1. Combination Approaches
GSICS Annual Meeting
29th Feb.- 4th March 2016, Tsukuba, Japan
Combination Approaches
Bertrand Fougnie
for CNES Calibration Center
(CNES-DCT/SI/MO + support from CNES-DCT/ME)

2. Introduction
From an initial question addressed on previous
GSICS Webmeeting /Workshop
GSICS intends to implement various reference methods
various results will be available
Synergy : How to merge results from various methods ?
Is it possible to derive a unique approach ?
Which answers could be derived ?
GSICS must define/construct a clear strategy

3. Observation Indicative behavior of targets
Several calibration methods could be available
Cloud-DCC, Moon, PICS-Desert, Rayleigh, Sunglint, PICS-Antarctica, SNO, Ray-matching, Land…
Each target has its own behavior :
Magnitude: from very dark to very bright
Spectral shape : from white to very pronounced
Angular signature : from nearly uniform to large BRDF
Polarized properties : from non-polarized to nearly fully polarized
Short-term stability : from variable to fully stable
Long-term stability : from seasonal variable to fully stable
So efficiency range …
Indicative behavior of targets
May sensitively vary with various parameters

4. Observation Several calibration methods could be available
Cloud-DCC, Moon, PICS-Desert, Rayleigh, Sunglint, PICS-Antarctica, SNO, Ray-matching…
Implement several methods will provide
various results which will differ (in general)
sometimes consistant, sometimes not at all
What’s the definition of « consistant » for GSICS needs ?
It is often called “calibration error” a radiometric artifact which is not a calibration error
 Straylight, spectral rejection, non-linearity, variation with scan, polarization…
They are radiometric errors, but not calibration errors
They are radiometric biases, but varying with every different situation
They are not full instrumental biases
 What do we mean by “calibration” ?
GSICS has to face the way to provide to users the results : alternatives
One single , the best
One blended
Multiple sets
 What’s the best alternative for GSICS needs ?

5. Example
The single
The multiple set
The Blend

6. To be evaluated for each band
The One-method matrix
Sensor to calibrate
To be evaluated for each band
Uncertainty from implemented method
(depending on data sampling)
Uncertainty from sensor
Characterization
to be addressed
Spectral response knowledge
Straylight
Linearity
Polarization
Radiometric noise …
Trending
Absolute
Interband
Cross-calibration …
DCC

7. To be evaluated for each band
The Synergy matrix
Sensor to calibrate
To be evaluated for each band
Uncertainty from implemented method
(depending on data sampling)
Uncertainty from sensor
Characterization
to be addressed
DCC
Moon
PICS-desert
PICS-snow
SNO
Rayleigh
Sunglint
Spectral response knowledge
Straylight
Linearity
Polarization
Radiometric noise …
Trending
Absolute
Interband
Cross-calibration …

8. Example : combination of Desert, DCC and Rayleigh
The Synergy matrix
Example : combination of Desert, DCC and Rayleigh
for MERIS and PARASOL
Sensor to calibrate
To be evaluated for each band
Uncertainty from implemented method
(depending on data sampling)
Uncertainty from sensor
Characterization
to be addressed
DCC
Moon
PICS-desert
PICS-snow
SNO
Rayleigh
Sunglint
Spectral response knowledge
Straylight
Linearity
Polarization
Radiometric noise …
Trending
Absolute
Interband
Cross-calibration …

9. 443nm band
as a function of VZA
MERIS
PARASOL
DCC
Desert
Rayleigh

10. Indicative map for a synergy
Generale baseline

11. Example Application to PARASOL (According Fougnie, IEEE, 2016)
Strategy used for the PARASOL revision
trending and variation within FOV for band 765 using desert and snow
trending for all other interbands based on a compromise between methods
Compromise = find for each band the adjustment that minimises all stdev with a high weight for methods with stdev < 0.6%
variation within FOV for all other interbands using clouds
validation with all other methods
checking of the consistency of absolute radiometric calibration between all methods
(According Fougnie, IEEE, 2016)

12. Ex: PARASOL - Monitoring
Best compromise
B490 = 0.16
D=0.018
B670 = 0.062
Synergy : calibration of the temporal monitoring
Calibration versus month
B565 = 0.11
B865 = 0.024
B1020 = 0.018
B765 = 0.01

13. Combination Combination for Trending Synergy for Absolute Calibration
Absolute adjusment (Fougnie et al., 2007):
mean of 490/565/670 over Rayleigh
NIR bands over mean of clouds/sunglint
Other methods for validation

14. Discussion – Questions
Derive results from various methods
How to define the mean value, or weighted value ? How to define corresponding weights ?
Consider theoretical accuracy of methods, as well as sensor behavior
Is the mean (or weighted) result the most optimal result for GSICS needs ?
How discrepancies will be interpreted ? Ignored ?
Will GSICS try to understand what radiometric artifacts could be ?
This is the job of every agencies or GPRC
Not only a question of maturity for the method, but also maturity for the result itself
If results are consistent between methods
Do we keep the best theoretical method ?
Do we derive a mean/blend in order to reduce residual bias ?
If results differ between methods :
Do we stay at the level where agency has to investigate the radiometric artifact
but try to select the “most relevant” result ?
Do we propose all results and clarifying the “applicability” range of each ?

15. Best method or weighted mean Radiometric explanation
The Synergy matrix
Sensor to calibrate
To be evaluated for each band
Uncertainty from implemented method
(depending on data sampling)
Uncertainty from sensor
Characterization
to be addressed
DCC
Moon
PICS-desert
PICS-snow
SNO
Rayleigh
Sunglint
Spectral response knowledge
Straylight
Linearity
Polarization
Radiometric noise …
Trending
Absolute
Interband
Cross-calibration … 1
One
calibration coefficient
per band per time …
Best method or weighted mean
GSICS Output
Radiometric explanation

16. Sorry, more questions than answers…
Wrap-up
In general, derive results from various (accurate) methods = provide different information from the radiometry
In general, results will slightly/largely differ
Probably, no universal approach can be defined in advance for combining
At least at the beginning, this would be a case-by-case analysis and conclusion
GSICS must define how the various calibration sets will be used (see previous list of questions)
Unclear information toward users may endanger the future of calibration correction proposed by GSICS
Sorry, more questions than answers…
« This is not Science, this is Art… » (Dave Doelling)

17. Back up slides