Connecting process-based vegetation model to data from the Biodiversity Exploratories to unravel the mechanisms that connect biodiversity, land-use and ecosystem function
Final Report Abstract
In this project, we developed and applied a modelling framework for calibrating a species-level, forest ecosystem model (LPJ-GUESS) with a parallelisable Bayesian algorithm (SMC) to data from the Biodiversity Exploratories. Thereby LPJ-GUESS can be used to synthesise information collected across a wider range of sub-projects into a process-based model. The first important step in this framework was to wrap the C++-coded LPJ-GUESS into an R-package (Rlpj) in order to facilitate calling it from R and thereby allowing sensitivity analyses and calibration from the R environment. This has not been done before in such an easy, reproducible and comprehensive way as with Rlpj. The second step was to implement and adapt a sequential Monte-Carlo sampler (SMC, also known as particle filtering), which, in contrast to traditional Markov Chains, can be parallelised and hence allows Bayesian calibration of models that have runtimes prohibitive of frequent (> 106) executions on a single core. The algorithm is implemented as part of R’s BayesianTools-package. A publication demonstrating the equivalence of its estimation to current standard MCMCs as well as their runtime efficiency in a parallelised setting is submitted. Finally, one of the important resources of forest growth is light, and we thus measured forest floor light conditions in 75 forest plots in 2017. The light transmission was a good (linear and positive) predictor of understorey plant species richness and provides one of very few actual light measurement-richness studies in the field. We are hoping to bring the calibration of LPJ-GUESS on data from an exploratory site in Haining to a successful conclusion. Overall, our project demonstrated the feasibility, but also the technical challenges, of linking a diverse and non-systematic set of field data to a detailed process-based ecosystem model. We think that this approach offers great potential for synthesis beyond correlative approaches.
Publications
- 2018. An R package facilitating sensitivity analysis, calibration and forward simulations with the LPJ-GUESS dynamic vegetation model. Environmental Modelling & Software 111, 55-60
Bagnara, M., Silveyra Gonzalez, R., Reifenberg, S., Steinkamp, J., Hickler, T., Werner, C., Dormann, C.F. & Hartig, F.
(See online at https://doi.org/10.1016/j.envsoft.2018.09.004)
