Upscaling of biogeochemical fluxes to characterize regional carbon balanc
Modeling of land surface fluxes (carbon, nitrogen, water and energy) is essential to improve the understanding and predictability of feedback mechanisms between climate change, ecosystem behavior, hydrological processes and land use. In order to support future decision making in climate politics and environmental planning, it is important to implement and enhance land surface modelling at the regional level. The central goal of this study was to upscale net ecosystem exchange (NEE) from eddy covariance (EC) footprint scale to the Rur catchment domain. This was done with help of the Community Land Model (CLM) and updated ecosystem parameter values, which were estimated with help of measured NEE data. Such model-data fusion methods and also the evaluation of the modelled NEE outputs require an accurate estimate of the measurement uncertainty.
Therefore, first the uncertainty of eddy covariance NEE measurements was investigated for one grassland site inside the Rur catchment. The classical two-tower approach estimates the measurement uncertainty or random error based on the standard deviations of the fluxes measured simultaneously at two nearby EC towers. The random error is estimated as function of the flux magnitude with help of a linear regression equation which allows then to estimate the random error as function of the measurement value. Various other uncertainty estimation approaches have been developed but none is generally accepted and applied. The two-tower approach assumes statistical independence of the measured data (non-overlapping footprints) and at the same time identical environmental conditions in the footprint of both EC towers. Because those two requirements seem contradictory, the applicability of the method might be hampered and the definition of an appropriate tower distance is difficult. To solve this issue, an extension of the classical two tower approach was formulated, which corrects for systematic differences of the synchronously measured EC data from two EC stations and thus reduces the overestimation of the uncertainty estimate. It was assumed that those systematic flux differences mainly arise from different environmental conditions in the footprint area of both EC stations and thus increase with the EC tower distance. The role of the EC tower distance was investigated by applying and evaluating the uncertainty estimation for five different EC tower distances ranging between 8m and 34km. The analysis was made for a dataset where (i) only similar weather conditions at the two sites were included, and (ii) an unfiltered one. The proposed extension of the two tower approach applied to weather-filtered data notably reduced the overestimation of the two-tower based NEE measurement uncertainty for all separation distances except for 8m. The NEE measurement uncertainty reduced by 79% (34km distance) to 100% (95m distance). A major conclusion of this study is that the extension of the two tower approach raised the applicability of the two-tower approach to more site pairs with less ideal conditions, i.e. not very similar environmental conditions.
In a simple sensitivity study, eight key ecological parameters were identified which strongly determine the simulated carbon fluxes in the Community Land Model v.4.5 (CLM). Those parameters were then estimated with the Markov Chain Monte Carlo method DREAM (DiffeRential Evolution Adaptive Metropolis) separately for four sites with different land cover types: C3-grass, C3-crop, evergreen coniferous forest and broadleaf deciduous forest. Those are the four main plant functional types (PFTs) in the Rur catchment and cover more than 90 % of the whole catchment area. The four sites were located inside or close to the Rur catchment. From the eight parameters estimated, five are PFT specific and three are hard coded in the CLM source code. Hard coded parameters such as the temperature coefficient Q10 do not vary among PFTs in CLM. However, various studies indicate that for example Q10 is not a constant but varies depending on the PFT and the environmental site conditions. Therefore the eight parameters were estimated jointly for each site (PFT). The parameters were constrained with half hourly NEE time series measured at the four EC sites. The NEE time series covered a whole year (including gaps). In addition, parameters were estimated separately for the single seasons of this one year period in order to test if accounting for seasonal variations of the CLM parameters would improve the simulated NEE. In addition, it was investigated how strongly the CLM parameter estimates and initial states are linked. This was done by means of an additional experiment where four initial state factors for the CLM carbon-nitrogen pools and the leaf area indices (LAI) were estimated jointly with the parameters. The maximum a posteriori (MAP) estimates obtained from DREAM were evaluated by i) half hourly NEE time series at the same sites, but the next year, and ii) NEE data from four FLUXNET sites situated more than 500km away from the original sites. To comprehensively evaluate the parameter estimates, different evaluation indices were calculated which express the mismatch of modelled and measured NEE data: (i) the relative difference of the annual NEE sum (RDsum), (ii) the root mean square error for the one-year time series of half-hourly NEE data (RMSEm), (iii) the mean absolute difference of the mean diurnal NEE cycle (MADdiur), and (iv) the mean absolute difference of the monthly mean NEE values over the evaluation year (MADann). Those indices were calculated and compared for the CLM runs with updated parameters and a reference run with CLM default parameters. It was found that parameter values underlie a strong seasonal variability and parameters estimated on seasonal basis outperformed the parameters estimated based on the whole one year period. With parameters estimated separately for the four seasons, RDsum was on average 32% lower and RMSEm 49% lower than for the reference run, averaged over all sites. The respective MADdiur and MADann indices were reduced by a factor of 1.6 (MADdiur) and 1.4 (MADann). This highlights that estimated parameters notably improved the constancy of modelled and measured NEE data as well as the model representation of the simulated mean diurnal and annual NEE cycle. The evaluation was more successful for the forest PFTs than for the C3-grass and C3-crop sites. Along with this result, correlations between parameters and the initial state factors were found to be higher for C3-grass and C3-crop than for the forest PFTs. It was concluded that parameter estimates compensate model structural deficits, particularly for C3-grass and C3-crop.
