Sensitivity Analyses

for each Rrs(&lambda) in NOMAD (also possible for satellite match-ups, Level-2 or -3 scene):
* calculate product(s) using reference model (ref)
* calculate product(s) using modified model(s) (mod)
* calculate relative % difference (RPD) = 200% x (mod - ref) / (mod + ref)
* retain all RPD for statistical analysis

generate frequency distributions, associated statistics, and series using compiled set of RPD

summary plots are presented below, with links to results for specific algorithms
algorithm names not supplied in summary plots to emphasize similarities in sensitivities

Rrs -> rrs(0-)

compare rrs(0-) = Rrs x 0.528 [mod] with = Rrs / (0.52 + 1.7 x Rrs) [ref]

observations:
* consistent bias, but most extreme RPD is < 10%
* little dependence on water type (see series plots)

see also this related forum post

T and S dependence of aw and bbw

compare aw(T,S) and bbw(T,S) [mod] with values from Pope and Fry 1997 and Morel 1974 [ref]
* consider 165 combinations of T and S ...
* ... 15 T from -2 to 40 C
* ... 11 S from 30 to 40 PSU

observations:
* adg(443) up to 20% lower and bbp(443) up to 20% higher using T,S dependent variables
* larger biases in bluer waters (see series plots)

(Un)correlated noise in Rrs

consider two scenarios where random noise is applied to Rrs(&lambda):
1. single value applied to all wavelengths (designated "correlated" noise below)
2. unique value applied to each wavelength (designated "uncorrelated" noise below)

noise distribution is normal with +/- 5% at 2-sigma
in IDL-ese: noise_multiplier = 1.0 - (randomn(seed,100) * 0.025)

uncorrelated spectral noise
consider 100 instances of Rrs x noise_multiplier with spectral dependence

observations:
* no directional bias
* consistent +/- 25% uncertainty for all products
* no dependence on water type (see series plots)
* Rrs(&lambda) shape more influential than magnitude (see below)

correlated spectral noise
consider 100 instances of Rrs x noise_multiplier with spectral independence

observations:
* no directional bias
* +/- 5% uncertainty for adg(443) and aph(443), but +/- 10% for bbp(443)
* no dependence on water type (see series plots)
* Rrs(&lambda) shape more influential than magnitude (see above)