Abteilung Systemanalyse, Integrated Assessment und Modellierung

Species distribution models (SDMs) are often criticised for lacking explicit linkage to ecological concepts. We aim to improve the ecological basis of SDMs by integrating prior knowledge about ecological preferences of organisms. Additionally, we aim to support a systematic, data-driven review of such prior knowledge by confronting it with independent monitoring data using Bayesian inference. We developed a series of multi-species distribution models (MSDMs) with increasing complexity to predict the probability of occurrence of taxa at sampling sites based on habitat suitability functions that are parameterized with prior ecological knowledge. We subsequently assessed the models` predictive performance with 3-fold cross-validation. So far, if ecological preferences or functional traits have been used in SDMs, they were mainly used as fixed inputs without considering their uncertainty. We take the additional step of considering uncertainty about preference parameters by including them as uncertain prior information that is subsequently updated with Bayesian inference. We apply the series of models in a case study on macroinvertebrates in Swiss streams. We analyse differences in the quality of fit, changes in predictive performance, and the potential to learn about the parameters from the data. We consider ecological preferences for natural and human modified environmental factors including temperature, flow velocity, organic matter concentration, insecticide pollution, and substratum. Results indicate that updating prior knowledge on ecological preferences with Bayesian inference, rather than using it as fixed input, improves model fit and predictive performance. For example, the predictive performance measured by the deviance for validation data improves by 17 % and the explanatory power increases 3.8 times from a model that treats ecological preferences as fixed scores to a model that treats them as uncertain parameters. The spatial distribution of many taxa, including rare taxa with frequencies of occurrence down to about 5 %, which are difficult to model with SDMs that do not consider prior information, can be captured by the new models. Integrating prior knowledge as uncertain parameters in a Bayesian framework establishes ecological interpretable links between taxa and their environment and supports a systematic revision and complementation of databases on ecological preferences, even in case of poor or missing prior knowledge. Model outputs need to be carefully interpreted by modellers and experts on ecological preferences. Increased exchange between these research fields will benefit further integration of ecological preferences into SDMs.

Understanding and predicting the ecological consequences of different management alternatives is becoming increasingly important to support environmental management decisions. Ecological models could contribute to such predictions, but in the past this was often not the case. Ecological models are often developed within research projects but are rarely used for practical applications. In this synthesis paper, we discuss how to strengthen the role of ecological modeling in supporting environmental management decisions with a focus on methodological aspects. We address mainly ecological modellers but also potential users of modeling results. Various modeling approaches can be used to predict the response of ecosystems to anthropogenic interventions, including mechanistic models, statistical models, and machine learning approaches. Regardless of the chosen approach, we outline how to better align the modeling to the decision making process, and identify six requirements that we believe are important to increase the usefulness of ecological models for management support, especially if management decisions need to be justified to the public. These cover: (i) a mechanistic understanding regarding causality, (ii) alignment of model input and output with the management decision, (iii) appropriate spatial and temporal resolutions, (iv) uncertainty quantification, (v) sufficient predictive performance, and (vi) transparent communication. We discuss challenges and synthesize suggestions for addressing these points.

Aim: Species distribution models (SDMs) are widely used to study geographic distributions of taxa in response to natural and anthropogenic environmental conditions. For a community, common approaches include fitting individual SDMs (iSDMs) to all taxa or directly modelling community properties such as richness. However, the parameters of iSDMs are difficult to identify for rare taxa, and community properties do not reveal taxon‐specific responses. Individual models can be combined into a hierarchical multispecies distribution model (mSDM) that constrains taxon‐specific parameters according to overarching community parameters, or a joint model (jSDM) in which interdependencies between taxa are jointly inferred. We compare how individual, hierarchical multispecies and joint SDMs differ in quality of fit, explanatory power and predictive performance, and analyse how these properties depend on the prevalence of taxa.
Taxa: Presence–absence observations of 245 benthic macroinvertebrate taxa identified at a mixed taxonomic resolution.
Location: Four hundred and ninety‐two sites in rivers throughout Switzerland.
Methods: Individual, hierarchical and joint hierarchical generalized linear models (GLM) were developed for all taxa. Parameters were estimated using maximum likelihood estimation or Bayesian inference with Hamiltonian Markov chain Monte Carlo simulations. Predictive performance was assessed with cross‐validation. In addition, the predicted family and species richness of the models was compared with a GLM for richness.
Results: Individual models show a slightly higher quality of fit largely due to overfitting for rare taxa. The mSDM achieves a similar quality of fit and explanatory power, mitigates overfitting for rare taxa and considerably improves predictive performance over the whole community. The joint models further improve the quality of fit, but decrease predictive performance and increase predictive uncertainty.
Main conclusions: We show that even a relatively simple mSDM combines many of the analytical capabilities of iSDMs and improves predictive performance. Increasingly complex mSDMs and jSDMs provide additional analytical possibilities, but depending on the data and research questions, different levels of complexity may be appropriate.

The widespread application of deterministic hydrological models in research and practice calls for suitable methods to describe their uncertainty. The errors of those models are often heteroscedastic, non-Gaussian and correlated due to the memory effect of errors in state variables. Still, residual error models are usually highly simplified, often neglecting some of the mentioned characteristics. This is partly because general approaches to account for all of those characteristics are lacking, and partly because the benefits of more complex error models in terms of achieving better predictions are unclear. For example, the joint inference of autocorrelation of errors and hydrological model parameters has been shown to lead to poor predictions. This study presents a framework for likelihood functions for deterministic hydrological models that considers correlated errors and allows for an arbitrary probability distribution of observed streamflow. The choice of this distribution reflects prior knowledge about non-normality of the errors. The framework was used to evaluate increasingly complex error models with data of varying temporal resolution (daily to hourly) in two catchments. We found that (1) the joint inference of hydrological and error model parameters leads to poor predictions when conventional error models with stationary correlation are used, which confirms previous studies; (2) the quality of these predictions worsens with higher temporal resolution of the data; (3) accounting for a non-stationary autocorrelation of the errors, i.e. allowing it to vary between wet and dry periods, largely alleviates the observed problems; and (4) accounting for autocorrelation leads to more realistic model output, as shown by signatures such as the flashiness index. Overall, this study contributes to a better description of residual errors of deterministic hydrological models.

Trade in commodities implies trade in virtual water (VW), which refers to the water that was used to produce the traded goods. Various studies have quantified international or inter-provincial virtual water (VW) flows related to the trade in crops and animal products. Until date, however, no effort has been undertaken to understand how the water embodied in traded feed crops (trade stage TS1) will be transferred further because of trade in animal products (trade stage TS2). This is the first study showing this mechanism, in a case study in China for maize (the major pig feed) and pork (the dominant meat), considering the period 2000–2013. We estimate the annual green and blue water footprints in maize production and then quantify the inter-provincial VW flows related to trade in maize (TS1) and trade in maize embodied in pork (TS2). Results show that in TS1, maize-related VW flowed from the water-scarce North to the water-rich South, with an increase of 40% over the study period (from 43 to 61 billion m³ y^-1). In TS2, about 10% of the water embodied in maize exports from North to South China returns in the form of pork, with an increase in the absolute amount of 25% (from 4.8 to 6.1 billion m³ y^-1). Considering blue VW flows specifically, we find that North-to-South blue VW flows decreased by 5% in TS1, while South-to-North blue VW flows increased by 23% in TS2.