% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Buffat:1018121,
author = {Buffat, Jim and Pato, Miguel and Alonso, Kevin and Auer,
Stefan and Carmona, Emiliano and Maier, Stefan and Müller,
Rupert and Rademske, Patrick and Scharr, Hanno},
title = {{A} {N}ovel {S}elf-{S}upervised {S}un-{I}nduced
{F}luorescence {R}etrieval {U}sing {S}imulated {H}y{P}lant
and {DESIS} {D}ata},
reportid = {FZJ-2023-04565},
year = {2023},
note = {The authors gratefully acknowledge the computing time
granted through JARA on the supercomputer JURECA[1] at
Forschungszentrum Jülich.[1] Jülich Supercomputing Centre.
(2021). JURECA: Data Centric and Booster Modules
implementing the Modular Supercomputing Architecture at
Jülich Supercomputing Centre Journal of large-scale
research facilities, 7, A182.
http://dx.doi.org/10.17815/jlsrf-7-182},
abstract = {Operationally efficient retrieval of sun-induced
fluorescence (SIF) from remote sensing data requires exact
atmospheric correction at low computational cost to
compensate for atmospheric effects at play in the radiative
transfer of light through the atmosphere. Incomplete
knowledge not only of the atmospheric state at acquisition
time, but also of surface conditions imply the formulation
of SIF retrieval as a simultaneous search of the
atmospheric, viewing geometry and surface related variables
describing best the observed scene.Such a problem can be
addressed by the construction of large look-up tables (LUT)
covering the space of possible input parameters describing
the hyperspectral observations. The SIF retrieval reduces to
spectrally fitting this LUT to the observations. In the
present contribution we investigate the use of a neural
network to perform this optimization step. We show-case the
possibility to tightly integrate a neural network with the
domain knowledge of radiative transfer codes simulating
observations of the airborne HyPlant instrument and the
ISS-based DESIS spectrometer in a spectral window around the
O2-A oxygen absorption band (740-780 nm). While DESIS has
not been designed for SIF retrieval, HyPlant is a testing
instrument for the FLORIS spectrometer onboard FLEX. The
present work can thus make use of the insights gained on
high-resolution HyPlant data for the more challenging SIF
prediction on DESIS acquisitions.We have developed a generic
simulation tool based on MODTRAN6 to simulate at-sensor
radiance in the O2-A oxygen absorption band. The tool
consists of two modules. A primary module simulates
distributions of atmospheric functions under varying
atmospheric state variables. The high spectral resolution of
DESIS (full width at half maximum 3.5 nm) and HyPlant (0.25
nm) required us to run MODTRAN6 at the finest available
resolution. A secondary module modeling the sensor and
surface characteristics factors in expert knowledge on DESIS
and HyPlant as well as constraints on the SIF signal. We
sampled an exhaustive range of observable atmospheric and
surface conditions based on an extensive sensitivity
analysis of key input parameters.Due to its computational
complexity and cost the simulation tool can not be included
in the network predictor directly. In order to tightly
integrate the physical domain knowledge with a neural
network predictor we tested a set of models to emulate
at-sensor radiance using the simulated database. The
high-dimensional nature of the regression problem (11d ->
13d for DESIS, 13d -> 349 for HyPlant) required us to test
relatively simple models. We found fourth-degree polynomials
to perform best with at-sensor radiance residuals of less
than 0.01 mW/cm2/sr/μm on average across our simulated test
set for both DESIS and HyPlant. Importantly, the per sample
prediction time is of the order 10 μs, i.e. 107 times
faster than the original simulation. We integrated this
emulator as a prediction head into an encoder-decoder type
neural network and trained the network in a first step on
observed HyPlant imagery. Application of this approach to
DESIS data is left for future work. A self- supervised loss
evaluating the radiance residual of the emulated
reconstruction and a physiological constraint was
implemented to this end. Furthermore, the inclusion of
topographic variation in the simulation data base as well as
spatial constraints in the network architecture allow for
SIF prediction even in topographically complex
terrain.Validation of our self-supervised method’s
predictions with ground based SIF measurements and
comparison with the state-of-the-art Spectral Fitting Method
(SFM) and Improved Fraunhofer Line Detection (iFLD) baseline
methods revealed that the presented approached yielded
comparable correlation scores. Linear calibration is,
however, still needed to reach state-of-the-art absolute
prediction accuracy as the SIF predictions suffer from
systematic biases. Further plausibility studies highlight
that the network yields physiologically plausible diurnal
SIF dynamics and compensates for changes in the atmospheric
path in topographically complex terrain.In summary, we have
developed a SIF prediction approach integrating a
computationally efficient and precise emulator in a neural
network architecture derived from extensive sampling and
high precision simulations of observational conditions.
Validation and plausibility studies could show- case its
successful application to HyPlant imagery.},
month = {Sep},
date = {2023-09-19},
organization = {Flex Fluorescence Workshop 2023,
Frascati (Italy), 19 Sep 2023 - 21 Sep
2023},
subtyp = {Invited},
cin = {IBG-2 / IAS-8},
cid = {I:(DE-Juel1)IBG-2-20101118 / I:(DE-Juel1)IAS-8-20210421},
pnm = {2173 - Agro-biogeosystems: controls, feedbacks and impact
(POF4-217) / 5112 - Cross-Domain Algorithms, Tools, Methods
Labs (ATMLs) and Research Groups (POF4-511)},
pid = {G:(DE-HGF)POF4-2173 / G:(DE-HGF)POF4-5112},
typ = {PUB:(DE-HGF)24},
doi = {10.34734/FZJ-2023-04565},
url = {https://juser.fz-juelich.de/record/1018121},
}