signal to Noise Ratio assessment
Focus assessment
Relative gain determination
Non-linearity assessment
Platform stability (geometric)

11. Absolute calibration drift
The methodology was developed initially under the ESA ALTS contract for AATSR data.
The results showed that using Level 0 data and simple calculations (a subtraction and division) with the underlying assumption that for a mission the earth radiance was constant over time we could derive the on-board calibrator drift.
The surface composition, atmosphere, presence of cloud, BRDF of the surface in theory have no discernible impact on the final result.
This can be extended from monitoring the on-board calibrator to monitoring the absolute drift directly.
Only problem is that we need to process ALL images, but this can be automated.

12. Results – Calibrator drift
This is being considered for future use on other sensors including S2 MSI, PROBA-V, S3 OLCI and S3 SLSTR

13. Signal to noise ratio (SNR)
We have developed several algorithms for SNR estimation, from the original coarse algorithm to refined algorithms that can assess single detector variations and determine differences across and along track due to residual striping
Further work is on-going. Some sensors are problematical such as S2-MSI which shows high levels of signal quantisation, while S3-OLCI shows a very good response even given its large pixel size (300m GSD). Which means it may be applicable to sensors with large footprints such as Proba-V.
The coarse algorithm uses many images, typically a couple of hundred (about two days of S3-OLCI data for example)
The finer algorithms use a lower number of images.

14. Early OLCI SNR slide
Data from two days using the coarse algorithm. Each dot is an image. Upper envelope defines SNR
Diffuser values are over 10% higher. Part of the difference is due to extra noise from persistent residuals due to differences between neighbouring detectors

15. Focus assessment
The current algorithm gives a good indication of relative differences across the swath of an instrument.
The absolute magnitude and some of the relative differences are sensitive to surface variability. More homogeneous surfaces give better relative differences, but not necessarily good absolute values.
Algorithm is under assessment to try and make it more robust in the presence of surface variability or even changes in solar illumination.
However, using a larger number of images (less than an orbit of S2 MSI data) we can get good relative estimates across track and some sense of absolute stability.

16. Focus Sentinel-2 results
Using less than 30 images we were also able to assess the across-track focus variation across the twelve detector sets
Some asymmetry can be seen, higher on the left (S2A blue, S2B orange).
Ideal would be a whole days worth of data from each sensor.
Possibility of trending focus change with time.

17. Relative Gain
Algorithm works very effectively in extracting the high frequency relative gain variations even in the presence of image heterogeneity for even single images.
Low frequency terms are more difficult to extract and require the use of multiple bands and multiple images to get approximate corrections. Further modifications to the algorithms are envisaged to improve this capability.
The algorithm works very well, but we believe that the accuracy can be improved by binning the observed relative gain values in different radiance bins, due to non-linearity effects we will discuss in the next section.
This can be used to determine multiplicative and additive corrections.

18. Relative gain - Persistent residuals
A persistent residual is a calibration residual found in radiometrically corrected and equalised imagery. In theory, in a perfectly calibrated image there should be no residuals.
X-axis is a relative detector number, 39 detectors centred on detector 711 for the Blue band detector set 7 (there are 12 detector sets that make up the swath)
Y-axis is percentage variation observed, where = 0.2%.
Each month is the average of five random images.

19. Correlated persistent residuals
Band 10 - Argentina (x) and California (y)
If we get any two images and plot the calibration residuals determined by our relative gain algorithm in a scattergram we see very high correlations in the imagery (blue band left, cirrus band right). The blue band values exceeding 0.3%, the cirrus band values exceeding 2% a huge discrepancy.

20. Persistence of residuals temporally
OLCI data also has residuals
Very consistent results showing the same persistent residuals over six months. Note the magnitude of the residuals less than 0.05%

21. Homogeneous surfaces give better results
And…the daily averages match up with the repeat data over Libya 4 which has a lower number of images, but more homogeneous surface. Especially good match for June data.

22. Non-linearity
Once we found persistent residuals and given the magnitude was exceeding requirements in some cases (for both S2 MSI and S3 OLCI), we had to question why they couldn’t be seen by the MPC using the on-board diffuser.
The conclusion was that at diffuser brightness the differences were very small, but at normal background brightness we could see differences, as each detector response varied a little due to non-linearity effects.
The updated relative gain algorithm which bins the features according to radiance should produce an improved estimate of non-linearity with less scatter and using less data.

23. Persistent residuals – Variable depths
Finding persistent residuals means we have to ask the question…
Why can we see these features in normal images but the MPC can not see them in diffuser images?
A clue to why could be what we see in the plot (left). Note that the grey line extremes are greater than the orange line extremes with the dark blue line extremes in-between.
We considered that this sort of behaviour, along with the fact that the MPC saw nothing could be related to a non-linearity error.

24. Example of non-linearity in S2A?
Bias term (subtracted)
Gain term
Radiance
Digital Number (DN) A B
Diffuser images
Used 1/6th of an orbit of data.
If we ratio the values A/B we get a distinct pattern of behaviour which shows the correction required to remove the persistent residuals
Assuming our non-linearity correction is not perfect (B)

25. Very large effects for several months
The same feature persists throughout June, until the 21st June, then it disappears. Note the variation in the depth of the feature, are the changes related to brightness variations?

26. Modelled additive effect
We created a very simple model assuming an additive term of 2.5 scaled radiance units (green dots) which accounted for almost all the variation seen.
Therefore we can distinguish quite clearly between additive terms where the model converges towards one, and multiplicative effects shown previously.

27. S3 OLCI data shows similar problems
Unexpected that the calibration is off by 0.5% across a large radiance range. We saw many examples of this. Also examples of sloping trends intersecting the one line but diverging at lower radiance values. Way outside requirements.

28. S3 OLCI data shows similar problems
The effects in imagery of the feature shown in the previous slide

29. Conclusions
The methodologies developed provide an alternative way of assessing calibration and data quality, sometimes providing information that can not be easily determined using on-board devices or established vicarious methods.
The methods provide a means of post-launch validation in many cases.
The algorithms can be improved, with the aim of developing an automated data quality assessment system to run in concert with the established methods.
The methods seem applicable to a range of sensors, from high resolution small satellite systems to lower resolution missions such as S2 MSI and S3 OLCI.

30. Platform stability (micro-Vibe)
Question: Can we use specific radiometric calibration images to determine very small changes in platform stability in an automated manner? I will leave you to consider how.
0.1 Pixel Shift ? ? ? 10 50
100