Evaluation and improvement of the Community Land Model ( CLM 4 . 0 ) in Oregon forests

Introduction Conclusions References


Introduction
Modeling and understanding the response of terrestrial ecosystems to changing environmental conditions and land use change are primary goals of climate mitigation policy (IPCC, 2007;Moss et al., 2010;NRC, 2010a;Pacala, 2010).The Intergovernmental Panel for Climate Change (IPCC) synthesizes estimates of future climate change impacts on terrestrial carbon cycling through the use of a specific set of global circulation models (IPCC, 2007).Among them is the Community Earth System Model (CESM) of the National Center for Atmospheric Research (NCAR).The land model component (Community Land Model; CLM) has been extensively used to evaluate and predict the net carbon uptake and loss from terrestrial biomes, particularly forests (Thornton and Zimmermann, 2007;Bonan et al., 2012).
Recent assessments of seasonal performance of multiple terrestrial biosphere models indicated better performance in forest ecosystems, particularly in evergreen forests, during the summer growing season than in other biomes and seasons (Schwalm et al., 2010).The highest skill across biomes was in models that prescribed canopy phenology and did not use a daily time step (Schwalm et al., 2010).In an assessment with FLUXNET tower data, Schaefer et al. (2012) found that none of the models estimated gross photosynthesis (GPP) at all sites within observed uncertainty.The most important factor influencing model performance was light-use efficiency computed from the slope of the GPP -light response curve.This is influenced by photosynthetic parameters, indicating better parameter values are needed for variables influencing LUE.The model evaluations also suggested more detailed assessments need to be performed with individual models.The photosynthesis model of (Farquhar et al., 1980) is widely used in vegetation models, and an evaluation at more than 200 FLUXNET sites showed that simple PFT classification of photosynthesis parameters introduces Introduction

Conclusions References
Tables Figures

Back Close
Full uncertainty in photosynthesis and transpiration fluxes (Groenendijk et al., 2012b), and including V cmax seasonality can improve predictions (Groenendijk et al., 2012a).The study suggested focusing on the effects of seasonal foliar nitrogen on V cmax .In an assessment of CLM4 with FLUXNET data, Bonan et al. (2011) also found that the bias in annual GPP could be reduced by including improved estimates of photosynthetic parameters.Finally, CLM4 was also found to overestimate shade-leaf photosynthesis leading to overestimation of canopy GPP when the nitrogen limitation functionality was inactive (Bonan et al., 2012) suggesting a multi-layer canopy could improve initial GPP calculations before downregulation due to nitrogen limitation is imposed.
In this paper, we evaluate the Community Land Model (version 4.0) portion of the Community Earth System Model (CESM1.0.4).CLM4 is the latest in a series of land models developed for the CESM and runs at a half-hourly timestep.CLM4 includes coupled carbon and nitrogen processes and examines the physical, chemical, and biological processes through which terrestrial ecosystems affect and are affected by climate across a variety of spatial and temporal scales.Recent model releases include improvements to hydrology and an integrated transient land cover and land use change dataset (Lawrence et al., 2011a).The transient land cover dataset includes historical wood harvest which is known to have a large influence on Pacific Northwest forest carbon storage and uptake (Harmon et al., 1990;Law et al., 2004).When the carbon-nitrogen biogeochemistry is active, CLM4 uses prognostic canopy phenology to determine leaf and stem area index (LAI and SAI) and vegetation heights.Potential GPP is calculated from leaf photosynthetic rate without nitrogen constraint.Actual GPP is computed from nitrogen limitations to potential GPP.
Our objective was to evaluate carbon stocks and fluxes simulated by CLM4 in a representative sample of Pacific Northwest forests using independent datasets with different spatial and temporal resolutions, rather than relying only on a few eddy-covariance tower sites as is typical.We utilized annual carbon stocks and net primary production (NPP) computed from over 3000 forest inventory plots that had been scaled to the study region with remote sensing data to produce a spatial dataset (Hudiburg et al., Introduction Conclusions References Tables Figures

Back Close
Full 2011; Law et al., 2012), and the seasonal data were gross photosynthesis (GPP) from two FLUXNET sites in the region (Krishnan et al., 2009;Thomas et al., 2009).Model evaluation with an observation-based spatially representative dataset such as federal forest inventory data (FIA) is critical in order to constrain model estimates, facilitate model development and ultimately reduce model uncertainty.Reduced model uncertainly is beneficial for forest policy development of land use practices aimed at increasing carbon sequestration.Combined with seasonal data and analysis over a strong climatic gradient, diagnostics can point to areas for model improvements.The specific objectives of this study are to (1) evaluate regional model performance against spatially representative FIA estimates of stem wood biomass and annual NPP; (2) evaluate seasonal model performance against monthly tower GPP utilizing observed physiological parameters; and (3) examine the roles of modeled LUE, temperature, precipitation, and nitrogen limitation in determining seasonal patterns of GPP.Temperature is a dominant control on the seasonal variation in GPP (Schaefer et al., 2012) whereas nitrogen and water limitation are major determinants of seasonal maximum GPP (Lawrence et al., 2011b;Bonan et al., 2011).
The accuracy and uncertainty associated with regional estimates of carbon fluxes by CLM4 is unknown because model output is not usually validated against spatially representative observations.The federal forest inventory (FIA) collects data on an annual basis on all forestland regardless of ownership or location resulting in an intensively sampled landscape through which modeled estimates of forest carbon stocks and fluxes could be evaluated.FIA data provide measurements that can be used to calculate the carbon density of live and dead trees, woody detritus, and understory shrubs.Wood increment data are also provided, allowing estimation of 5-10 yr average bole wood production, depending on the repeat measurement cycle.Combined with remote sensing land cover products, reliable maps of wood carbon stocks, net primary production (NPP), heterotrophic respiration (R h ), and net ecosystem production (NEP) can be produced (Hudiburg et al., 2011)  providing a range of values or baseline conditions modeling activities should be constrained by before making predictions about future conditions.
Here, we evaluate CLM4 with FIA data maps of wood carbon stocks and NPP averaged for the years 2001-2006 in Oregon forests.We improve on model estimates by calibrating with supplemental plot data collected at over 100 sites across the study area.The region is characterized by tremendous climate variation and it has been recommended that a minimum of 15 km resolution be utilized in such heterogeneous terrain (Salath é et al., 2010).We utilize a new downscaled 200 yr 1/8-th • (15 km) resolution dataset specifically designed for use in the Pacific Northwest (Salathe et al., 2007).Furthermore, we also test the model's ability to determine the seasonal timing and magnitude of GPP by comparing modeled estimates with eddy-covariance data in the region, and evaluate tower annual NPP and GPP for consistency (Luyssaert et al., 2009).Finally, we make modifications to model algorithms and plant functional type (PFT) physiological parameterization to improve model performance.By using existing datasets and uncertainty quantification, this research provides a solid scientific foundation for evaluating carbon dynamics under conditions of future climate change and land management practices at local and regional scales.

