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

Abstract. Accurate representations of the surface and vegetation are critical for simulating the terrestrial CO₂ cycle in response to climate and meteorological conditions. To meet this challenge, an increasing number of satellite missions are being launched which can monitor vegetation conditions and biomass. One is the Copernicus Sentinel-5P mission, which carries the TROPOMI instrument and retrieves Solar-Induced Chlorophyll Fluorescence (SIF). As an indicator of plant photosynthetic activity, SIF can provide critical information for evaluating and parameterising gross carbon flux dynamics in surface models. This study aims to assimilate TROPOMI SIF data into the ISBA (Interactions between Soil, Biosphere and Atmosphere) land surface model developed by Meteo-France, with the objective of directly correcting the representation of Leaf Area Index (LAI) and Gross Primary Production (GPP). To achieve this, we have developed a dedicated observation operator that links the modelled LAI to the TROPOMI SIF daily product. This neural network operator was developed using deep learning and was trained using observations over Europe. This operator achieved good accuracy, was implemented in a land data assimilation system (LDAS), and was used to assimilate TROPOMI SIF in ISBA using a sequential simplified extended Kalman filter. Specific experiments were conducted to study the assimilation process over the Ebro basin in Spain. This area is known for its irrigated croplands, which are not well represented by ISBA. Some experiments assimilate TROPOMI SIF, some assimilate a Copernicus Land Monitoring Service (CLMS) LAI 10-day product, and some assimilate both. This provided us with a useful point of reference for improving the vegetation simulation. Following SIF assimilation, LAI representation improved across the domain, highlighting heavily irrigated croplands. The gross primary production (GPP) derived from the analysis is closing the gap between the simulated and observed values, though a significant difference remains. When compared with other assimilation experiments, assimilating SIF alone provides a similar benefit to standard assimilation of a CLMS LAI product on LAI and GPP. The best improvements to the LAI and GPP results come from co-assimilating TROPOMI SIF with the CLMS LAI product, which combines the advantages of high-frequency SIF observations and robust 10-day LAI assimilation.