Loading…
Developing models for winter processes in grasslands – Bayesian calibration and sensitivity analysis
For grass-based agriculture at high latitudes, poor overwintering of perennial forage grasses often has economic consequences due to yield loss and re-establishment of grass fields. In order to assess the performance of grass cultivars currently used in Norway under a future changing climate, a whol...
Saved in:
| Main Author: | |
|---|---|
| Format: | Dissertation |
| Language: | English |
| Online Access: | Request full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | For grass-based agriculture at high latitudes, poor overwintering of perennial forage grasses often has economic consequences due to yield loss and re-establishment of grass fields. In order to assess the performance of grass cultivars currently used in Norway under a future changing climate, a wholeyear grassland model has been developed. The basis of this whole-year model was a grassland model developed for the growing season. In order to incorporate the winter season, this grassland model needed additional sub-models for simulating snow cover, soil frost, ice encasement and the development of frost tolerance in the plants. The main objective of this thesis has been to develop these additional sub-models, calibrate them using Bayesian methods and identifying key parameters using sensitivity analysis. The sub-models were also used to construct agroclimatic indices in order to assess the impact of climate change on the winter survival of two forage grasses.
There are several challenges emerging when applying Bayesian calibration to a dynamic model. The Bayesian approach regards parameters as random and allows integration of prior knowledge. Using the snow cover sub-model as case study, it is here demonstrated how prior information and new data affect the calibration process, parameters and model outputs, with focus on uncertainty. Point estimates and uncertainties are calculated and visualized for both parameters and model outputs. Generally, uncertainty decreased when new data were incorporated. Uniformly distributed priors gave the best fit for this model according to root mean square error, while the more informative beta distributed priors gave more physically meaningful parameter estimates. Markov chains of samples from the posterior distribution of the parameters were obtained by the random walk Metropolis-Hastings algorithm. Crucial points when using these methods are reaching and determining convergence of these chains. In order to reach convergence faster, informative priors, Sivia’s likelihood, reflection and updating the proposal distribution with parts of the data gave successful results. To determine convergence objectively and correctly, the use of multiple chains and the Gelman Rubin method was found useful. Several decisions must be made when implementing Bayesian calibration, and we highlight and visualize the choices that were found to be most effective.
We developed a simple model SnowFrostIce which simulates depth of snow cover, the low |
|---|