Model description
The model used to simulate these processes is the Community Land Model (CLM4) portion of the Community Earth System Model (CESM 1.0.4) of the National Center for Atmospheric Research (Oleson et al., 2010).CLM4 uses hourly climate data, ecophysiological characteristics, site physical characteristics, and site history to estimate the daily fluxes of carbon, nitrogen, and water between the atmosphere, plant state variables, and litter and soil state variables.State variables are the live and dead carbon pools.CLM4 examines the physical, chemical, and biological processes through Introduction

Conclusions References
Tables Figures

Back Close
Full which terrestrial ecosystems affect and are affected by climate across a variety of spatial and temporal scales.The basic assumption is that terrestrial ecosystems, through their cycling of energy, water, chemical elements, and trace gases, are important determinants of climate.Model components consist of biogeophysics, hydrologic cycle, biogeochemistry and dynamic vegetation.The land surface is divided into five subgrid land cover types (glacier, lake, wetland, urban, vegetated) in each grid cell.The vegetated portion of a grid cell is further divided into patches of plant functional types (PFTs), each with its own leaf and stem area index and canopy height.The multiple PFTs within the grid cell compete for water and nutrients on a single soil column.
Recent model improvements include updates to hydrology and an integrated transient land cover and land use change dataset that accounts for wood harvest.Updates to the hydrology include improved ground evaporation parameterization and snowpack heating and aging resulting in higher snow cover, cooler soil temperatures in organic-rich soils, greater river discharge, and lower albedos over forests and grasslands (Lawrence et al., 2011a).The transient land use dataset has been formatted for use by CLM4 from a global historical transient land use and land cover change dataset (Hurtt et al., 2006) covering the period from 1850-2005.The dataset describes land cover and its change via four classes of vegetation (crop, pasture, primary vegetation, and secondary vegetation).The dataset also describes the annual fraction of land that is transformed from one PFT to another including the conversion of primary forest to secondary forest which is essential for studies attempting to track forest carbon storage and uptake over time.This dataset essentially provides a human-induced disturbance history of forests which can be used to test the model's ability to simulate carbon dynamics following harvest.

Model calibration and forcing datasets
CLM4 has over 40 physiological parameters for each of the 17 plant functional types and a default constant mortality rate of 2 % for all PFTs.The PFTs include 8 different forested and 3 different shrub land PFTs (Appendix A; Tables A1 and A2).We use Introduction

Conclusions References
Tables Figures

Back Close
Full data collected on 100 supplemental plots located throughout the study region (Fig. 1) to calibrate the physiology file, mortality rate, and wood allocation algorithms.Default physiology values were used for parameters where data was unavailable and for parameters which are known to have little influence on biomass and NPP.Assessment and sensitivity analysis of CLM4 to the parameter values has been tested and described (White et al., 2000;Lawrence et al., 2011a) and we incorporate this knowledge to facilitate our regional calibration.The supplemental plot dataset includes measurements of specific leaf area (SLA; projected), foliar carbon nitrogen ratios, litter carbon nitrogen ratios, and leaf longevity for the major tree species across the study region.
We use the PFT mean for each of these parameters in the calibrated physiology file for default model configuration evaluation, heretofore indicated as "CLM4" in figures and text.Prior modeling studies in this region indicate that a dynamic mortality rate that varies with age and/or disturbance type is necessary to predict the correct seasonal and annual carbon fluxes (Turner et al., 2007), especially in the drier forest types where mortality decreases as a percentage of live biomass as stands age.Mortality rates in Oregon forests range from 0.5 to 2 % (Turner et al., 2007;Hudiburg et al., 2009) in the absence of stand replacing disturbance.Since dynamic mortality algorithms have not been incorporated into CLM4, we chose to use 1.0 % as a static mean value for the default configuration simulations.To improve on this, we parameterized CLM4 with PFT mortality rates within each ecoregion based on inventory data (Hudiburg et al., 2009).
Ecoregions are areas within which ecosystems (and the type, quality, and quantity of environmental resources) are generally similar (Omernik, 1987).
Offline simulations of CLM4 are typically forced with the NCEP reanalysis dataset (Qian et al., 2006) provided by NCAR.While this dataset includes the required climate and forcing variables at a sub-daily timestep, the resolution (∼ 2 degree) is not adequate for regional simulations in Oregon.For this reason, we forced the model with a 1/8-th degree regional dataset assembled by the Climate Impacts Group at the University of Washington (http://www.cses.washington.edu/data/ipccar4/)downscaled from the ECHAM5 20th century climate simulations used in the IPCC 4th assessment.Introduction

Conclusions References
Tables Figures

Back Close
Full The regional dataset includes daily precipitation, minimum and maximum temperature, and wind speed.Because CLM4 also requires shortwave radiation and relative humidity we calculated values incorporating algorithms from DAYMET (Thornton et al., 1997) and methods for sub-daily calculations as described by (G öeckede et al., 2010; Appendix A).

