% 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:1018123,
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 Rascher, Uwe and Scharr,
Hanno},
title = {{E}mulator-{B}ased {N}eural {N}etwork {P}rediction for the
{R}etrieval {O}f {S}un-{I}nduced {F}luorescence in the
{O}2-{A} {A}bsorption {B}and},
reportid = {FZJ-2023-04567},
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 = {Remote sensing applications rely on precise and efficient
corrections of atmospheric effects to retrieve surface
parameters such as reflectance, physiological quantities or
emission features. Often the retrieval is made more
difficult by the need for additional correction of the
illumination-viewing geometry. A well known approach to
solve such inversion tasks is to generate look-up tables of
simulated at-sensor radiance spectra with precise radiative
transfer models. The retrieval may then be expressed as a
spectral fitting of the observations. This contribution
investigates the performance of an encoder-decoder neural
network architecture to learn this optimization step in the
context of the retrieval of Sun-induced fluorescence (SIF)
in the O2-A oxygen absorption band for the ISS- based DESIS
spectrometer [1] and the airborne HyPlant instrument
[2].Recently, SIF has gained much interest in the wake of
the selection of the FLEX satellite mission by the European
Space Agency to be the first dedicated Earth Explorer
satellite mission for global SIF retrieval. The value of SIF
maps at multiple spatial scales to environmental and
agricultural use- cases are due to the close causal link
between SIF and plant photosynthesis. The present work aims
at show-casing the possibility to tightly integrate a neural
network with the domain knowledge of radiative transfer
codes. While DESIS has not been designed a-priori 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 two-module simulation tool
to simulate DESIS and HyPlant at-sensor radiance spectra in
the O2-A band, which is particularly sensitive to SIF
changes. The first module runs MODTRAN6 on a set of
atmosphere and geometry parameters from which atmospheric
functions can be derived. The second module models the SIF
emission as well as surface and sensor properties. A
detailed analysis of appropriate input parameter ranges and
dense sampling allowed to generate spectra covering the most
common acquisition conditions of HyPlant and DESIS.We
tightly integrate the physical domain knowledge intrinsic in
the simulation data base and the network architecture by
using an emulator of the simulations as non-trainable output
layer. Simple emulator models have been tested in order to
reduce the complexity of the regression problem due to its
high dimensionality (11d -> 13d for DESIS, 13d -> 349d for
HyPlant). The training of our encoder-decoder network on
simulated and observed at-sensor spectra is conducted in a
semi- supervised fashion, where we use the network to
estimate appropriate emulator input parameters. The
label-free loss evaluates the residuals between the
corresponding emulator outputs and the data. This
semi-supervised training set-up allows a direct comparison
of the performance on simulated and observed data, where no
SIF labels are available. Moreover, it has the advantage of
allowing emulator transformations to reduce domain gaps
between simulations and observations. A combined training on
simulated and observed data with label regularization is
left for future work.[1] K. Alonso et al, “Data Products,
Quality and Validation of the DLR Earth Sensing Imaging
Spectrometer (DESIS)”, Sensors, 19(20), 4471, 2019.[2] B.
Siegmann et al, “The high-performance airborne imaging
spectrometer HyPlant – From raw images to top-of-canopy
reflectance and fluorescence products: Introduction of an
automatized processing chain.”, Remote Sensing, 11, 2760,
2019.s},
month = {Jun},
date = {2023-06-12},
organization = {Helmholtz AI Conference, Hamburg
(Germany), 12 Jun 2023 - 14 Jun 2023},
subtyp = {After Call},
cin = {IBG-2 / IAS-8},
cid = {I:(DE-Juel1)IBG-2-20101118 / I:(DE-Juel1)IAS-8-20210421},
pnm = {5112 - Cross-Domain Algorithms, Tools, Methods Labs (ATMLs)
and Research Groups (POF4-511) / 2173 - Agro-biogeosystems:
controls, feedbacks and impact (POF4-217)},
pid = {G:(DE-HGF)POF4-5112 / G:(DE-HGF)POF4-2173},
typ = {PUB:(DE-HGF)6},
doi = {10.34734/FZJ-2023-04567},
url = {https://juser.fz-juelich.de/record/1018123},
}