by Andrey Lessa
As you might already know from reading through our webpage, the ConFoBi projects do extensive field surveys in the search for connections between measures of forest structure and the various taxa. This invaluable information are used among others by the subproject C1 to conduct economic analyses and to plan conservation actions. The analyses and planning are done with the help of models, which are a simplification of the ecosystem under study. Nowadays one of the most important challenges in this area is the handling of uncertainty related to the systems under study. Climate change, sampling coverage, and numerous model parameters are only a few of the various uncertainties that affect the responses of our models. In order to provide robust recommendations we need to properly tackle and communicate them.
Recently, the use of Bayesian inference is gaining popularity among ecologists of various areas and has largely contributed to uncertainty assessment in ecological models. Among the various advantages of Bayesian calibration is the possibility to get a direct estimation of the parameter uncertainty related to our model predictors. In a recent paper arising from a joint effort from the subprojects A2, B6 and C1, Marco Basile has calibrated bird responses to forest structures using a hierarchical Bayesian model. It is of interest to identify the reasons why some parameters in the model have wider ranges than others. With that knowledge we can focus our efforts on collecting data that help to reduce this uncertainty and finally produce more accurate forecasts of bird responses to forest management. A simple option to realize that is to use a Monte Carlo simulation of the posterior parameter distribution and see how these changes the model’s responses. Afterwards, we can identify the most important parameters affecting the model responses, which will depend on how uncertain the parameters are and how sensitive the model is to these parameters. Here we did this exercise, sampling 1000 times the parameter distribution and calculating the model response, aiming to find the parameters causing the greatest variation in the estimated bird abundance (figure 1).
Our study area is dominated by Norway spruce and thus, not surprisingly, the parameter related to the share of conifers caused most variation on the model’s responses, since the sampling intensity in this category was smaller and the bird assemblage is sensitive to this landscape feature. Altitude was the second most important parameter. This suggests that increasing the sampling intensity to cover a wider variation of forest cover in terms of species composition (i.e. plots with larger share of broadleaves) and could reduce the uncertainty related to the model’s responses. This is straightforward using the Bayesian updating here to update the posterior distribution of the parameters.
In this simple example we could identify key sources of uncertainty for bird abundance forecasts. A targeted data collection to reduce the confidence interval of the abundance predictions can be of great value for conservation planning. Taking into account that tree species conversion is costly, more confident estimates on bird responses can help to define narrower targets for mixed forests, increasing the efficiency of conservation policies.