Model simulations and evaluation
All model simulations are summarized in Table 1.Model spinups required 650 yr to reach equilibrium conditions and were performed using regional physiology parameters and the downscaled climate forcing dataset (PNW).The fire module was turned "off" for spinup because in some ecoregions, biomass would "burn up" before reach- file is a single value for the entire region which changes annually.For the FLUXNET tower sites, we used the observed meteorological data for all tower site simulations including spinup.Gap-filled hourly data included air temperature, precipitation, shortwave solar radiation, wind speed, and specific humidity.

Observations
We combined spatially representative observational data from 3125 FIA plots measured during 2001 to 2006 with remote sensing products on forest cover, ecoregion, and age and a global data compilation of wood decomposition data (Wirth et al., 2010) to provide current maps of the Pacific Northwest state of Oregon (∼ 12 million hectares) forest biomass carbon stocks, heterotrophic respiration (R h ), net primary production (NPP), and net ecosystem production (NEP).We included all forestland in our analysis across all age classes (20-800 yr old) and management regimes.Plot values were aggregated by ecoregion, age class, and forest type and this look-up table was used to assign a value to each associated 30 m pixel.Methods and uncertainty of the inventory biomass and carbon flux calculations are fully described in (Hudiburg et al., 2009(Hudiburg et al., , 2011)).Tree biomass is calculated using ecoregion-and species-specific allometric equations from diameter at breast height (DBH) and height.Woody detritus length and diameter are recorded along transects for each plot in the FIA dataset.Woody detritus biomass is then calculated from piece volume and decay class density.Wood NPP is calculated using increment data from wood cores collected on the FIA plots, foliage NPP is foliage biomass divided by the average leaf retention time measured at the intensive plots, and fine root NPP is calculated from fine root biomass multiplied by the average root turnover rates from intensive plot data.
Monthly mean values of GPP from eddy-covariance tower data and the associ- 2003; Law et al., 2004).Seasonal patterns of modeled GPP are compared with the flux tower observations.

Uncertainty analysis
We calculated the total uncertainty in the current FIA estimates using the propagation of error approach (Williams et al., 2012).The propagation of error approach is a method for combing error terms to calculate total uncertainty in an estimate.Monte Carlo simulations were used to estimate the uncertainty due to variation in region-and/or speciesspecific allometry by using the mean and standard deviations for stem biomass and NPP calculated using three alternative sets of allometric equations.The full suite of species-specific equations that use tree diameter (DBH) and height (preferred) were compared to a DBH-only national set (Jenkins et al., 2003) and to a grouped forest type set.Total uncertainty in FIA estimates was calculated as the combined uncertainty of the allometric equations and land cover estimates (10 %) using the propagation of error approach (NRC).Ecoregion level uncertainty estimates are represented in the figures by the grey error bars and in tables with "±" symbols.For finer scale evaluation (i.e.grid cell size), the uncertainty in the observations was used to evaluate model performance using the reduced chi-square (x 2 ) statistic (Schwalm et al., 2010) and model bias defined as the mean of the model-data residuals for different groups of interest.Inventory plot data were grouped by ecoregion and then aggregated by precipitation regime, biome type, and age group within each ecoregion.Plots with annual precipitation greater than 800 mm were considered "Wet" and less than 800 mm as "Dry".FIA forest types are either evergreen needleleaf (ENF), deciduous broadleaf (DBF), or mixed ENV/DBF (MXD).Complete analysis and statistics were performed for all biomes however since 92 % of the forested area in the study region is ENF, we include only the ENF biome in the figures and tables of this document.Age groups are defined as young (< 60 yr) and mature (> 60 yr).
The reduced x 2 is the squared sum of the residuals normalized by the observation uncertainty and divided by the total number of samples (i.e. the mean x 2 for an arbitrary Introduction

Conclusions References
Tables Figures

Back Close
Full group of interest): where, δ is the uncertainty in the i -th observation, "2" normalizes the uncertainty to correspond to a 95 % confidence interval, and the summation is across the aggregated data groups within each ecoregion (Schwalm et al., 2010).Reduced x 2 values close to 1 indicate model-data unity or agreement.Model bias can be estimated as the mean of the residuals: where i is the group of interest and positive values indicate average overestimation by the model and negative values indicate average underestimation.

