Processing arctic eddy‐flux data using a simple carbon‐exchange model embedded in the ensemble Kalman filter

• Published in: Ecological Applications
• Main Authors: Rastetter, Edward B., Williams, Mathew, Griffin, Kevin L., Kwiatkowski, Bonnie L., Tomasky, Gabrielle, Potosnak, Mark J., Stoy, Paul C., Shaver, Gaius R., Stieglitz, Marc, Hobbie, John E., Kling, George W.
• Format: Article in Journal/Newspaper
• Language: English
• Published: Wiley 2010
• Subjects: Brooks Range; Tundra; Alaska
• Online Access: http://dx.doi.org/10.1890/09-0876.1; https://api.wiley.com/onlinelibrary/tdm/v1/articles/10.1890%2F09-0876.1; https://esajournals.onlinelibrary.wiley.com/doi/pdf/10.1890/09-0876.1

Continuous time‐series estimates of net ecosystem carbon exchange (NEE) are routinely made using eddy covariance techniques. Identifying and compensating for errors in the NEE time series can be automated using a signal processing filter like the ensemble Kalman filter (EnKF). The EnKF compares each measurement in the time series to a model prediction and updates the NEE estimate by weighting the measurement and model prediction relative to a specified measurement error estimate and an estimate of the model‐prediction error that is continuously updated based on model predictions of earlier measurements in the time series. Because of the covariance among model variables, the EnKF can also update estimates of variables for which there is no direct measurement. The resulting estimates evolve through time, enabling the EnKF to be used to estimate dynamic variables like changes in leaf phenology. The evolving estimates can also serve as a means to test the embedded model and reconcile persistent deviations between observations and model predictions. We embedded a simple arctic NEE model into the EnKF and filtered data from an eddy covariance tower located in tussock tundra on the northern foothills of the Brooks Range in northern Alaska, USA. The model predicts NEE based only on leaf area, irradiance, and temperature and has been well corroborated for all the major vegetation types in the Low Arctic using chamber‐based data. This is the first application of the model to eddy covariance data. We modified the EnKF by adding an adaptive noise estimator that provides a feedback between persistent model data deviations and the noise added to the ensemble of Monte Carlo simulations in the EnKF. We also ran the EnKF with both a specified leaf‐area trajectory and with the EnKF sequentially recalibrating leaf‐area estimates to compensate for persistent model‐data deviations. When used together, adaptive noise estimation and sequential recalibration substantially improved filter performance, but it did not improve performance ...