Constraining terrestrial biosphere models with forest inventory data

Background/Question/Methods

Uncertainty in the response of vegetation to climate change accounts for much of the uncertainty in predicting the future state of the Earth system.  Recent evidence suggests that current global terrestrial biosphere models (TBMs) overestimate ecosystem response to climate and CO₂ change.  This is likely due to the emphasis in TBMs on fast time-scale physiological processes (e.g., biochemical temperature sensitivities of photosynthesis and respiration), rather than slow time-scale processes (e.g., temperature acclimation, local adaptation, and species turnover).  Spatial patterns of forest biomass and productivity observed in geographically extensive forest inventories, such as the U.S. Forest Service’s Forest Inventory and Analysis (FIA) program, integrate biological mechanisms across a range of time scales and provide an important but largely neglected constraint on TBMs.  We compared patterns in forest biomass and wood production (hereafter, ‘growth’) in FIA data across the eastern U.S. to output from the NOAA Geophysical Fluid Dynamics Laboratory LM3V land model.  We also developed an optimization system to fit vegetation parameters in LM3V to forest inventory or other data.  Finally, we show that ignoring errors in climate and soil data at measurement sites can bias model-data comparisons, and we propose an unbiased optimization approach that accommodates such errors.

Results/Conclusions

Biomass and growth both increased linearly with precipitation and soil water holding capacity in both FIA data and LM3V simulations, but the slopes of these relationships were steeper for LM3V.  This could indicate that LM3V vegetation is too sensitive to water.  However, errors in climate and soil maps (derived from one-degree-scale LM3V forcing data) at FIA plots biases the FIA regression slopes towards zero and provides an alternative explanation for the model-data mismatch.  A standard optimization approach (minimization of sum of squared deviations between growth in LM3V and FIA data) significantly improved the model-data fit, but yielded unrealistic estimates for vegetation parameter values (e.g., root:leaf allocation).  We used data from individual meteorological towers to quantify point-scale errors in the LM3V meteorology.  Accounting for these errors in FIA-data regressions yielded precipitation responses that were not significantly different from the LM3V precipitation responses.  Incorporating such errors into standard optimization approaches would greatly increase the computational expense of the optimization algorithm.  As an alternative, we suggest optimizing TBM parameter values to maximize agreement between the functional response of the model (e.g., ecosystem response to precipitation) and the error-corrected functional response of data.