Model development
We aimed to reduce overall uncertainly by calibrating with known regional mean values and evaluating with inventory data.This was in part due to an interest in evaluating the default model configuration with the fewest adjustments possible and to utilize the rich regional plot dataset.Testing the performance in a diverse region aids diagnosis of parameter or structural deficiencies.As stated above, the default configuration includes PFT specific physiological variables (no seasonal variation), a constant mortality rate for all forest PFTs, and a single stem wood allocation equation for all woody PFTs.After evaluation with the default configuration, we tested the use of modified model versions where the physiological parameters and mortality rates were allowed to vary by PFT within ecoregion (CLM4 eco) and where the stem wood allocation equation in addition to the physiological and mortality rates also varied by ecoregion (CLM4 stem).Introduction

Conclusions References
Tables Figures

Back Close
Full For CLM4 eco and CLM4 stem, parameter values for foliar CN ratios, foliar N content in Rubisco, leaf longevity, fine root CN ratios, and specific leaf area (SLA) were adjusted according to field plot data from sites in each of the tested ecoregions in the study area.The PFT physiological variable input file was restructured so that two PFTs were assigned to each ecoregion and surface datasets were modified to reflect the new PFT assignments (Appendix A, Tables A1 and A2).The CLM4 mortality module was then modified to assign a different mortality rate based on the PFT.For CLM4 stem, the inventory data was used to construct ecoregion-specific allocation to stem wood equations and the equations were added to the CLM4 allocation module (Appendix A, Table A4 and Fig. A1).
Finally, CLM4 nitrogen fixation is controlled by a single equation that determines the rate as a function of NPP yielding nitrogen fixation rates up to 1.8 g N m −2 yr −1 , whereas nitrogen fixation at the Campbell river site is reportedly 3 g N m et al., 2006).Because this is not a parameter in the physiology file, we needed to modify the equation in the source code for the Campbell River spinup and simulations (Appendix A).

Modeled results and regional totals
In general, CLM4 stem modeled estimates of stem biomass carbon (Fig. 2a) followed a west to east gradient with higher values in the productive mesic western ecoregions in the Coast Range (Fig. 2d).Total simulated regional tree carbon stocks are estimated to be 1.2 Pg C with 61 % in the Coast Range and West Cascades alone.For the period from 2001-2006, total forest NPP and NEP averaged 57.6 and 12.7 Tg C yr −1 respectively indicating a strong sink in the region.Total modeled harvested carbon equaled 4.9 Tg C yr −1 and total fire emissions equaled 2.4 Tg C yr −1 .After accounting for these removals, the forest carbon sink (net biome production, NBP) is still positive at 5.4 Tg C yr −1 .

Model improvement and testing
We compared regional plot data of foliar chemistry and morphological characteristics with the physiological file constants used for the modeled runs (Fig. 3).There is significant variation in both foliar carbon to nitrogen ratios and specific leaf area (SLA, projected) between ecoregions according to plot data.The default configuration for CLM4 allows one foliar CN ratio and one SLA value per PFT.We experimented with physiological parameter spatial variation (Fig. 4; CLM4 eco), and found that for the Blue Mountains, overall mean biomass decreased to 71 Mg C ha −1 and mean NPP decreased to 307 g C m −2 yr −1 , both falling within the observed range of uncertainty.The stem carbon reduced x 2 statistic decreased from 15 to 2, and the NPP x 2 decreased from 5 to 2. Stem carbon bias was reduced from 48 to 9 Mg C ha −1 and NPP bias decreased from 112 to a slight underestimation of −3 g C m −2 yr −1 .We found similar improved relationships in the other ecoregions for NPP with CLM4 eco, but stem carbon stocks were subsequently underestimated.CLM4 includes dynamic stem wood allocation, but it is a single equation for all woody PFTs globally.Finally, using CLM4 stem where we also changed the dynamic stem allocation equation to vary by ecoregion, we found an overall improvement in stem carbon stocks compared to CLM4.Introduction

Conclusions References
Tables Figures

Back Close
Full

Spatial evaluation with inventory data
We show ecoregion means of the evergreen needleleaf (ENF) biome because it represents 92 % of the forested area in the region and we do not have sufficient regional plot data for the deciduous broadleaf (DBF) biome for evaluation.CLM4 regional modeled mean stem biomass for the period from 2001-2006 fell within the observed range of uncertainty in the Coast Range and the West Cascades, was close to observed range of uncertainty in the East Cascades and was underestimated in the Klamath Mountains (Fig. 4a; grey error bars indicate observation uncertainty).However, mean simulated biomass was twice the observed mean in the Blue Mountains.Simulated regional mean NPP was within the observed range of uncertainty in the Coast Range, East Cascades and Klamath Mountains, was slightly overestimated in the West Cascades and was again almost twice the observed mean in the Blue Mountains.After modifying the model for ecoregion differences in input physiological parameters (CLM4 eco), ecoregion mean stem biomass was within the observed uncertainty in the Blue Mountains, but dropped below the average in the West Cascades.Ecoregion mean NPP was within observation uncertainty in all ecoregions after modification.After modifications to the stem allocation equations, modeled ecoregion means of NPP were not significantly different from observed means (two-sided p-value 0.8, respectively) and modeled means of stem biomass fell within the observed range of uncertainty in the majority of the ecoregions.
Ecoregion means are a good first order approximation of model performance and regional total evaluation, but cannot be used to determine the dynamics that may be causing bias and reveal nothing about model fidelity at a smaller scale.Statistical tests of model performance for stem carbon biomass revealed overall better agreement with inventory data using the CLM4 stem version of the model.At 15 km spatial resolution, reduced x 2 statistics indicated adequate model performance for the region as a whole for both stem biomass and NPP (overall reduced x 2 was 3 and 2, respectively).This was an improvement over the default configuration by 50 % for stem biomass and 30 % Introduction

