Seminar: Ranit De

Terrestrial ecosystems can act as a net sink of carbon, as it mediates exchange of carbon and water fluxes between land and atmosphere through photosynthesis. Ongoing measurements of carbon fluxes using the eddy covariance technique have shown large interannual variability (IAV) and its changing response with climate change. However, current models simulating gross primary productivity (GPP) fail to capture the IAV due to limitations in model structure and parametric uncertainty. In previous studies, we have shown these limitations can be partially alleviated with spatially and annually varying (both in terms of sites and site-years) model parameters, such as inverting values of maximum light use efficiency (ε_max) per year to better simulate annual GPP. Here, we wanted to explain and predict the spatial-temporal variation of ε_max mainly from annual statistics of meteorological variables, such as temperature, vapor pressure deficit (VPD), carbon dioxide concentration, a proxy of soil water content, etc. Moreover, we used features explaining site characteristics, such as vegetation type, climate classes, and high-resolution earth embeddings (which combine and distill high variety and volumes of remote sensing data). First, we tuned the hyperparameters, and trained three tree-based machine learning models, including Random Forest, XGBoost, and LightGBM. We also trained a foundation model for tabular data (TabPFN). Then we evaluated the performance of these models in predicting both spatial and temporal variability of ε_max in a nested cross-validation framework, as well as evaluated the predictive performance on temporal variability alone in a post-processing step. These models produced an average coefficient of determination (R²) of 0.6 in training sets, and an R² of 0.35 in validation sets. However, the predictive power of IAV or annual variability of ε_max remains poor (R² in the range of -0.23 to -0.14). Furthermore, we performed post-hoc analysis of these models to find important features explaining the spatio-temporal variability of ε_max, and found features related to annual statistics of temperature and VPD as important. Currently, we are working on expanding our methodologies to interpretable models, such as symbolic regression and generalized additive models, to better explain the spatio-temporal variability of ε_max, as well as discover physical relationships between meteorological features and annually varying ε_max.