A long-standing challenge in studying the global carbon cycle has been understanding the factors controlling inter–annual variation (IAV) of carbon fluxes related to vegetation photosynthesis and respiration, and improving their representations in existing biogeochemical models. Here, we compared an optimality-based mechanistic model and a semi-empirical light use efficiency model to understand how current models can be improved to simulate IAV of gross primary production (GPP). Both models simulated hourly GPP and were parameterized for (1) each site–year, (2) each site with an additional constraint on IAV (CostIAV), (3) each site, (4) each plant–functional type, and (5) globally. This was followed by forward runs using calibrated parameters, and model evaluations at different temporal scales across 198 eddy covariance sites. Both models performed better on hourly scale than annual scale for most sites. Specifically, the mechanistic model substantially improved when drought stress was explicitly included. Most of the variability in model performances was due to model types and parameterization strategies. The semi-empirical model produced statistically better hourly simulations than the mechanistic model, and site–year parameterization yielded better annual performance for both models. Annual model performance did not improve even when parameterized using CostIAV. Furthermore, both models underestimated the peaks of diurnal GPP in each site–year, suggesting that improving predictions of peaks could produce a comparatively better annual model performance. GPP of forests were better simulated than grassland or savanna sites by both models. Our findings reveal current model deficiencies in representing IAV of carbon fluxes and guide improvements in further model development.

Vegetation plays a fundamental role in modulating the exchange of water, energy, and carbon fluxes between the land and the atmosphere. These exchanges are modelled by Land Surface Models (LSMs), which are an essential part of numerical weather prediction and data assimilation. However, most current LSMs implemented specifically in weather forecasting systems use climatological vegetation indices, and land use/land cover datasets in these models are often outdated. In this study, we update land surface data in the ECMWF land surface modelling system ECLand using Earth observation-based time varying leaf area index and land use/land cover data, and evaluate the impact of vegetation dynamics on model performance. The performance of the simulated latent heat flux and soil moisture is then evaluated against global gridded observation-based datasets. Updating the vegetation information does not always yield better model performances because the model’s parameters are adapted to the previously employed land surface information. Therefore we recalibrate key soil and vegetation-related parameters at individual grid cells to adjust the model parameterizations to the new land surface information. This substantially improves model performance and demonstrates the benefits of updated vegetation information. Interestingly, we find that a regional parameter calibration outperforms a globally uniform adjustment of parameters, indicating that parameters should sufficiently reflect spatial variability in the land surface. Our results highlight that newly available Earth-observation products of vegetation dynamics and land cover changes can improve land surface model performances, which in turn can contribute to more accurate weather forecasts.

This study analyzed highly-correlated, feature-rich datasets from hyperspectral remote sensing data using multiple machine and statistical-learning methods. The effect of filter-based feature-selection methods on predictive performance was compared. Also, the effect of multiple expert-based and data-driven feature sets, derived from the reflectance data, was investigated. Defoliation of trees (%) was modeled as a function of reflectance, and variable importance was assessed using permutation-based feature importance. Overall support vector machine (SVM) outperformed others such as random forest (RF), extreme gradient boosting (XGBoost), lasso (L1) and ridge (L2) regression by at least three percentage points. The combination of certain feature sets showed small increases in predictive performance while no substantial differences between individual feature sets were observed. For some combinations of learners and feature sets, filter methods achieved better predictive performances than the unfiltered feature sets, while ensemble filters did not have a substantial impact on performance. Permutation-based feature importance estimated features around the red edge to be most important for the models. However, the presence of features in the near-infrared region (800 nm - 1000 nm) was essential to achieve the best performances. More training data and replication in similar benchmarking studies is needed for more generalizable conclusions. Filter methods have the potential to be helpful in high-dimensional situations and are able to improve the interpretation of feature effects in fitted models, which is an essential constraint in environmental modeling studies.