Conclusions References
Tables Figures

Back Close
Full for NPP.Agreement was best in the West Cascades, Blue Mountains, and Coast Range (x 2 ≤ 2; Table 3) followed by the Klamath Mountains and East Cascades (x 2 < 7).The largest improvements were in the Blue Mountains where x 2 was reduced from 15 to 2. Initial CLM4 stem carbon biomass was underestimated by 11 to 41 Mg C ha −1 in all ecoregions except the Blue Mountains.After modification, stem carbon biomass was consistently underestimated in all ecoregions by an average of 18 Mg C ha −1 .Within ecoregions, performance was generally better in older age classes (Table 3).NPP also improved when using CLM4 stem, primarily because overestimation was reduced in the Blue Mountains and West Cascades (Table 3).Overestimation in the Blue Mountains exceeded 100 g C m −2 yr −1 with CLM4 before modifications were made.However, reduced x 2 statistics for the Coast Range, East Cascades and Klamath Mountains did not significantly improve.NPP was underestimated in the Coast Range and East Cascades by CLM4, improved for the Coast Range with CLM4 stem, but was then overestimated for the East Cascades.As with stem biomass, there was better performance in wet versus dry stands, except in the Klamath Mountains.Performance was better for both metrics in mature stands.This was especially true for the Coast Range and West Cascades where the x 2 values for CLM4 and CLM4 stem were close to 1 for the mature age classes, indicating model-data unity.

Evaluation with tower and supplemental plot data
Simulated monthly values for GPP were compared with 6 yr of eddy-covariance tower data at two sites in the region.At the semi-arid mature ponderosa pine site (Metolius), a reduced x 2 value of 1.06 indicated good overall model performance (values close to 1 indicate good model performance) after accounting for observation uncertainty (Schaefer et al., 2012).Observation uncertainty ranged from 12-100 % for the Metolius site and 13-90 % for the Campbell River site.At the semi-arid Metolius site, maximum Introduction

Conclusions References
Tables Figures

Back Close
Full mature Douglas-fir site (Campbell River), a reduced x 2 of 1.5 indicated weaker model performance than for the Metolius site, although estimates were still within the observed range of uncertainty for most months and annually (Fig. 5c and 5d).In contrast to the Metolius site, model data mismatch occurred primarily in spring with timing of peak GPP occurring one month early and overestimated spring values.This resulted in annual estimates that were on average 17 % greater than observed GPP.CLM4 includes direct down-regulation of the photosynthetic rate under nitrogen limitation, which effectively reduces GPP.We also compared GPP before down-regulation (Fig. 5a, b; "initial GPP") with tower data and found decreased model performance at both sites.Initial GPP was overestimated by up to 200 g C m −2 month −1 at the Metolius site, especially in the late summer months.Results were similar at the Campbell River site where simulated summer maximum initial GPP was much higher than observed (Fig. 5c).We compared modeled and observed monthly light use efficiency (LUE) with monthly average air temperature at each tower site (Fig. 6).We define LUE as the monthly GPP as a function of monthly average incident solar radiation (shortwave radiation).At the Metolius site, observed LUE increased with air temperature to about 10 • C, then declines (Fig. 6a).In other words, higher GPP is attained at lower radiation levels at temperatures up to 10 • C.This pattern was also true for the modeled relationship, but occurred less rapidly (Fig. 6b).Simulated LUE is lower than observed LUE for lower radiation and temperature conditions, but similar to observed at higher light and temperature conditions.At the Campbell River site, there is a smaller increase in LUE with temperature than the Metolius site (Fig. 6c), and there is much less seasonal variation in temperature.There appears to be a slight decrease in LUE with higher temperature in the modeled results (Fig. 6d) whereas there is a consistent increase with temperature in the observations.Simulated LUE is higher than observed for lower radiation and temperature conditions.Annual biometric estimates of NPP were available from the Metolius site and are often used as cross-checks with tower data for consistency (Luyssaert et al., 2009).Introduction

Conclusions References
Tables Figures

Back Close
Full Modeled annual NPP was underestimated by an average of 17 % when compared to biometric estimates of NPP at the Metolius site while modeled GPP was underestimated by an average of 18 % compared to the eddy-covariance data.However, the ratio of NPP/GPP was nearly identical for both equaling 34 % (observed) and 33 % (modeled).
We also compared observed and simulated monthly GPP with growing season monthly precipitation (Fig. 7).For both sites, winter precipitation is much higher and summers are generally dry.At the Metolius site, 2002-2004 were particularly dry while 2005-2007 received average precipitation.There does not appear to be a relationship with precipitation at the mesic Campbell River site, whereas at the semi-arid Metolius site, observed GPP increases for months with less than 30 mm of precipitation.Simulated GPP is similar to observations for the higher precipitation months for both Metolius and Campbell River, but underestimated in the low precipitation months at the Metolius site.

Modeled results and regional totals
The regional patterns of carbon stocks, NPP, and NEP are similar to the observed values in this study and prior studies.Stem carbon stocks from over 8000 inventory plots and 200 plots in Oregon showed stem carbon stocks range from 60-420 Mg C ha age classes typically show a decrease in tree wood biomass or it reaches an asymptote.However, observations in the Oregon suggested otherwise (Hudiburg et al., 2009) and modeled allocation to wood biomass in these regions was modified in CLM4 to reflect these regional stand dynamics (CLM4 stem).
Modeled NPP ranges from 100-1100 g C m −2 yr −1 (CLM4 stem) across the study region agreeing with the reported regional values of 100-900 g C m −2 yr −1 (Hudiburg et al., 2009) and falling within the range of 100 to 1600 g C m −2 yr −1 reported for temperate and boreal forests (Luyssaert et al., 2008).The regional total NPP for forests is 57.6 Tg C yr −1 , a value very close to the observed total of 58.2 ± 6.5 Tg C yr −1 calculated from inventory data (Law et al., 2012).