In a following study, DREAM-CLM parameter estimates were applied to the whole Rur catchment domain and thus used to upscale NEE data from EC footprint to catchment scale. New parameter sets were estimated using the same data and the same DREAM-CLM setup as in the previous study, but this time only considering the five PFT specific parameters. A PFT specific definition of the three hard wired parameters for a regional model setup would require major changes in the CLM source code structure. To evaluate the performance of the new parameter estimates, different evaluation indices including MADdiur and RDsum were calculated for (i) the mean of a 60 member CLM ensemble with parameters sampled from the joint posterior probability distribution function, and (ii) a reference run with default parameters. The CLM performance with and without updated parameters was evaluated based on NEE data of seven EC towers inside the Rur catchment. The difference between the observed and simulated NEE sum for the evaluation period (Dec. 2012 – Nov. 2013) was 23% smaller if DREAM-parameters instead of default parameters were used as input. This indicates that parameter estimates can provide a more reliable estimate of the annual NEE balance for the catchment than global default parameters. However, results suggested that parameters estimated for a particular crop type cannot be transferred to grid cells where very different crops types are grown, such as sugar beet instead of winter wheat. In addition it was tested if parameter estimates improve simulated LAI for the catchment. This was evaluated using LAI data obtained from remotely sensed RapidEye images. Results showed that the misfit between modelled and observed LAI data was notably reduced if estimated
parameters instead of CLM default parameters were used, particularly in case of C3-grass and C3-crop. For those PFTs, the mean absolute difference between observed and modelled LAI data (MADLAI) was about 52% lower with parameter estimates. Evaluation indices calculated for the forest PFTs were considered less meaningful because the error in RapidEye based LAI data is assumed high for those PFTs. For the whole Rur catchment, independent of PFT distribution within the grid cell, MADLAI was reduced by 59% on average. Predicted LAI with estimated parameters was not only improved in terms of magnitude but in some cases also in terms of timing (beginning of plant onset in spring). This was strongly liked to improvement of NEE.
Weblink to former H2020-project (2010-2015):
Other relevant link:
Prof. Dr. Harrie-Jan Hendricks Franssen
Forschungszentrum Jülich GmbH
Tel. +49 02461 4462
Post, H., H.J. Hendricks Franssen, X. Han, R. Baatz, C. Montzka, M. Schmidt, and H. Vereecken, 2017. Evaluation and uncertainty analysis of regional scale CLM4.5 net carbon flux estimates. Biogeosciences Discussions, https://doi.org/10.5194/bg-2016-540.
Post, H., J.A. Vrugt, A. Fox, H. Vereecken, and H.J. Hendricks Franssen. 2017. Estimation of Community Land Model parameters for an improved assessment of net carbon fluxes at European sites. Journal of Geophysical Research - Biogeosciences, doi:10.1002/2015JG003297.
Post, H., H.J. Hendricks Franssen, A. Graf, M. Schmidt, and H. Vereecken. 2015. Uncertainty analysis of eddy covariance CO2 flux measurements for different EC tower distances using an extended two-tower approach. Biogeosciences, 12, 1205-1221.