Conference Agenda

The upcoming FLEX mission will provide spectrally resolved Sun-induced fluorescence (SIF) observations, opening new opportunities to quantify photosynthetic function and stress from space. To fully exploit FLEX data, these SIF measurements must be interpreted together with robust, uncertainty-aware essential vegetation traits (EVTs) products derived from the tandem-orbiting Sentinel-3 (S3) mission. These EVTs provide key structural, biochemical, and radiative context for FLEX Level-2 SIF retrievals and FLEX-related Level-3/4 products on photosynthetic activity and ecosystem productivity.

Here, we present three Gaussian Process Regression (GPR) model families, each tailored to a distinct training dataset released between 2022 and 2025, and designed for operational retrieval of EVTs from S3: (1) simulation-based top-of-atmosphere models (S3-TOA-GPR), (2) hybrid models trained on S3 SYNERGY surface reflectance (SCOPE-SYN-GPR), and (3) empirical models trained on upscaled GBOV reference data (GBOV-SYN-GPR). The models target leaf chlorophyll content (LCC), leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), and fractional vegetation cover (FVC), which jointly constrain canopy structure, light absorption, and photosynthetic capacity.

The S3-TOA-GPR models are trained on radiance simulations from the coupled SCOPE–6SV radiative transfer framework and implemented directly in Google Earth Engine (GEE). This allows direct retrieval of EVTs from OLCI TOA data, bypassing explicit atmospheric correction. Multi-site validation against MODIS products, comparison with Copernicus Global Land Service (CGLS) maps, and evaluation using VALERI field data demonstrate coherent phenological dynamics and spatial consistency.

To exploit the operational S3 SYNERGY Level-2 surface reflectance product (SY_2_SYN, 300 m), SCOPE-SYN-GPR models were trained using SCOPE simulations, and FAPAR/FVC models were validated against the global Copernicus Ground-Based Observations for Validation (GBOV) dataset. Intercomparison with CGLS FAPAR/FVC products shows strong agreement, with SCOPE-SYN-GPR slightly outperforming existing solutions. These models are deployed on cloud infrastructures through the openEO platform, enabling continental-scale mapping with Bayesian uncertainties and standardized programmatic access.

Finally, to address the limitations of SCOPE in capturing conditions characteristic of heterogeneous landscapes, we have transitioned to fully empirical GPR models. GBOV-SYN-GPR models (LAI/FAPAR/FVC) were trained on more than 3,650 GBOV samples across Europe and validated with independent GBOV data (2022–2024, over 1700 GBOVs samples). Performance is biome dependent, with high accuracy for croplands and persistent challenges in dense evergreen forests, consistent with expected saturation and separability limitations (as also observed in previous GPR models). A key feature of this framework is the possibility for explicit decomposition of predictive uncertainty into epistemic (model-driven) and aleatoric (observation-driven) components, revealing structured spatial patterns linked to training representativeness and observation conditions.

All three model families are implemented in the open-source PyEOGPR Python package (https://pypi.org/project/pyeogpr/), providing a reproducible, extensible, and cloud-ready toolkit for scalable vegetation trait retrieval on GEE, openEO and other EO platforms. Together, these S3 GPR-based products deliver the structural and functional information needed to contextualize FLEX SIF signals, support algorithm evaluation at Level-2, and enable the generation of FLEX-relevant Level-3/4 indicators of photosynthetic stress and productivity.

Atmospheric radiative transfer models (RTMs) simulate how light propagates through the Earth’s atmosphere by accounting for interactions with gases, aerosols, and clouds. They are fundamental to remote sensing and satellite data processing because of their physical accuracy. However, this rigor comes with a high computational cost, which limits their operational use. A common workaround is the use of look-up tables (LUTs) of stored RTM simulations. While LUT interpolation speeds up computations, it requires very large datasets to preserve accuracy, leading to high generation times and memory demands. These issues become particularly severe for hyperspectral missions, which often need mission-specific LUTs and limit the generalization of processing algorithms. A prominent example is ESA’s FLEX Earth Explorer mission, which aims to retrieve sun-induced fluorescence using high spectral resolution observations in the O2 absorption bands. Achieving the required atmospheric correction accuracy for FLEX demands densely sampled LUTs based on high-resolution RTM simulations, posing a major challenge for operational processing.

Machine-learning-based emulators have emerged as a promising alternative. Instead of interpolating LUTs, emulators statistically approximate RTM outputs through regression models, offering large gains in speed while maintaining high accuracy. However, the high dimensionality and physical complexity of RTM data make emulation challenging, leading to a wide range of proposed methods that trade off accuracy, interpretability, and computational efficiency.

To advance this field, the second Atmospheric Machine Learning Emulation Challenge (AMLEC-2) is being organized in collaboration with ESA’s phi-lab. The goal of AMLEC-2 is to engage the remote sensing and machine learning communities in developing and benchmarking RTM emulators using common datasets and standardized evaluation protocols. This edition focuses on the FLEX mission and evaluates emulators based on prediction accuracy, runtime performance, and uncertainty quantification.