Spatial evaluation
After modifications were made, stem carbon stocks continued to be underestimated in all ecoregions.Performance was generally better in older age classes, and underestimation of stem biomass in old age classes was balanced by overestimation of stem biomass in young age classes in some ecoregions.Stem wood allocation is a dynamic function of NPP rather than age in CLM4.In a modeling assessment of US forest NEP following disturbance, the ecosystem model (CASA) was altered so that live biomass estimates were constrained by age-accumulation trajectories calculated from FIA data (Williams et al., 2012).An approach similar to this could be used in CLM4 to remove the age-related bias in stem biomass and would improve on the ecoregion-specific allocation we implemented in this study.In the Pacific Northwest, the age-related biomass accumulation trajectories are known to vary by ecoregion and management regime (Van Tuyl et al., 2005;Hudiburg et al., 2009) with higher potential stem biomass in the more productive ecoregions.Introduction

Conclusions References
Tables Figures

Back Close
Full Implementation of ecoregion physiological variation and mortality had a significant impact on NPP performance, with better performance in mature forests for CLM4 eco and CLM4 stem.This is not surprising considering NPP tends to peak early in stand development, decline and then remain at a consistent level as stands age.The agerelated variation in this response varies by ecoregion and disturbance regime (Law et al., 2004) and improvement to age-related allocation may help to correct part of the discrepancies.When using single-value parameterization of foliar CN ratios and SLW, we found underestimation of NPP by CLM4 in the Coast Range and overestimation in the Blue Mountains (Fig. 3), which are much higher in mesic versus semi-arid forests (Matson et al., 1994).Maximum rates of carboxylation (V cmax ) are determined by SLA and foliar nitrogen content in CLM4.Prior modeling efforts with CLM4 also found significant improvement in GPP estimates by using site-specific parameterization of V cmax (Bonan et al., 2011).
After implementing CLM4 eco and CLM4 stem, NPP continued to be overestimated while stem biomass continued to be underestimated in the West Cascades, although both were improved.Foliar nitrogen content in the Cascade Mountain ecoregion is lower than that of the Coast Range and Klamath Mountain ecoregions and has higher specific leaf area.The positive bias in NPP was reduced with ecoregion-specific parameterization, but this also reduced stem biomass.However, observed West Cascade stem biomass is equal to or higher than Coast Range mean biomass.Stem biomass density is affected by removal rates through fire and harvest.The CLM4 input harvest rates are higher than observed values in the West Cascades (1.9 Tg C yr −1 vs. Fortunately, these values can be easily adjusted to the known rates for future simulations.Total regional modeled fire emissions are also overestimated by 1.0 Tg C yr −1 which could also be contributing to the underestimation bias.Finally, nitrogen availability has recently been shown to significantly impact the NPP to GPP ratio, with higher rates of biomass production on sites with higher fertility (Vicca et al., 2012).Our results support this hypothesis, with higher biomass production in the Introduction

Conclusions References
Tables Figures

Back Close
Full forests with high foliar nitrogen content (Figs. 2 and 3).Thus, the model represents nitrogen influences on allocation and productivity quite well.

Seasonal evaluation
Simulating GPP correctly is important because most of the subsequently calculated carbon fluxes are calculated from GPP.In a model-data synthesis activity including 39 flux sites and 26 models (CLM4 was not included), GPP was generally found to be overestimated in the spring and fall and underestimated in the summer resulting in poor model performance at most flux sites (Schaefer et al., 2012).We found CLM4 simulated GPP at the Metolius site to be close to the observed range of uncertainty (reduced x 2 = 1.06) and GPP was not over-predicted in fall and spring resulting in overall good model performance.However, GPP was under-predicted in the summer months for some of the modeled years (Fig. 5a).LUE was determined as the primary driver of underestimation of GPP in summer months in the model-data synthesis.Our analysis agrees with this as we found LUE to be on average less efficient in the modeled results than the observations, primarily in lower light and temperature conditions.
However, this does not necessarily mean the low GPP bias in the model is caused by incorrect representation of LUE.We also found a different range of temperature thresholds for the range of LUE values indicating PFT-specific temperature thresholds may vary, especially in the drier more extreme climates found in the eastern portion of our study region.GPP was also found to be underestimated for months with little to no precipitation (i.e.summer; Fig. 7a) indicating soil water availability or plant water use efficiency need further investigation.When we did not include nitrogen in the simulations, the "initial GPP" before down-regulation occurred as a result of nitrogen limitation, leading to a positive summer bias.When nitrogen was included in the simulations, down-regulation brought the majority of the simulated GPP monthly values closer to observed values, but there was too much reduction in GPP in the summer months due to nitrogen constraints.

BGD Introduction Conclusions References
Tables Figures

Back Close
Full At the Campbell River site, modeled GPP was overestimated in spring and maximum GPP was reached one month early, resulting in overall overestimation of annual GPP.Foliar nitrogen content and availability are both higher at the Campbell River site than the Metolius site as well as precipitation.Downregulation of GPP due to nitrogen limitation does not result in underestimated summer GPP, however spring initial GPP and downregulated GPP are both too high and peak too early, indicating the V cmax response to temperature and daylength may need revising.

