The study, published in the journal Biogeosciences by Pierre Vanderbecken of Météo-France and colleagues from institutions including the European Centre for Medium-Range Weather Forecasts, the Laboratoire des Sciences du Climat et de l’Environnement, and Forschungszentrum Jülich, tackles a long-standing problem in Earth system science. Land surface models, which simulate the exchange of water, energy and carbon between the ground, vegetation and the atmosphere, are the backbone of climate monitoring and carbon budget accounting. Yet the natural carbon flux from the land surface remains the most uncertain component of the global carbon budget, a gap that matters enormously for tracking whether countries are meeting their emission reduction commitments under the Paris Agreement.

The satellite at the heart of the work is the Copernicus Sentinel-5 Precursor, which carries an instrument called TROPOMI. With a vast 2,600-kilometre-wide swath and a heliosynchronous orbit, TROPOMI passes over any given region at least once a day, retrieving SIF in the near-infrared part of the spectrum within a narrow window between 743 and 753 nanometres. The retrieval works by statistically modelling the top-of-atmosphere radiance over vegetation and subtracting the contribution of the underlying bare surface, which means the resulting distribution of observations is centred on zero over barren ground and can occasionally dip into small negative, non-physical values. The instrument’s nadir measurement crosses the equator at 13:30 local solar time with a resolution of 7 by 3.5 kilometres, offering a daily, near-global picture of photosynthetic activity that no previous instrument could match.

There was, however, a catch. The ISBA land surface model, developed at Météo-France and distributed through the SURFEX modelling platform, simulates variables such as leaf area index, soil moisture and gross primary production, but it does not simulate SIF itself. To assimilate the satellite observations into the model, the researchers needed an observation operator: a mathematical bridge that translates model variables into the space of the observations. A physically based SIF operator would have to solve complex radiative transfer processes and would demand heavy computational resources and careful calibration, making it impractical for an operational system. The team’s alternative was a machine learning approach, adapting a methodology previously used to assimilate radar backscatter observations from the ASCAT satellite.

The resulting neural network is elegantly simple in concept but sophisticated in its training. It is a feed-forward network with two hidden layers of 128 neurons each, using ReLU activation, and a single linear output neuron. A Gaussian noise layer, active only during training, was inserted before the first activation to reproduce the effect of noisy inputs and to prevent overfitting, while batch normalisation layers at the end of each hidden layer regularised the learning process. Because the distribution of TROPOMI SIF values resembles a log-normal distribution with a mode close to zero, the network was trained on a transformed version of the signal to avoid non-physical negative predictions. The loss function was a Huber loss, which behaves like a mean-square loss for small errors but switches to a more linear behaviour for extreme differences, making the training less sensitive to outliers in the noisy satellite data.

Crucially, the network takes only four inputs: the day of the year, the latitude, the longitude, and the observed leaf area index from the Copernicus Land Monitoring Service. This minimal design means the operator can be plugged into any land surface model and will remain robust as those models improve. It was trained on a vast database covering Europe from 26 degrees west to 46 degrees east and 28 to 72 degrees north on a regular 0.1-degree grid, drawing on TROPOMI SIF retrievals from the TROPOSIF project spanning 2018 to 2021. After screening out oceans, lakes, snow, urban areas, high-altitude regions and frozen ground, the network learned from roughly 14 million input-output pairs, trained on one year of data and tested on a completely different year to verify that it could generalise. The results were strong: a Pearson correlation above 0.8 between the network’s estimates and the observed SIF, and a root mean square error of around 0.15 milliwatts per square metre per steridian per nanometre, with no evident overfitting between the training and test years.

With the trained operator in hand, the team embedded it in a land data assimilation system that uses a simplified extended Kalman filter to nudge the ISBA model towards the observations on a daily cycle. The control vector comprises the leaf area index and soil moisture in different soil layers, and the Kalman gain weighs the background error against the observation error to compute each day’s analysis. The proving ground was the Ebro basin in Spain, a region famous for its heavily irrigated croplands, which conventional land surface models struggle to represent because irrigation is a human decision, not a natural process the model knows about. The researchers ran a suite of experiments: an open-loop simulation with no assimilation, a baseline assimilating the standard 10-day leaf area index product with a fixed 20 percent relative error, several experiments assimilating TROPOMI SIF alone with different observation error assumptions, and co-assimilation experiments combining both data streams.

The results were striking. After SIF assimilation, the leaf area index improved across the domain, with analysis increments reaching up to 0.2 square metres per square metre over the C3 croplands. Most tellingly, the assimilation revealed a confined area around the Ebro basin where the leaf area index increased during the summers, a pattern repeated in every year of the experiment and corresponding to artificially irrigated land. Irrigation sustains both the fluorescence and the foliage while surrounding areas dry out, and the satellite observations captured this signature even though the model never simulated it. The residuals, meaning the differences between the analysed SIF and the observations, were consistently smaller than the innovations, the differences before assimilation, confirming that the filter was genuinely pulling the model towards reality. Over irrigated pixels, the late-summer regrowth of vegetation that the open-loop simulation completely missed was faithfully reproduced in the analysis.

Perhaps the most surprising finding concerned gross primary production, the gross carbon flux that SIF is famous for tracking. Although assimilating SIF improved the correlation with the independent FLUXCOM-X GPP product on average across the domain, the local gains were modest, and a marked difference remained between simulated and observed values. The reason lies in the internal logic of the model: simulated GPP depends not only on the vegetation state but also on soil water content. When SIF assimilation added extra leaf area in summer, the model’s soil dried out faster and evapotranspiration increased, which prevented the computed GPP from rising, even though the observed product captured a second late-summer surge driven by irrigation. Because the assimilation cannot infer an increase in soil water supply that the model does not represent, SIF alone could not close that final gap. The comparison with airborne SIF measurements from the HyPlant campaign also showed that the neural network is an emulator, not a substitute for a physically based fluorescence model.

The broader lessons are twofold. First, assimilating daily TROPOMI SIF alone delivered benefits comparable to the standard assimilation of the 10-day leaf area index product, while providing roughly five times more observations to the system, and an observation error of 20 percent proved close to optimal. Second, and most powerfully, co-assimilating SIF with the leaf area index product outperformed either observation alone, combining the high-frequency, day-to-day sensitivity of fluorescence with the robust, low-noise anchor of the 10-day vegetation product. In the co-assimilation framework, a more complex error description for SIF even improved the results further. The authors suggest the approach could be extended to other SIF missions and to additional satellite products such as land surface temperature. For a scientific community racing to build a near-real-time monitoring and verification capacity for the carbon cycle, the message is clear: the faint glow of photosynthesis, filtered through a modest neural network, can now help models see the human fingerprints on the land surface, one irrigated field at a time.

Subject of Research: Deep learning-based assimilation of satellite sun-induced chlorophyll fluorescence into a land surface model to improve vegetation and carbon flux simulation

Article Title: Using deep learning to assimilate sun-induced fluorescence satellite observations in the ISBA land surface model

Article References: Using deep learning to assimilate sun-induced fluorescence satellite observations in the ISBA land surface model. (n.d.). https://doi.org/10.5194/bg-23-6687-2026

Image Credits: AI Generated

DOI: 10.5194/bg-23-6687-2026

Keywords: sun-induced fluorescence, TROPOMI, Sentinel-5P, data assimilation, deep learning, neural network, land surface model, ISBA, leaf area index, gross primary production, irrigation, carbon cycle