Integrating Machine Learning, Remote Sensing, and Eddy Covariance Data to Assess Vegetation Parameters and Carbon Fluxes

Understanding vegetation structure and carbon fluxes is essential for assessing ecosystem functioning and biogeochemical cycles. Alpine ecosystems pose particular challenges due to extreme environmental conditions and limited data availability. This thesis presents two complementary research lines integrating remote sensing, machine learning, and in situ measurements to improve the estimation of key vegetation parameters and carbon fluxes. The first line focuses on estimating Leaf Area Index (LAI) and Fractional Cover (FC) using the SBG-TIR mission optical data and a machine learning algorithm trained on synthetic reflectance from the SCOPE radiative transfer model. The algorithm achieved excellent performance on synthetic validation data, with RMSE of 0.046 for FC and 0.053 for LAI. When applied to real observational data, the retrieval yielded RMSE of 0.19 for FC and 1.02 for LAI, consistent with state-of-the-art operational methods. These results confirm that combining physically-based simulations with data-driven learning allows robust and transferable estimation of vegetation structural parameters in spectrally complex and heterogeneous environments. The second line addresses local carbon flux quantification in alpine ecosystems, focusing on the Torgnon site (Aosta Valley). A local optimization of the FLUXCOM framework, trained on fewer than 20 alpine sites with environmental conditions similar to Torgnon, outperformed the global FLUXCOM model in reproducing gross primary productivity (GPP). This demonstrates that regionally tuned models better capture underrepresented ecosystems, improving the accuracy of regional carbon budget estimates. Overall, this thesis shows that physically-informed machine learning approaches can provide reproducible, interpretable, and scalable estimates of vegetation structure and carbon dynamics. The methodologies developed bridge the gap between theoretical modeling and operational applications, enhancing our understanding of alpine ecosystem functioning and supporting informed environmental monitoring and management.