Summary of results and recommendations for model improvement
CLM4 was evaluated against inventory data in Oregon forests using the default configuration and regional parameterization.With default parameterization, initial results indicated overall underestimation of stem biomass except in the semi-arid Blue Mountains where it was overestimated by 48 Mg C m −2 .There was good general agreement with observed NPP values.However, modeled ecoregion mean NPP was overestimated in the Blue Mountains and mesic West Cascades.Following initial default parameterization, model improvements were made to account for ecoregion differences in the physiological variables, foliar N content and mortality, and this resulted in an overall improvement in NPP estimates (all ecoregions fell within the observed range of uncertainty).Changing the stem wood allocation algorithm further improved the results, however wood carbon stocks were still underestimated in the West Cascades and Klamath Mountains (CLM4 stem).Adjusting stem wood allocation changes with age should improve results, as well as adjusting known harvest rates and reducing the amount of wood combusted in wildfire to match historical rates.Using ecoregion-specific parameterization at 15 km spatial resolution, reduced x 2 statistics indicated adequate model performance in all ecoregions for both stem biomass and NPP (reduced x 2 was equal 3 and 2, respectively).This was an improvement from the default configuration by 50 % for stem biomass and 30 % for NPP.Within ecoregions, there was good performance (x 2 < 2) in the Coast Range and West

BGD Figures Back Close
Full Cascades.There was generally better performance in mature stands and apart from the Klamath Mountains, better performance in wet stands.Evaluation of CLM4 monthly GPP with eddy-covariance tower data revealed good model-data agreement from July to December at the mesic site (ca1; Fig. 5c) and September-May at the semi-arid site (me2; Fig. 5a).Summer GPP was underestimated at the semi-arid site due to several possible reasons including variation in seasonal V cmax response to temperature and soil water.We found a different range of temperature thresholds for the range of LUE values indicating PFT-specific temperature thresholds may vary, especially in the drier more extreme climates found in the eastern portion of our study region.At the mesic site, overestimation of spring GPP indicates that improvement to V cmax seasonality in CLM4 will require adjustment to other factors in addition to temperature such as daylength in order to improve summer GPP.
Appendix A A1 Methods for formatting regional downscaled climate forcing dataset for use by CLM4 The regional downscaled dataset includes daily precipitation, minimum and maximum temperature, and wind speed.To format the datasets for use by CLM4, the required shortwave radiation and relative humidity were calculated incorporating algorithms from DAYMET (Thornton et al., 1997), and methods for sub-daily calculations as described by (G öeckede et al., 2010).We created hourly atmospheric forcing data files to be used in offline CLM4 simulations.Source data files are from the climate impacts group (Salathe, 2008) and are downscaled historical and future ECHAM5 AR2 files specifically designed for the Pacific Northwest http://www.cses.washington.edu/data/ipccar4/.
The specific variable calculations are as follows:

BGD Figures Back Close
Full 1. Air temperature: The regional input files provide daily minimum and maximum temperature.Daylength was used to apply a diurnal pattern to the minimum and maximum temperatures.
2. Wind speed: The regional input files provide daily estimates of wind in m s −1 .Wind speed was assumed to be constant for sub-daily time steps.
3. Relative humidity: Percent relative humidity was calculated using the hourly mean temperatures calculated above and vapor pressure.

Precipitation:
The input files provide daily sums of precipitation.This needed to be distributed over the day, but not evenly.CLM will evaporate off the water too quickly and none of it will reach the plant roots.The precipitation was split into 3 equal amounts of precipitation and dropped at 8 h intervals similar to the NCEP dataset where it is dropped at 6 h intervals.We recognize more sophisticated diurnal precipitation algorithms using site observations could be developed, but more locations with sub-daily patterns of rainfall would be necessary for the region.
5. FSDS: Incoming shortwave radiation or incident solar is not provided in the input files.Again, DAYMET algorithms were used.The inputs required are daily minimum and maximum temperature, precipitation, latitude, longitude, and elevation all of which are available from the downscaled regional dataset.

BGD Introduction Conclusions References
Tables Figures

Back Close
Full

A2 Methods for making PFT modifications in CLM4
There are 17 PFTs in CLM4, eight of which are forested.Because we did not need the tropical or boreal PFTs, we reorganized the PFT physiological file to represent the variation between and within the temperate evergreen needleleaf PFT among the ecoregions.The old and new PFT assignments are shown in Tables A1 and A2.We also show the parameters that varied for each new PFT.The surface datasets and the dynamic PFT file were altered to match the new PFT assignments.Inventory and remote sensing data were used to identify within region PFT percentage cover for "fir" versus "pine" where necessary.The FLUXNET tower sites were also parameterized using site-specific observations and regional observations where site-specific observations were not available.Tower physiological parameterizations are in Table A3.

A3 Methods for stem allocation modification in CLM4
The PFT physiology file includes an option to have the fraction of annual NPP allocated to stem wood versus foliage change with the increasing annual sum of NPP throughout the year.The equation is for all forested PFTs: Allocation ratio = (2.7/(1.0+ exp(−0.004* (annsum npp(t) − 300.0)))) − 0.4 (A1) Where "annsum npp" is the total PFT NPP summed over the year at the current timestep (t).The ratio is constrained to be 0.20 when NPP = 0 and does not exceed 2.2 for NPP values greater than 1000 g C m −2 yr −1 .Using supplemental plot data, we found annual stem wood to foliage NPP allocation ratios as high as 3.7 for the mesic ecoregions and no higher than 2.0 for the semi-arid ecoregions (Fig. A1a).We were unable to investigate seasonal allocation ratios with the plot data data.We modified the equation to increase allocation to stem wood versus foliage as stands age in the mesic ecoregions compared to the default equation and decrease allocation to stem wood versus foliage in the semi-arid ecoregions compared to the default equation (Fig. A1b and Table A4).Introduction

Conclusions References
Tables Figures

Back Close
Full Where NPP is the annual sum of net primary production and the function represents smaller increases in biological nitrogen fixation at higher levels of NPP to incorporate the hypothesis that N fixation is eventually limited by other nutrients such as phosphorus.Maximum fixation rates do not exceed 1.8 g N m −2 yr −1 .The Campbell River fir site reports rates of 3 g N m −2 yr −1 (Arain et al., 2006).Pacific Northwest forests have reported additions of up to 15 g of N m −2 yr −1 from native symbiotic nitrogen fixing plant species which became established on cut or burned forested areas (Cleveland et al., 1999).However, the abundance and distribution of these species is limited to early successional forests following disturbance and would be difficult to predict by both ecoregion and plant functional type for the regional simulations.Because we knew the exact disturbance history and the reported rate for the Campbell River mature fir site (ca1) we modified the equation to yield up to 2.5 g N m −2 yr −1 retaining the limitation by other nutrients: where, the rate yields higher rates of fixation for lower values of annual NPP than the original equation to better represent successional dynamics.Introduction

