Using Remote Sensing Data to Help Constrain Fluxes of CO2 in a Geostatistical Inverse Modeling Framework

Current methods for estimating surface fluxes of CO2 can be divided into "bottom-up" approaches which quantify fluxes based on inventories of variables with a known influence on carbon fluxes, and "top-down" or inverse modeling approaches, which use information derived from atmospheric concentration measurements and atmospheric transport models to infer these fluxes. However, there is a growing need to estimate fluxes and their uncertainties at a finer spatial and temporal resolution than current models provide while limiting the reliance on inventory estimates in inverse modeling studies. The overall goal of this work is to estimate carbon dioxide fluxes using the recently developed geostatistical approach (Michalak et al, 2004) that replaces the use of inventory data as priors with a spatial and/or temporal covariance based on the separation distance between flux locations. As such, geostatistical flux estimates are not subject to some of the limitations of traditional Bayesian inversions, such as potential biases created by the choice of prior fluxes and aggregation error resulting from the use of large regions with prescribed flux patterns. The method is also ideally suited to inversions at fine spatial scales. In this presentation, we develop the tools necessary to incorporate auxiliary environmental information to help constrain the geostatistical inversion. Many of the auxiliary data are derived from remote sensing measurements, allowing these products to be used directly to help calculate the estimates of the fluxes. It is important to note that although this additional information is included in the inversion setup, no correlation between these data and carbon fluxes is assumed a priori, unlike in classical Bayesian approaches. Instead, this correlation is derived from the atmospheric data themselves. Auxiliary data collected by NOAA's Advanced Very High Resolution Radiometer (AVHRR) used in the geostatistical model include: leaf area index (LAI), normalized difference vegetation index (NDVI), fraction of photosynthetically active radiation (fPAR), and land cover type. Other datasets were obtained from a variety of sources (e.g. NOAA's NOAA-CIRES Climate Diagnostic Center) and include: sea surface temperature, air temperature, population, palmer drought index and precipitation. The final set of auxiliary variables was identified by performing a Variance Ratio Test that quantified the significance of the inferred correlation between the environmental variables and estimated carbon fluxes, based on the information contained in the atmospheric concentration measurements. Preliminary results show that LAI, fPAR, population, and certain land use categories significantly influence the mean flux value at a given location. Covariance parameters for the model of the mean and the model-data mismatch used in the inversion were optimized using the atmospheric data themselves, through the Restricted Maximum Likelihood Approach. The newly developed method is used to estimate global monthly-averaged fluxes of carbon dioxide for 1997 through 2001 on a 3.75o by 5.0o scale. These results are presented in an accompanying presentation, "Fine spatial resolution global CO2 flux estimates for 1997 to 2001 obtained using remote-sensing derived environmental data within a geostatistical inverse model," by S. Gourdji et al.