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

Results on predictive mapping 2010

What can be modelled?

Task 2.4: Predictability of different types of response variables

A total of 70-80 response variables from different parts of the Baltic Sea region have been modelled. The majority of these have been modelled both quantitatively and qualitatively (presence/absence). Within all the case study areas three main groups of response variables has been modelled: individual species, functions of habitats and benthic communities/biotopes.

In the western part of Sweden, 10 algal species and 3 functional groups were modelled. RF models were the most accurate ones.

In the eastern part of Sweden, six fish species and one functional group were modelled. RF, GAM and MAXENT models were the most accurate ones.

In Lithuanian waters, 23 zoobenthos species, 1 macrophytobenthos species and 6 functional groups were modelled. RF, MARS and KRIGING models were the most accurate ones.

Detailed and comprehensive conclusions about predictability of different response variables, including practical recommendations for mapping efforts in the Baltic will be summarised in a scientific paper as well as incorporated in the web-resource which is developed within task 4.2.

Which type of information can be used for modeling?

Task 2.3: Environmental predictors for modelling benthic habitats

A total of more than 50 predictor variables from different parts of the Baltic Sea region have been modelled. For purposes of overall comparison and synthesis, they have been classified according as either of the types location, bathymetry, substrate, exposure, hydrography and biotic. The majority of these have been modelled both quantitatively and qualitatively (presence/absence).

According to preliminary results the importance of predictors differed among study areas:

In western part of Sweden substrate and bathymetry were the most important environmental predictors, where location were the least significant one.

Location was the most important predictor in models from eastern part of Sweden, followed by exposure and hydrography.

Bathymetry, hydrography and exposure were the most important predictors in Östergötland, whereas water current was the least important.

In case of Lithuanian waters, bathymetry and substrate predictors were the most important in both regression and classification models. Hydrography and exposure predictors were more important for classification models than for regression models.

Detailed and comprehensive conclusions about efficiency of different predictors, including practical recommendations for mapping efforts in the Baltic, will be summarised in a scientific paper as well as incorporated in the web-resource which is developed within task 4.2.

Which statistical tools can be used for modeling?

Task 2.1: Analysis of scale-dependent performance of predictive habitat models

Sets of data on suitable for modelling scale-dependence of response variables in predictive power has been established by GU. Simulation and analytical modelling of precision of quantitative response variables is ongoing.

This work involves modelling of precision and empirical limits of regression models at different spatial scales and effects of variable grid size in predictive models. During the development of this task the significance of measurement- and other errors related to predictor variables such as depth and substrate, which is also scale-dependent, has become evident. Therefore the work within this task will also involve analyses of errors due to uncertainty in predictors.


Task 2.2: Comparing modelling techniques

Five different statistical techniques have been evaluated for predicting spatial occurrence of species and habitats. Approximately 700 modelling runs were performed for all partners, techniques and variables.

Results show that in general no differences were found among the modelling methods, whereas from data of eastern part of Sweden models differed among the species.

Accuracy of models was relatively good among case study areas (generally >0.8 AUC):

In western part of Sweden RF (Random forest) was the best model for both classification and regression models, followed by GAM, MARS and Kriging.

GAM models were the most accurate ones for the data from Östergötland.

In case of Lithuanian waters (KU) the accuracy of regression models based on r2 was similar among the methods, where RF was the best according to normalized RMSE.

Accuracy of classification models based on AUC was relatively high for all models, except MAXENT (ca. 0.6 AUC).
 

Detailed and comprehensive conclusions about efficiency of different techniques, including consequences for their practical application in mapping efforts in the Baltic, will be summarised in a scientific paper as well as incorporated in the web-resource which is developed within task 4.2.

 

© 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?