Browsing by Subject "Bayesian calibration"

Now showing 1 - 2 of 2

Bayesian multi-purpose modelling of plant growth and development across scales
(2024) Viswanathan, Michelle; Streck, Thilo
Crop models are invaluable tools for predicting the impact of climate change on crop production and assessing the fate of agrochemicals in the environment. To ensure robust predictions of crop yield, for example, models are usually calibrated to observations of plant growth and phenological development using different methods. However, various sources of uncertainty exist in the model inputs, parameters, equations, observations, etc., which need to be quantified, especially when model predictions influence decision-making. Bayesian inference is suitable for this purpose since it enables different uncertainties to be taken into account, while also incorporating prior knowledge. Thus, Bayesian methods are used for model calibration to improve the model and enhance prediction quality. However, this improvement in the model and its prediction quality does not always occur due to the presence of model errors. These errors are a result of incomplete knowledge or simplifying assumptions made to reduce model complexity and computational costs. For instance, crop models are used for regional scale simulations thereby assuming that these point-based models are able to represent processes that act at regional scale. Additionally, simple statistical assumptions are made about uncertainty in model errors during Bayesian calibration. In this work, the problems arising from such applications are analysed and other Bayesian approaches are investigated as potential solutions. A conceptually simple Bayesian approach of sequentially updating a maize phenology model, an important component in plant models, was investigated as yearly observation data were gathered. In this approach, model parameters and their uncertainty were estimated while accounting for observation uncertainty. As the model was calibrated to increasing amounts of observation data, the uncertainty in the model parameters reduced as expected. However, the prediction quality of the calibrated model did not always improve in spite of more data being available for potentially improving the model. This discrepancy was attributed to the presence of errors in the model structure, possibly due to missing environmental dependencies that were ignored during calibration. As a potential solution, the model was calibrated using Bayesian multi-level modelling which could account for model errors. Furthermore, this approach accounted for the hierarchical data structure of cultivars nested within maize ripening groups, thus simultaneously obtaining model parameter estimates for the species, ripening groups and cultivars. Applying this approach improved the model's calibration quality and further aided in identifying possible model deficits related to temperature effects in the post-flowering phase of development and soil moisture. As another potential solution, an alternative calibration strategy was tested which accounted for model errors by relaxing the strict statistical assumptions in classical Bayesian inference. This was done by first acknowledging that due to model errors, different data sets may yield diverse solutions to the calibration problem. Thus, instead of fitting the model to all data sets together and finding a compromise solution, a fit was found to each data set. This was implemented by modifying the likelihood, a term that accounts for information content of the data. An additive rather than the classical multiplicative strategy was used to combine likelihood values from different data sets. This approach resulted in conservative but more reliable predictions than the classical approach in most cases. The classical approach resulted in better predictions only when the prediction target represented an average of the calibration data. The above-mentioned results show that Bayesian methods with representative error assumptions lead to improved model performance and a more realistic quantification of uncertainties. This is a step towards the effective application of process-based crop models for developing suitable adaptation and mitigation strategies.
Mid-infrared spectroscopy and enzyme activity temperature sensitivities as experimental proxies to reduce parameter uncertainty of soil carbon models
(2021) Laub, Moritz; Cadisch, Georg
Models that simulate the dynamics of soil organic carbon (SOC) are crucial to understand the global carbon cycle, but current generation models are subject to major uncertainties due to two model shortcomings. Firstly, their different carbon pools are not connected to measurable SOC fractions. Secondly, there is uncertainty about the response of the different carbon pools to an increasing temperature. The aim of this thesis was thus to link the SOC model pools of the Daisy model to measurable proxies for SOC quality and pool specific temperature sensitivity. In the first study, the drying temperature for soil samples assessed by diffuse reflectance mid infrared Fourier transform spectroscopy (DRIFTS) was optimized to assure optimal representativeness of aliphatic and aromatic-carboxylate absorption bands as proxies for fast- and slow-cycling SOC pools. Their ratio was termed the DRIFTS stability index (DSI). In the second study, the DSI was used to distinguish fast- and slow-cycling SOC model pools at model initialization. In the third study, model initialization using DSI was performed to infer pool specific temperature sensitivities for the different Daisy carbon pools. Furthermore, it was tested whether the measured temperature sensitivities of different extracellular soil enzymes could be used as proxies for pool specific temperature sensitivity. Using a global collection of soil samples revealed that the absorption of all studied DRIFTS absorption bands increased significantly (p < 0.0001) with increasing drying temperature from 32°C to 105°C. This effect was disproportionally strong for the aliphatic absorption band. Due to the strong interference of the residual soil sample moisture content with the aliphatic absorption band, drying at 105°C and storage in a desiccator prior to measurement would be necessary for representative spectra for model pool initialization. In the following, a combination of medium to long-term bare fallow experiments was used, to test the utility of the DSI for SOC pool initialization. Pool partitioning by the DSI was superior to using a fixed pool partitioning under the assumption that SOC was at steady state. The DSI contained robust information on SOC quality across sites. Therefore, in the majority of cases, the application of the DSI led to significantly lower model errors than the steady state assumption. Furthermore, the application of the DSI in Bayesian calibration led to a reduced parameter uncertainty for the turnover of the slow-cycling SOC pool and the humification efficiency. The 95% credibility interval of the slow-cycling SOM pools’ half-life between 278 and 1095 years suggested faster SOC turnover than earlier studies. The DSI used for SOC model pool initialization was then combined with the lignin-to-nitrogen ratio for litter pool initialization to infer pool specific temperature sensitivities. The simulations of five field studies and laboratory incubations with fallow soil and crop-litter inputs were combined. Based on a clear pool definition, pool specific temperature sensitivities could be inferred by Bayesian calibration. However, differences in temperature sensitivities of the same pools between experiments suggested that carbon stability was not the main driver of temperature sensitivities. Instead, the main difference was found between the laboratory incubations (higher Q10 values up to 3) and the field (lower Q10 values centered around 2). In a second approach, the measured Q10 value of phenoloxidase (1.35) was used as Q10 value of the temperature function of both SOM pools and the slow crop-litter pool while ß glucosidase (1.82) was used for the fast crop litter pool. This improved field simulations by 3 to 10% compared to assuming a standard Q10 of 2 for all pools. Thus, site specific Q10 of different soil enzymes showed potential as proxy for site and pool specific temperature sensitivities. Important state variables that explain the observed Q10 value differences between experiments were identified as physical protection of SOC, substrate availability and environmental stress for microorganisms due to fluctuating state variables in the field. In conclusion, the usefulness of the DSI as an indicator of SOC stability and proxy for pool initialization was demonstrated for several soils in central Europe. In addition, it was shown that pool partitioning proxies can help to infer pool specific temperature sensitivity by Bayesian calibration. However, temperature sensitivity was not mainly a function of carbon stability.