CoAuthors

Jean-Marc Delouis (LOPS - CNRS , France)

Data from space observations have instrumental effects or foregrounds
that degrade the signal of interest recovery. The SRoll software
developed initialy for cosmology (Planck Project) simultaneously inverts
this signal and the instrumental effects including noises. Therefore,
SRoll uses the all available data set to minimise the variance of the
measured calibration parameters. This problem is common to astrophysical
and oceanographic data as with CFOSAT, but the geophysical signal
evolves over time on the duration of the measurement, unlike astrophysical
processes. Therefore, we present a new version of SRoll based on
Scattering Transforms (ST) to model the signal dynamics thanks
to its reduction dimension abilities. STs have been
used successfully on interstellar medium to statisticaly describe the intermittent
structure of the turbulent processes. These statistical
constraints are measured on the data itself and therefore there is no learning phase as
with usual neural networks. This work presents the encouraging results
obtained with SRoll on one month of CFOSAT data to correct
the calibration of the SWIM instrument.