The gross primary production (GPP) of the terrestrial biosphere is a key source of variability in the global carbon cycle. It is modulated by hydrometeorological drivers (i.e. short-wave radiation, air temperature, vapour pressure deficit and soil moisture) and the vegetation state (i.e. canopy greenness, leaf area index) at instantaneous to interannual timescales. In this study, we set out to evaluate the ability of GPP models to capture this variability. Eleven models were considered, which rely purely on remote sensing data (RS-driven), meteorological data (meteo-driven, e.g. dynamic global vegetation models; DGVMs) or a combination of both (hybrid, e.g. light-use efficiency, LUE, models). They were evaluated using in situ observations at 61 eddy covariance sites, covering a broad range of herbaceous and forest biomes. The results illustrated how the determinant of temporal variability shifts from meteorological variables at sub-seasonal timescales to biophysical variables at seasonal and interannual timescales. RS-driven models lacked the sensitivity to the dominant drivers at short timescales (i.e. short-wave radiation and vapour pressure deficit) and failed to capture the decoupling of photosynthesis and canopy greenness (e.g. in evergreen forests). Conversely, meteo-driven models accurately captured the variability across timescales, despite the challenges in the prognostic simulation of the vegetation state. The largest errors were found in water-limited sites, where the accuracy of the soil moisture dynamics determines the quality of the GPP estimates. In arid herbaceous sites, canopy greenness and photosynthesis were more tightly coupled, resulting in improved results with RS-driven models. Hybrid models capitalized on the combination of RS observations and meteorological information. LUE models were among the most accurate models to monitor GPP across all biomes, despite their simple architecture. Overall, we conclude that the combination of meteorological drivers and remote sensing observations is required to yield an accurate reproduction of the spatio-temporal variability of GPP. To further advance the performance of DGVMs, improvements in the soil moisture dynamics and vegetation evolution are needed.

The processes involved in the exchange of water, energy and carbon in terrestrial ecosystems are strongly intertwined. To accurately represent the terrestrial biosphere in land surface models (LSMs), the intrinsic coupling between these processes is required. Soil moisture and leaf area index (LAI) are two key variables at the nexus of water, energy and vegetation. Here, we evaluated two prognostic LSMs (ISBA and ORCHIDEE) and a diagnostic model (based on the LSA SAF, Satellite Application Facility for Land Surface Analysis, algorithms) in their ability to simulate the latent heat flux (LE) and gross primary production (GPP) coherently and their interactions through LAI and soil moisture. The models were validated using in situ eddy covariance observations, soil moisture measurements and remote-sensing-based LAI. It was found that the diagnostic model performed consistently well, regardless of land cover, whereas important shortcomings of the prognostic models were revealed for herbaceous and dry sites. Despite their different architecture and parametrization, ISBA and ORCHIDEE shared some key weaknesses. In both models, LE and GPP were found to be oversensitive to drought stress. Though the simulated soil water dynamics could be improved, this was not the main cause of errors in the surface fluxes. Instead, these errors were strongly correlated to errors in LAI. The simulated phenological cycle in ISBA and ORCHIDEE was delayed compared to observations and failed to capture the observed seasonal variability. The feedback mechanism between GPP and LAI (i.e. the biomass allocation scheme) was identified as a key element to improve the intricate coupling between energy, water and vegetation in LSMs.