Conclusions References
Tables Figures
Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | to validate model output and identify model biases.The associated uncertainty in the FIA estimates can also be calculated Figures Back Close Full Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Finally, the modified version of the code was used for spinup at the tower sites because the biological nitrogen fixation equation was changed for the Campbell River site (Arain et al., 2006; see Sect.2.6).Discussion Paper | Discussion Paper | Discussion Paper | ated uncertainty for the period from 2001-2006 were downloaded from the FLUXNET database for two tower sites in the region, one in a mesic fir forest of the Coast Range ecoregion (Campbell River, BC) and one in a dry pine forest of the East Cascades (Metolius, OR).The provided uncertainty estimates for the flux tower observations were calculated for the model-data synthesis project by (Schwalm et al., 2010).These sites represent mature secondary forest under private ownership (50-80 yr) with contrasting rates of biomass accumulation and NPP primarily due to climatic conditions (Law et al., Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | (i.e.Coast Range, West Cascades) and lower values in the less productive dry eastern ecoregions (East Cascades, Northern Basin).Stem wood biomass peaks in the West Cascades at 360 Mg C ha −1 and is lowest in the Northern Basin at 10 Mg C ha −1 .NPP is highest in the Coast Range with values up to 1100 g C m −2 yr −1 and lowest in Northern Basin at less than 100 g C m −2 yr −1 (Fig. 2b).Forest NEP ranges from −200 g C m −2 yr −1 in the Blue Mountains to 350 g C m −2 yr −1 in the Coast Range Introduction Discussion Paper | Discussion Paper | Discussion Paper | (Fig. 2c) and R h ranges from 20 g C m −2 yr −1 in the Northern Basin to 750 g C m −2 yr −1 Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

− 1 (
Hudiburg et al., 2009) in the more mesic ecoregions (Coast Range, Klamath Mountains, West Cascades).Reported densities from the same studies for the drier ecoregions (East Cascades, Blue Mountains) range from 20 to 100 Mg C ha −1 .The CLM4 stem estimates in this study range from 10-400 Mg C in agreement with the reported values.Pacific Northwest temperate forests have some of the highest carbon stocks in the world (Keith et al., 2009), with most of the carbon stored in the tree wood.Modeled relationships of biomass accumulation over time and with increasing Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |
Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | a. DAYMET and MTClim algorithms were used to calculate vapor pressure from temperature, precipitation, and solar radiation.Vapor pressure was then used to calculate relative humidity (RH): i. RH = 100 • (VP/SVP); where VP = the average daily vapor pressure in Pascals and SVP = the saturation vapor pressure.SVP varies with temperature.ii.SVP = 610.78• exp(T/(T + 238.3) • 17.2694); where T is the current temperature in • C.
Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | A4 Methods for biological nitrogen fixation modification in CLM4 Biological nitrogen fixation in CLM4 is a dynamic equation based on the annual sum of NPP and is updated on an annual time-step.The equation is for all PFTs (not PFT dependent): Nitrogen fixation (g N m −2 yr −1 ) = (1.8• (1.0 − exp(−0.003• NPP))) (A2) Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Table

Fig. 1 .
Fig. 1.Study region area (Oregon) divided by ecoregions (BM = Blue Mountains, CR= Coast Range, CP = Columbia Plateau, EC = East Cascades, KM = Klamath Mountains, NB = Northern Basin, SR = Snake River, WC = West Cascades, WV = Willamette Valley).Green circles represent FIA plots used for evaluating annual model output, blue triangles are the supplemental plots used for parameterization, and red squares are the FLUXNET tower sites available for seasonal validation (Campbell River site in British Columbia, Canada not shown).

Fig. 5 .Fig. 7 .
Fig. 5. Simulated monthly GPP versus observed GPP at two FLUNET eddy-covariance tower sites in the study region.Solid black circles and bars represent tower observations, grey squares are modeled values of GPP before downregulation due to nitrogen limitation, and blue crosses and bars are modeled GPP after downregulation.Black and grey error bars represent observed estimate uncertainty.(A) Monthly GPP for the years 2002-2007 at the Metolius mature pine site in Oregon, USA, (B) annual GPP at the Metolius mature pine site for 2002-2007, (C) monthly GPP for 1998-2003 at the Campbell River fir site, British Columbia, Canada, (D) annual GPP at the Campbell River fir site.

Table 2 .
Regional and FLUXNET tower site simulations.Transient CO 2 , nitrogen deposition (Ndep), and land cover files are annual files covering the period from 1850-2006.Ndep and land cover are spatially interpolated using the CLM toolkit from a half-degree global dataset while the CO 2 file is a single value for the entire region which changes annually.Simulations marked with and "*" indicate tower/point simulations (ca1 = Campbell River and me2 = Metolius Mature Site).

Table 3 .
Stem carbon and NPP reduced x 2 statistics and model bias by ecoregion, annual precipitation regime a , and age group b for CLM4 and CLM4 stem.Values marked with an "*" indicate good model performance (x 2 < 2).Reduced x 2 values close to 1 are within observation uncertainty.

Table A2 .
Model physiological parameters that vary for subgroups of PFTs in each ecoregion.