The separation of the relative contributions of anthropogenic factors and stochastic natural factors is of particular importance for long-term sustainable management of semi-natural ecosystems. Due to the long inherent time-scales of vegetation change and a lack of long-term monitoring data, the separation of the effects of the two basic factors affecting vegetation in semi-arid ecosystems, i.e., grazing and highly variable rainfall has not been possible empirically and was not explicitly addressed by modelling studies. The general aim of this PhD is to provide an understanding of the small-scale processes involved in degradation. More specifically, I present an individual- and rule-based stochastic and spatially-explicit simulation model to investigate the effect of grazing under stochastic rainfall on the perennial tussock grass Festuca pallescens and to separate the causal effects on F. pallescens dynamics. One essential characteristic of the simulation model is that both exogenous drivers – grazing and precipitation – act on each demographic process of each individual grass tussock. This property of the model will finally facilitate the separation of the relative effects of both drivers for each simulated time step on F. pallescens dynamics. To respond to its aim, the model needs to include a number of detailed factors affecting the Festuca population dynamics under grazing. This is reflected in an intermediate complexity with some thirty model parameters. Due to the lack of field data most of these parameters could not be estimated directly. To calibrate the simulation model I followed the indirect multi-criterial pattern-oriented approach, which was developed by Wiegand et al. (2003). Within this PhD thesis I further developed this approach, using a step wise and cross calibration. I showed that the medium complex Festuca model with 30 free parameters can be calibrated with a small field data set to produce behaviour in accordance with field observations. An extensive sensitivity analysis showed the novel result that lateral local water redistribution has a relevant impact on the behaviour of the dynamics of the grazed ecosystem. It further revealed that the model system is highly sensitive against the senescence and the littering rate. Both parameters compete with the herbivores for green biomass. After model calibration (or with the expert parameterisation) I performed various simulation experiments to investigate the behaviour of the system and to investigate the effect of grazing on vegetation under stochastic rainfall. I found that the Festuca steppe shows an event-driven behaviour which is modified by grazing gradually as well as qualitatively. The vegetation cover showed threshold behaviour under the grazing gradient which is also reflected by the recovery times of the grazed system. I found that the beginning of vegetation decrease depends on both climatic and biological uncertainty. I proposed a method for determining risk levels for degradation and for the determination of long-term sustainable stocking rates. I analysed temporal autocorrelations of essential variables of F. pallescens to show how memory effects influence its complex vegetation dynamics. Finally I separated the relative effects of grazing and precipitation on vegetation for specific precipitation time-series. This elucidated the link between the short-term interaction of grazing and precipitation and the observed long-term grazing threshold. The presented simulation model improves the understanding of the effect of small-scale biological processes on patterns emerging at larger scales as e.g. the landscape scale. This PhD-thesis contributes significant new insights into the interaction of grazing and stochastic precipitation in semi-arid systems and provides instruments to estimate degradation risk considering biological and climatic uncertainty. Furthermore the presented indirect multi-criterial pattern-oriented calibration method helps to bridge the gap between theoretical and empirical ecology as it enables us to gain strong confidence into simulation models even if we dispose only over scarce evidence from empirics. This aspect leads not only to a better understanding of ecosystems which are endangered by land use and strongly affected by stochastic environmental processes, but is also of general interest for simulation models facing a high degree of uncertainty because this method allows to tie the model closely to the data, i.e. ensuring a biologically reasonable behaviour and parameter values.