Here is some news.
I finaly applied the sun glint removal technique from this paper http://www.marinespatialecologylab.org/wp-content/uploads/2010/11/Hedleyetal_2005.pdf. The process seems easier but it is also the one used in the sen2coral toolbox of snap with sentinel 2 imagery (However it does not work on my Snap ...).
It consist of extracting a sample area displaying range of sun glint in homogeneous water such as deep water. A linear regression is computed between each bands to correct and one NIR band. Lastly a minimum NIR value is extrated from the sample. Correction is then processed as followed :
R'i = Ri - b_i*(R_nir - Min_R_nir)
with R'_i, the corrected pixel in band i, R_i the pixel to be corrected in band i, b_i the slope of regression between NIR and band i, R_nir the near-infrared signal of the pixel, and Min_R_nir the minimun of the near infrared signal over the extracted sample.
In my case I computed a separated regression for each detector footprint but choose the global Min_R_nir over all samples(one per footprint) in order to obtain same range of intensity.
Results are pretty good regarding the sun glint removal :
However the detectors footprints are still visible:
The difference are slight but big enough to expect statistical difference between the footprint which risk to be problematic for the next parts of my work(involving classifications).
The red component seems to be the more distingish diffenrence between footprints so I took a look on the regressions :
Indeed, regressions of Red band over NIR band shows the lowest coefficients of determination(despite the fact that their wavelength are the closest ...).
At the end, I'm asking myself if the specular reflection behaving differently due to the detectors angles is the only cause responsible for detector footprints difference. I imagine that each detectors are undergoing instrumental noise. Is this noise corrected ?