Exploring robustness and uncertainties of projections with forest ecosystem models

Forests act as important CO2 sinks and might help to reduce the impacts of global and climate change. We can explore such scenarios with forest ecosystem models as their mechanistic struc- ture in principle allows forecasting into never-observed conditions. However, to make realistic projections, we have to adjust the model to fit the observed data. To do so and to assess uncer- tainties of projections, researchers use methods like a sensitivity and uncertainty analysis but also Bayesian calibration. However, the naive application and and the associated assumptions of these methods do often not reflect the empirical knowledge about forest ecosystems. To adress these issues, this doctoral thesis analyzes the robustusness of and applies these numerical meth- ods to state-of-the-art forest ecosystem models. We asked the following questions: The first one was: What are the main contributors of uncertainty in forest ecosystem models? Can we use uncertainty analysis and calibration of forest ecosystem models to analyze ecological patterns on environmental gradients? The next question deals with the remaining variance: To what ex- tent can random effects be used to represent ecological variation and how much data points are required to estimate these variations precisely? And the last question investigates if the findings above are robust when we have structural model error: What are the consequences and solutions of calibrating of and projecting with models with structural errors? The first chapter introduces the role of forest ecosystem models and their associated uncer- tainities for projecting forest dynamics under climate change and emerging challenges. In the second chapter, we explain key concepts and methods which are essential to understand our research results. In particular these are types of forest ecosystem models, their associated un- certainties (due to initial conditions, model inputs, model structure and parameters), sensitivity and uncertainty analysis and Bayesian calibration. In the third chapter, we analyze sensitivi- ties and uncertainties of carbon projections across European forests under climate change with a dynamic vegetation model (LPJ-GUESS 4.0) addressing the effect of both model parameters and environmental drivers. We find that carbon projections are most sensitive to photosynthesis- related parameters, while environmental drivers induce most uncertainty. Moreover, environ- mental drivers modify the uncertainties of other parameters. This study shows that environmen- tal drivers are strong contributors and modifiers of uncertainties in other ecosystem processes. In the fourth chapter, we analyze the consequences and possibilities to represent intraspecific variation in the calibration of a forest ecosystem model. To do so, we calibrate the 3-PG model against biomass derived from inventory data across Germany and Sweden with a hierarchical Bayesian calibration scheme. We find evidence for intraspecific variation that can be partly ex- plained by environmental conditions. This study shows the potential of using forest ecosystem models to infer not measurable ecological information. In the fifth chapter, we analyze if with a low number of levels it is better to model a grouping variable as a random or as a fixed-effect. We find with varying intercepts and slopes in the data-generating process, using a random slope and intercept model, and, in case of a singular fit switching to a fixed-effects model, avoids over- confidence in the results. This study shows how to make ecological inference with mixed-effects models more robust for a small number of levels. In the sixth chapter, we explain why model error causes bias and underestimated uncertainties, especially when calibrated against unbal- anced data, and propose a framework for robust inference with complex computer simulations. As possible solutions we discuss data rebalancing and adding bias corrections during or after the calibration procedure. We illustrate the methods in a case study, using a dynamic vegetation model. From this, we conclude that developing better methods for robust inference of complex computer simulations is essential for generating reliable predictions. The last chapter discusses the relevance and significance of our studies for forecasting and inference with forest ecosystem models and outlines further research questions