Computational protocol: The Consequences of Precipitation Seasonality for Mediterranean-Ecosystem Vegetation of South Africa

Similar protocols

Protocol publication

[…] Conservation areas within South Africa are dominated by large reserves within savannas. To eliminate this sample bias, the data were averaged for each unique ‘vegetation unit’, i.e. as described by Mucina and Rutherford [], occurring with a particular geological stratigraphic description. As a consequence, some vegetation units were sampled repeatedly where these were associated with a number of geological formations. This resulted in 1 116 unique combinations of vegetation and stratigraphy occurring within conservation areas. These vegetation unit-stratigraphic combinations were used directly for analyses across all vegetation types. The climatic and NDVI data were summarized for each vegetation type and subjected to one-way ANOVA.Predictor variables were screened for colinearity using the “select07” procedure outlined by Dormann et al. [], and less powerful variables removed where colinearity existed. The retained variables were included in multiple regression analysis of the quadratic predictors of NDVI and BS with stepwise backward-elimination based on the Akaike Information Criteria (AIC). We also examined dependence of NDVI on water availability and BS for each vegetation type individually using multiple regression.Boosted regression trees (BRT) provide a machine learning-based model of response variables, and do so without involving normal null-hypothesis significance testing. BRT model construction was performed, as detailed by Elith et al. [] and implemented in R []. Models for NDVI were constructed with water availability and the edaphic variables (see above) using the ‘dismo’ package version 0.7–23 (Hijmans et al. 2012). Tree complexity (5) and learning rate (0.01) were optimised for the analysis and a bagging fraction of 0.5 was used. After initial BRT analysis, the model was simplified following procedures outlined by Elith et al. []. The BRT analysis was used to rank the importance of different predictor variables in determining the NDVI.A structural equation model (SEM) was constructed to evaluate the relative strength of the influence of water and soil properties on NDVI. The SEM was produced using backward-simplification based on the AIC score of quadratic multiple regression of NDVI on water-related measures (P-PET, precipitation seasonality and PCI) and soil related measures (BS, TEB, CEC and organic C and clay) to yield two composite variables which were included in the SEM implemented using the package Lavaan [] in R []. […]

Pipeline specifications

Software tools dismo, lavaan
Application Phylogenetics