Till startsida
To content Read more about how we use cookies on gu.se

PREHAB top-10 conclusions for predictive modelling and mapping

We have used four fundamentally different approaches for modelling distribution and abundance of vegetation, invertebrates and fish. We have used data from four case-study areas, representing a wide range of conditions in the Baltic Sea. The quality and quantity of both biological and environmental data has been highly variable. This broad assessment has resulted in a number of general conclusions and recommendations about sampling and modelling for predictive mapping of coastal habitats in the Baltic Sea.

  1. Predictive modelling based on species-environment relationships can successfully be used for mapping the distribution and abundance of coastal species and habitats in the Baltic Sea. The performance of the models typically meet the standards what is considered useful by scientific criteria.
  2. There are no strong differences in the usefulness of predictive modelling among different parts of the Baltic Sea.
  3. It is possible to model and map distribution and abundance of different types of biodiversity, e.g. vegetation, benthic invertebrates and fish. 
  4. Overall, there are large differences in predictability among species. Some of these differences are explained by types of organisms and ecological features, while other by characteristics of the data. For example, rare species (lower prevalence) distribution is easier to predict than widespread species, whereas rare species abundance tend to be more difficult to predict.
  5. Several alternative statistical methods for modelling can be used. Of the ones tested within PREHAB, Maxent and RandomForest performed best for models of distribution while RandomForest or GAM were the best for models of abundance. Differences were, however, relatively small and as a general recommendation we suggest that more than one method is applied, in order to minimize modelling uncertainty.
  6. Bottom topography (bathymetri) and bottom substrate were the most powerful and important predictors in general. Access to high resolution data on depth and substrate can greatly improve modelling and mapping, and is more or less a requirement for fulfilling the potential of predictive modelling.
  7. The importance of some environmental predictors varied among organisms. Wave exposure tended to be more important for models of vegetation while hydrographic factors tended to be more important for models of fish.
  8. The quality and amount of field data is a very important determinant of the average and consistency in performance of models. Small sample sizes tend to be more variable in performance. Sample sizes of around 100 may sometimes produce useful models, but those based on 500-1000 samples appear to be more reliable.
  9. Biological data for predictive modelling need to collected using (a) representative and independent samples, (b) at a spatial resolution suitable for modelling and (c) with appropriate and reliable taxonomic resolution. This means that we recommend sampling in points (corresponding to model grid size) rather than transects or polygons. If models are based on such data, e.g. because it is available or because other priorities than modelling is accounted for, proper care needs to be taken against bias due to spatial autocorrelation.
  10. The spatial resolution of models greatly affects the predictive power of models. Maximum resolution (e.g. <1 m) is not always better than aggregation at a larger scale (e.g. 10 m). A priori definition of resolution, with known sampling error, is important for the design of efficient sampling programmes for predictive purposes.


© University of Gothenburg, Sweden Box 100, S-405 30 Gothenburg
Phone +46 31-786 0000, About the website

| Map

The University of Gothenburg uses cookies to provide you with the best possible user experience. By continuing on this website, you approve of our use of cookies.  What are cookies?