AMLEC-2 is structured into four phases. First, the challenge design, datasets, and evaluation methodology are defined. Second, RTM simulation datasets are generated. Third, over three months, participants train their emulators and submit predictions, with automated evaluations and regularly updated rankings encouraging iterative improvement. In the final phase, results are analyzed and potentially presented in a peer-reviewed publication. Participants are involved only in the third phase, which includes downloading data, training models, generating predictions, and submitting results.

Although the challenge is still in the design stage at the time of writing, it will be in its third phase at the time of the FLEX Fluorescence Workshop. We will present the organisation of the challenge, the evaluation methodology, and the test datasets. We expect that this contribution will encourage the FLEX community to participate in AMLEC-2, ultimately supporting the development and implementation of novel data processing algorithms for the FLEX mission.

The retrieval of solar-induced fluorescence (SIF) from remote sensing measurements requires accurate modeling of surface, atmospheric and sensor effects in order to separate the feeble fluorescence signal from the dominant reflected light component. Radiative transfer modeling has been routinely used to achieve the required level of accuracy for SIF retrieval, but it involves time-consuming calculations typically performed offline. The necessarily slow radiative transfer step presents a fundamental obstacle to the full exploitation of modern machine and deep learning SIF retrieval algorithms, which imply training or processing of large amounts of data. This challenge is addressed here by developing a fast and accurate physics-based emulator of at-sensor radiances. We start by introducing a general-purpose simulation framework and then apply it to simulate at-sensor radiances around the O₂-A absorption band (740–780 nm) for two representative sensors: HyPlant (airborne, limited spatial coverage, spatial resolution of ~1 m, spectral resolution of ~0.25 nm) and DESIS (space-based, global spatial coverage, spatial resolution of 30 m, spectral resolution of ~3.5 nm). Millions of HyPlant and DESIS spectra are simulated across a meaningful range of atmosphere, geometry, surface and sensor characteristics. We then show that a simple polynomial model trained on these extensive datasets is an excellent forward emulator of at-sensor radiances at the O₂-A band for both HyPlant and DESIS with prediction times per spectrum as low as 10 µs and typical errors of at most 10% of a typical SIF signal. Importantly, the emulator can closely reproduce measurements acquired by HyPlant and DESIS thereby underlining its suitability for interpreting real data. These findings are of relevance for novel SIF retrieval schemes, opening the way for the swift generation of high-fidelity training datasets, the realization of a fast and accurate simulation step and the performance assessment of any retrieval method. The recently developed SFMNN SIF retrieval algorithm implements a self-supervised method employing the emulator presented here as the forward simulation step and has been shown to successfully retrieve SIF from HyPlant and DESIS data (see contribution by Buffat et al). Additionally, we quantitatively assess the SIF retrieval performance of the 3FLD method on simulated data and illustrate the optimization of a retrieval method using our framework. Finally, we comment on how our emulators can be extended to the FLORIS instrument and beyond the O₂-A band providing a promising avenue to derive SIF from FLEX data with SFMNN or other novel SIF retrieval methods.

The development, test and validation of retrieval methods for Solar Induced Fluorescence (SIF), especially full spectrum retrievals, require accurate radiative transfer models that incorporate fluorescence. Due to the small contribution of SIF to the overall signal, very high model accuracy is required.

SLOPE is a physics based, deterministic, leaf level radiative transfer model using independently derived specific absorption coefficients for leaf pigments and other constituents. Therefore, SLOPE produces accurate outputs using only biophysical variables like pigment concentrations as inputs. It does not require adjustment of tuning parameters which would require a number of simultaneous measurements of pigment concentrations and reflectance and fluorescence spectra.

SLOPE takes into account the non-homogeneous distribution of pigments across the leaf cross-section that is exhibited by most dicotyledon and some monocotyledon leaves. It also takes into account that pigments like chlorophylls are not homogeneously distributed laterally but rather concentrated (clumped) in chloroplasts and thylakoids. Also in SLOPE emphasis has been placed on accurate representation of the red absorption maximum and red-edge spectral regions.

This presentation will firstly provide a brief description of the model structure. Secondly, results of intercomparisons with reflectance, transmittance and fluorescence measurements will be given. The third part will demonstrate the influence of pigment concentrations on reflected light and fluorescence emission. Furthermore, the impact of non-homogeneous vertical and lateral distribution of pigments on fluorescence emission will be shown.

Originally, SLOPE was implemented in C. To facilitate easier and more widespread use it is currently being re-implemented in Phyton. All source code is fully open source and is not subject to any licensing restrictions.

The use of SIF to evaluate land surface models shows considerable promise to help constrain estimates of, and elucidate the processes that control Gross Primary Productivity (GPP) on large spatial scales. To use SIF effectively for this purpose, we argue that forward modelling of the observations from the land surface model – as opposed to, say, relying on empirical relationships with the modelled GPP – is desirable if we wish to understand structural deficiencies in the land surface model.

