Constraining Forest Carbon and Water Fluxes in the Italian Alps by Coupling National Forest Inventory Data, Process-Based Modeling, and Earth Observation

Mountain forests play a key role in the terrestrial carbon cycle, yet their contribution as carbon sinks remains highly uncertain, particularly in alpine regions. Steep elevation gradients, complex topography, and highly heterogeneous forest structure generate strong spatial variability in meteorological conditions and ecosystem processes, making high spatial resolution essential for both observation and modeling. While process-based forest models provide valuable insight into carbon and water fluxes, their application in mountain environments is often constrained by sparse observations and difficulties in scaling plot-level processes to the landscape. Integrating plot-scale modeling with high-resolution spatial information is therefore critical to better constrain model estimates in these systems. In our study, we developed a model–data integration framework for the Italian Alps (~52,000 km²), in which the 3D-CMCC-FEM process-based forest model was parametrized and run at National Forest Inventory (NFI) plot level. Plot-scale simulations of Gross Primary Production (GPP), Net Primary Production (NPP), and Evapotranspiration (ET) were then spatialized to continuous 30 m resolution maps using machine learning. The spatialization combined NFI-derived forest structural variables with high-resolution meteorological data, topographic predictors, and satellite-based vegetation indices. Four machine learning algorithms—Random Forest, Artificial Neural Networks, Extreme Gradient Boosting, and Support Vector Machines—were evaluated to extend plot-scale model outputs across the landscape. Model performance was assessed using k-fold cross-validation. Random Forest consistently achieved the highest predictive accuracy for all target variables, explaining approximately 27–47% of the variance in GPP, NPP, and ET across k-fold cross-validation and showing 3–15% lower prediction errors compared to the other machine learning methods. Variable importance analyses indicated that forest structural attributes derived from NFI data, elevation-related topographic metrics, and temperature- and precipitation-based meteorological predictors together accounted for the majority of the explained variance, emphasizing their dominant control on the spatial variability of forest carbon and water fluxes in alpine terrain. The resulting maps show clear spatial patterns in productivity and water use across alpine forest types and elevational gradients, providing spatially continuous, wall-to-wall information that complements plot-based National Forest Inventories. By linking plot-scale forest processes to landscape-scale patterns, this approach supports improved estimation, spatial consistency, and upscaling of forest carbon fluxes and stocks for measurement, reporting, and verification activities in heterogeneous mountain landscapes under ongoing climate change.