Assimilation of solar‐induced fluorescence satellite observations in the <scp>European Centre for Medium‐Range Weather Forecasts</scp> integrated forecast system

Abstract The objective of this work is to assess the potential of the assimilation of satellite solar‐induced fluorescence (SIF) retrievals at eight‐day and 0.1° resolutions, derived from the TROPOMI instrument on board Copernicus Sentinel‐5P, in the integrated forecast system (IFS) developed at the European Centre for Medium‐Range Weather Forecasts (ECMWF), to update the IFS leaf area index (LAI) climatology and improve the IFS forecasts of gross primary productivity (GPP). This work represents a first attempt to assimilate SIF in a numerical weather prediction (NWP) framework at global scale by leveraging a machine‐learning (ML)‐based observation operator. Two configurations of the ML‐based observation operator, based on the XGBoost gradient‐boosted trees technique, were assessed: one trained using the IFS model fields which include the IFS LAI climatology, and one trained using only the actual satellite LAI observations along with spatiotemporal localization variables. The model directly trained on the satellite LAI shows the highest prediction scores. Besides, it produces more meaningful spatial patterns of LAI increments in response to climate anomalies compared to the model trained using the IFS predictors which do not include vegetation interannual variability. A key outcome is that the assimilation of SIF in the IFS improves LAI over cropland, which is promising for the monitoring of anthropogenic emissions over agriculture regions. However, at these spatial and temporal resolutions, the SIF ML model mainly resolves the seasonal variability of LAI which explains the degradations of LAI over tropical rainforests where the SIF signal is mainly driven by the variability in light use efficiency and the satellite LAI retrievals are more uncertain. Finally, the use of the updated LAI in the IFS coupled model leads to limited improvement of GPP forecasts, which is likely due to the prevailing effects of other sources of biases in the coupled IFS model.