The Limitations and Robustness of Data Assimilation in Terrestrial Ecosystem Modelling.

Accurately estimating how much carbon is leaving the atmosphere and being taken up by plants, by processes such as photosynthesis, is critical in order to make accurate climate forecasts. There is a large uncertainty of this atmosphere-plant carbon flux, sometimes referred to as Net Ecosystem Exchange (NEE), therefore reducing this is essential. One way of doing this is through Data Assimilation (DA), the framework by which data and models are combined together in a statistically optimal way. While DA has gained much interest as a powerful tool in estimating model parameters and quantities such as NEE, there are a number of issues surrounding its use which include: (1) What quantity and quality of data are required in order for DA estimated parameters to be close to the truth? (2) Should the initial conditions of a model be treated as parameters to be estimated or fixed values? (3) Is DA robust against unrealistic features of satellite measurements (e.g. noisy data), and what can be done to correct for these? The aim of this PhD was to address and learn more about these issues. This was done by using the evergreen and deciduous versions of the Data Assimilation Linked Ecosystem (DALEC) model. Addressing question (1), our results showed that between two and five years of NEE data were required in order for the DA estimated parameters and NEE forecasts to be close to the truth. There was for the most part very little difference in the parameter estimates and NEE forecasts when very noisy or very non-noisy observations were used, or whether 20% or 100% of the daily observations were present in the dataset. For question (2), it was found that parameter estimates were sensitive to the initial value of the labile carbon store. Moreover, the parameters were close to their true values if the true initial value of the labile C pool was used. It was also found that when these initial conditions were treated as parameters, although the modal value of the corresponding marginal posterior distributions were far from the truth, every other aspect of the model (parameters and trajectories of the model states) agreed well with the truth. This supported the common approach by many of the DA community that treating initial conditions as parameters is preferable than keeping them fixed (using site inventory data or from model spin-up). The novelty of this part of the thesis was the use of a surrogate function (a Gaussian Process emulator) in place of the DA scheme for the purposes of carrying out the sensitivity analysis. Addressing question (3), DA was used to estimate the leaf area index (LAI) and NEE model states, using a fixed parameter set and LAI data from the MODIS satellite sensor. It was found that processing this satellite data in order to correct for unrealistic features of the dataset, such as excessive temporal variation and inaccurate estimates of the uncertainties, improved the fit of the modelled to observed NEE after assimilation. The improvement in the fit was significantly better for Gross Primary Production (GPP).