Publication details
Imputation of environmental variables for vegetation plots based on compositional similarity
| Authors | |
|---|---|
| Year of publication | 2010 |
| Type | Article in Periodical |
| Magazine / Source | Journal of Vegetation Science |
| MU Faculty or unit | |
| Citation | |
| Field | Botany |
| Keywords | Conductivity; Ellenberg indicator values; Fens; Phytosociology; Water pH |
| Description | Question: Vegetation plot data combined with measured environmental variables such as soil pH or conductivity are often used for gradient analyses, species response modelling, quantifying recent vegetation change and prediction of plant species composition. Large vegetation databases contain high numbers of plots, but only small subsets with measured environmental data. To obtain broader datasets, researchers often use expert-based plant indicator values as surrogates of measured factors. Alternatively, missing environmental factors for vegetation plots may be estimated by imputation. In this study we tested whether imputation provides more exact approximations than do indicator values. Location: Fens in the West Carpathians (Slovakia, Poland, Czech Republic) and Bulgaria. Methods: We developed a simple imputation method based on vegetation plot similarity that estimates the missing environmental variables for vegetation plots, and named it the MOSS (mean of similar samples) method. The method was tested for water pH and conductivity, the most important environmental factors influencing vegetation composition and structure within wetlands, on two datasets of 485 (West Carpathians) and 118 (Bulgaria) vegetation plots for which directly measured values were available. The West Carpathian dataset was used as a source of calibration. Imputation was based on calculating the mean of the measured factor from a group of the most similar vegetation plots. According to pre-defined similarity criteria we selected subsets of both datasets for which we compared estimated and measured values. Using the root mean squared error of prediction we compared the predictive power of the method with the widely used averages of Ellenberg indicator values as well as with other recently published methods. Results: Within one study region, the MOSS method predicted the sample pH and conductivity more precisely than Ellenberg indicator values and similar calibration methods. The predictive power slightly decreased when the method was transferred to a distant region. Conclusions: Imputation using the MOSS method appears to be the best way to predict pH or conductivity values from existing composition data within a single geographical region and thus increase the number of replicates. The method does not require expert-based indicator values, which may contain considerable imprecision. We provide examples of situations in which our method can be utilised without the risk of circular reasoning or introducing pseudo-replications. |
| Related projects: |