College of Engineering | Apply to CSU | Disclaimer | Equal Opportunity Statement | Privacy | Search CSU

Time-Filtered Inverse Modeling of Land-Atmosphere Carbon Exchange

You must be on the CSU network—either physically or using VPN—to watch this or any of the videos on this site.

April 3, 2015
Nick Geyer
Hosted by Scott Denning (advisor), Chris O'Dell, Jennifer Hoeting (Statistics)


The locations and changes of biospheric carbon dioxide sources and sinks represents one of the least understood processes regarding carbon science. The diagnosis of these sources and sinks come from carbon inversion models, which employ aggressive statistical regularizations in both space and time for stable flux estimation. These regularizations are focused toward estimating fluxes biases on weekly timescales over fine spatial scales, which causes estimates to suffer from statistical issues of observational and model constraints. This study addressed these problems by refocusing the observational information toward estimating longer lived biospheric biases rather than faster better understood ecosystem properties. The approach suggested here was to use the longer lived component CO2 fluxes in a harmonic Kalman filter-based inversion framework designed to characterize and "learn" about multiple timescale flux biases. This study used model estimates from the Simple Biosphere Model 4 (SiB4) and observations from North American Carbon Program (NACP) site synthesis from 8 eddy covariance towers to over constrain this framework. Three experiments were carried out to test this algorithm's performance against a control algorithm under various amounts observational uncertainty. The first experiment subjected the algorithm to optimize modeled net ecosystem exchange (NEE) from observed NEE. The result of this experiment was that the time smoothing algorithm was unable to make a better annual NEE estimation over the control, but its flux estimations were more robust to increasing observational uncertainty. The second experiment optimized modeled gross primary production (GPP) and total respiration (RESP) from observed NEE. This experiment concluded that both the annual and seasonal signals of NEE were estimated better than the control and very robust, but at the cost of physically unrealistic optimized component fluxes. The third experiment tested the optimization of modeled GPP and RESP when observed NEE and GPP are the constraints. This experiment showed that when a component flux was added as a constraining variable, NEE was better estimated than the control and now the component fluxes were both accurately and robustly estimated. Future work should be to use this framework as a prototype to be implemented inversion systems involving regional and global CO2 flux estimation.