Gast ::
| Hauptseite > Publikationsdatenbank > Towards fast and sensor-independent retrieval of sun-induced fluorescence fromspaceborne hyperspectral data |
| Hauptseite > Publikationsdatenbank > Towards fast and sensor-independent retrieval of sun-induced fluorescence fromspaceborne hyperspectral data |
| Poster (After Call) | FZJ-2024-06514 |
; ; ; ; ; ; ; ; ;
2024
Abstract: A corner stone to mapping photosynthetic dynamics efficiently over large areas of land is theretrieval of sun-induced fluorescence (SIF) from passive remote sensing data. In this contributionwe present a novel method to retrieve SIF from hyperspectral imagery that tightly integratesradiative transfer simulations and self-supervised neural network training. Differently to otherphysically constrained retrieval methods that optimize the parameters to a radiative transfermodel (RTM), it reduces the prohibitive computational cost of a physical model deployed tocontinuous data streams. To achieve this, it couples an emulator of large-scale radiative transfersimulations with a lightweight encoder-decoder neural network architecture and is trained byoptimizing a constraint based loss formulation.This method was developed and tested in the spectral region around the O2-A absorption bandon high-quality data acquired by the HyPlant sensor, the airborne demonstrator sensor for ESA’supcoming Earth Explorer satellite mission FLEX that aims to provide global hyperspectral imageryfor SIF retrieval. In a validation study with in-situ SIF measurements we find better performancethan the traditional Spectral Fitting Method (Cogliati et al. 2019). Furthermore, an adapted versionof our approach yields consistent SIF estimates on hyperspectral data of the spaceborne DESISsensor onboard the International Space Station (ISS). This result is encouraging since DESIS onlyprovides spectrally low-resolved imagery (2.55 nm SSD, 3.5 nm FWHM) compared to HyPlant(0.11 nm SSD, 0.25 nm FWHM). In a unique data set consisting of quasi-simultaneous, spatiallymatching DESIS and HyPlant acquisitions, the DESIS SIF estimates achieve a mean absolutedifference of less than 0.5 mW nm-1 sr-1 m-2 with respect to HyPlant derived estimates.Furthermore, the method yields SIF estimates that align well with the equally ISS-based OCO-3SIF product.The proposed methodology could benefit research in computationally efficient full-spectrum SIFprediction from FLEX data. While our method has been tested only in the O2-A absorption bandof HyPlant and DESIS acquisitions, principally it can be adapted in a straightforward fashion forretrieval in other spectral regions and in data from different sensors. Future work will thus includerecently published simulated FLEX imagery and our simulation tool developed for DESIS SIFprediction to gauge the method’s applicability in FLEX-like data.
|
The record appears in these collections: |