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

Modelling vegetation

In PREHAB we have compared the ability of selected modelling methods to classify distribution and abundance of 32 species of vegetation from four areas of the Baltic Sea. Four methods were evaluated for predicting distribution (presences and absences), and three methods were evaluated for predicting abundance (biomass and cover). All measures of performance represent results of validation using external test datasets.

1. Methods for modelling vegetation distribution

In general we found that classification accuracy of each method was on average "good" (AUC >0.8). Although there were small differences among techniques, RF had the highest average performance, followed by MaxEnt, GAM and MARS. See Figure 1.

However, the range and variance of classification accuracy differed among the methods. The most consistent technique was MaxEnt, ranging from good (AUC >0.8) to excellent (AUC >0.9) levels, while RF, GAM and MARS were slighly more variable. These differences in stability appears to be explained by differences in sensitivity to data properties. Analyses show that MARS was highly sensitive to total number of samples and the prevalence (the ratio of presences and the total number of samples) of the modelled species. Maxent was the least sensitive.

In conclusion, the obtained results from PREHAB indicate that good to excellent classification accuracy of vegetation distribution by the four evaluated methods is highly probable, where data consists of more than 400 samples and 30% presences.Figure 1. Classification accuracy (AUC) of four methods for modelling vegetation distribution. RF and MaxEnt provided the best models, but all approaches were useful. Dotted line represent level of relative good classification accuracy (AUC= 0.8); values above the line are considered as good and below it as poor. Grey dots=values; squares=means; whiskers =standard deviations.

2. Methods for modelling vegetation abundance

All three evaluated methods for modelling vegetation abundance provided models with average deviations of 10-20% of the observed variable range (NRMSE), see figure 2. Despite the lack of generally accepted guidelines, these levels of error must be considered relatively good accuracy of predictions. Random Forest (RF) provided models with the smallest deviations and the most consistent results, but particularly GAM also provided models which were as accurate as RF but the performance was more variable. Similarly to models for distribution, these differences in consistency appear to be explained by differences in sample size and species prevalence.

In conclusion, the obtained results from PREHAB indicate that relatively good prediction performance by the four evaluated methods is highly probable, where data consists of more than 400 samples and 40 percent presences.

Figure 2. Performance of three methods for modelling vegetation abundance. The prediction error expressed as root mean square error normalized by range (NRMSE) was smallest and most consistent for random forest (RF) of vegetation abundance by three regression methods. Grey dots=values; squares=means; whiskers=standard deviation.

Important assumptions

All four selected species distribution modelling methods have 3 important assumptions and they must be tested.

More

Modelling techniques - description

  • Generalized additive models (GAM)
  • Multivariate adaptive regression splines (MARS)
  • Random forest or random forests (RF)
  • Maximum entropy (MaxEnt)

Four modelling techniques evaluated by PREHAB are briefly described here

© 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?