Seasonality of Overstory and Understory Fluxes in a Semi-Arid Oak Savanna: What can be Learned from Comparing Measured and Modeled Fluxes?

Semi-arid climates experience large seasonal and inter-annual variability in radiation and precipitation, creating natural conditions adequate to study how year-to-year changes affect atmosphere-biosphere fluxes. Especially, savanna ecosystems, that combine tree and below-canopy components, create a unique environment in which phenology dramatically changes between seasons. We used a 10-year flux database in order to define seasonal and interannual variability of climatic inputs and fluxes, and evaluate model capability to reproduce observed variability. This is based on the perception that model capability to construct the deviation, and not the average, is important in order to correctly predict ecosystem sensitivity to climate change. Our research site is a low density and low LAI (0.8) semi-arid savanna, located at Tonzi Ranch, Northern California. In this system, trees are active during the warm season (Mar - Oct), and grasses are active during the wet season (Dec - May). Measurements of carbon and water fluxes above and below the tree canopy using eddy covariance and supplementary measurements have been made since 2001. Fluxes were simulated using bio-meteorological process-oriented ecosystem models: BEPS and 3D-CAONAK. Models were partly capable of reproducing fluxes on daily scales (R2=0.66). We then compared model outputs for different ecosystem components and seasons, and found distinct seasons with high correlations while other seasons were purely represented. Comparison was much higher for ET than for GPP. The understory was better simulated than the overstory. CANOAK overestimated spring understory fluxes, probably due to the capability to directly calculated 3D radiative transfer. BEPS underestimated spring understory fluxes, following the pre-description of grass die-off. Both models underestimated peak spring overstory fluxes. During winter tree dormant, modeled fluxes were null, but occasional high fluxes of both ET and GPP were measured following precipitation events, likely produced by an adverse measurement effect. This analysis enabled to pinpoint specific areas where models break, and stress that model capability to reproduce fluxes vary among seasons and ecosystem components. The combined response was such, that comparison decreases when ecosystem fluxes were partitioned between overstory and understory fluxes. Model performance decreases with time scale; while performance was high for some seasons, models were less capable of reproducing the high variability in understory fluxes vs. the conservative overstory fluxes on annual scales. Discrepancies were not always a result of models' faults; comparison largely improved when measurements of overstory fluxes during precipitation events were excluded. Conclusions raised from this research enable to answer the critical question of the level and type of details needed in order to correctly predict ecosystem respond to environmental and climatic change.