This presentation describes the prediction of SIF from JULES, the Joint UK Land Environment Simulator, which is the land surface scheme of the Hadley Centre climate models, and the UK Earth System Model (UKESM). We explain how we couple leaf-level SIF models to the biochemistry routines in JULES, and how we scale the emitted SIF to the canopy level using a vegetation radiative transfer scheme (L2SM) that is consistent with the physics inside JULES but also allows for radiative emissions within the canopy. The SIF scheme includes attenuation within the leaf, utilizing either modelled or observed leaf reflectance and transmittance spectra and can make predictions of the canopy leaving SIF at arbitrary wavelengths. Downregulation of fluorescence by water stress is also included.

We show results JULES-SIF from the recent SIFMIP exercise at site level, and also regional comparisons against TROPOSIF data. The results show a generally good agreement and are sufficiently aligned with the observations that they are able to highlight areas where JULES is not correctly modelling the relevant environmental processes. Future directions for the JULES SIF module are explained, including accounting the directional component of the canopy leaving SIF.

We present the Terrestrial Carbon Community Assimilation System (TCCAS), funded by the European Space Agency within its Carbon Science Cluster. TCCAS is built around the newly developed D&B terrestrial biosphere model (Knorr et al., 2025). D&B builds on the strengths of each component model in that it combines the dynamic simulation of the carbon pools and canopy phenology of DALEC with the dynamic simulation of water pools, and the canopy model of photosynthesis and energy balance of BETHY. A suite of observation operators allows the simulation of solar-induced fluorescence (SIF), fraction of absorbed photosynthetically active radiation, vegetation optical depth from passive microwave sensors, the slope of the backscatter-incidence angle relationship of an active microwave sensor, surface layer soil moisture, and Net Ecosystem Production. The model is embedded into a variational assimilation system that adjusts a control vector to match the observational data streams. For this purpose TCCAS is provided with efficient tangent and adjoint code. The control vector consists of a combination of initial pool sizes and process parameters in the core model and in the observation operators. The observation operator for SIF simulates the radiative transfer within the canopy-soil system by the layered two stream model (Quaife, 2025). It offers leaf-level SIF formulations according to Gu et al. (2019), van der Tol et al. (2014) or Li et al. (2022). We show assimilation experiments of the TROPOSIF product derived from Sentinel 5P using each of the three formulations and discuss their performance. TCCAS and D&B are available as open source community tools with documentation and training events.

Gu, L., Han, J., Wood, J.D., Chang, C.Y.Y. and Sun, Y., 2019. Sun‐induced Chl fluorescence and its importance for biophysical modeling of photosynthesis based on light reactions. New Phytologist, 223(3), pp.1179-1191.

Knorr, W., Williams, M., Thum, T., Kaminski, T., Voßbeck, M., Scholze, M., Quaife, T., Smallman, T. L., Steele-Dunne, S. C., Vreugdenhil, M., Green, T., Zaehle, S., Aurela, M., Bouvet, A., Bueechi, E., Dorigo, W., El-Madany, T. S., Migliavacca, M., Honkanen, M., Kerr, Y. H., Kontu, A., Lemmetyinen, J., Lindqvist, H., Mialon, A., Miinalainen, T., Pique, G., Ojasalo, A., Quegan, S., Rayner, P. J., Reyes-Muñoz, P., Rodríguez-Fernández, N., Schwank, M., Verrelst, J., Zhu, S., Schüttemeyer, D., and Drusch, M.: A comprehensive land-surface vegetation model for multi-stream data assimilation, D&B v1.0, Geosci. Model Dev., 18, 2137–2159, https://doi.org/10.5194/gmd-18-2137-2025, 2025.

Li, R., Lombardozzi, D., Shi, M., Frankenberg, C., Parazoo, N.C., Köhler, P., Yi, K., Guan, K. and Yang, X., 2022. Representation of Leaf‐to‐Canopy Radiative Transfer Processes Improves Simulation of Far‐Red Solar‐Induced Chlorophyll Fluorescence in the Community Land Model Version 5. Journal of Advances in Modeling Earth Systems, 14(3), p.e2021MS002747.

Quaife, T. L. (2025). A two stream radiative transfer model for vertically inhomogeneous vegetation canopies including internal emission. Journal of Advances in Modeling Earth Systems, 17, e2024MS004712. https://doi.org/10.1029/2024MS004712

van der Tol, C., Berry, J. A., Campbell, P. K. E., and Rascher, U.: Models of fluorescence and photosynthesis for interpreting measurements of solar-induced chlorophyll fluorescence, Journal of Geophysical Research: Biogeosciences, 119, 2312–2327, doi:10.1002/2014JG002713, URL: https://onlinelibrary.wiley.com/doi/abs/10.1002/2014JG002713, 2014.