TY  - CONF
AU  - Buffat, Jim
AU  - Pato, Miguel
AU  - Alonso, Kevin
AU  - Auer, Stefan
AU  - Carmona, Emiliano
AU  - Müller, Rupert
AU  - Rademske, Patrick
AU  - Rascher, Uwe
AU  - Scharr, Hanno
TI  - Leveraging deep learning for the retrieval of sun-induced fluorescence in the O$_2$-A absorption band from space
M1  - FZJ-2025-05291
PY  - 2025
AB  - Efficiently monitoring the photosynthetic activity of plants across large areas is a challenging task that has been addressed in the last decades as part of an ongoing research interest in the remote sensing of the environment and agricultural areas. The closest measurable variable to photosynthetic activity is the emission of sun-induced chlorophyll fluorescence (SIF). While SIF data retrieved from spaceborne hyperspectral sensors with low spatial resolution (> 1 km pixel size) has been used for drought monitoring, yield prediction, and the estimation of global gross primary productivity in the past, these sensors provide only limited insight into the fine-scale dynamics of photosynthesis. This has led to the preparation of the ESA Earth Explorer FLEX - the first spaceborne satellite mission dedicated to SIF retrieval - which promises to address this gap, offering SIF estimates at a significantly higher spatial resolution (300 m). In this contribution, we summarize our results on a novel deep learning based methodology to retrieve SIF from hyperspectral imagery at 760 nm that could benefit the research in computationally efficient SIF retrieval and validation of FLEX data.Our approach was developed and validated using high-quality data from the HyPlant FLUO sensor, the airborne demonstrator for FLEX FLORIS, which has a full width at half maximum (FWHM) of 0.25 nm. In this contribution, we show that the method can be adapted to data acquired by the spaceborne DESIS sensor (3.5 nm FWHM) onboard the International Space Station (ISS) providing for the first time spaceborne SIF estimates at 30 m. In a validation study conducted with a unique benchmark dataset consisting of HyPlant and DESIS at-sensor radiance acquired in a time interval of less than 25 minutes we find r$^2$ = 0.60 and a mean absolute difference between the SIF products of the two sensors of 0.78 mW nm$^{-1}$ sr$^{-1}$ m$^{-2}$ at 740 nm. Furthermore, we find r$^2$ = 0.20 in a cross-comparison of DESIS SIF with OCO-3 estimates.Our methodology involves the training of an encoder-decoder type neural network to perform a decomposition of the at-sensor radiance into surface and atmospheric parameters in the spectral window around the O2-A absorption band. Predicted parameters are evaluated in the observation space according to a novel label-free reconstruction based loss formulation. In the case of DESIS, we additionally introduce supervised regularizing loss terms to enhance the integration of existing L2A products. In order to be able to perform radiative transfer simulation as part of the training process we make use of a computationally efficient radiative transfer emulator that we derive from extensive hyperspectral radiative transfer simulation for both HyPlant and DESIS sensor configurations. This approach contrasts with traditional spectral fitting methods by its adoption of a feature-based strategy, enabling faster inference times.Due to the exceptional performance of this approach on DESIS data as well as its beneficial properties showcased in HyPlant data, our contribution has the potential to enhance research on computationally efficient SIF retrieval from FLEX data. Furthermore, its DESIS SIF product may provide a valuable high-resolution dataset for cross-comparison of FLEX SIF products in regions where spatial mixing impedes other validation approaches.
T2  - Living Planet Symposium 2025
CY  - 23 Jun 2025 - 27 Jun 2025, Vienna (Austria)
Y2  - 23 Jun 2025 - 27 Jun 2025
M2  - Vienna, Austria
LB  - PUB:(DE-HGF)24
DO  - DOI:10.34734/FZJ-2025-05291
UR  - https://juser.fz-juelich.de/record/1049209
ER  -