Data from 'Wetland plant hydrological zonation predicted from observed water depth'

Description

This repo included four files: two datasets (one used to build, the other to validate, a model of wetland plant zonation), the fitted model object ('topmod') of class(brmsfit), and, an R script to reproduce the analysis. Each dataset consists of presence-absence of plants within quadrats, within wetlands, where plant species identity has been used to assign one of three hydrological response groups: T = terrestrial, A = amphibious, S = submerged. These are based on the water plant functional group classification of Casanova & Brock (2000, Plant Ecology V 147, pp 237-250). Accompanying the species / hydrological group composition for each quadrat is an observed depth measurement. All quadrats were observed in the local (Austral) Spring. The build data (n = 813 observations of 1 x 1 m quadrats across 12 wetland complexes in temperate Southeast Australia) were collected in October 2013; the validation data (n = 198, 23 wetland complexes) are completely independent of the model build data and were compiled from multiple monitoring programs over the period 2008-2012. Wetlands were from the same region, but were largely not sampled in the surveys used in the model build dataset. The central premise of the study is that wetland hydrological niche space can be conceptualized as comprising three distinct zones, respectively dominated by fringing terrestrial, amphibious, and submerged species. This offers a framework to predict high-level impacts on wetland vegetation, characterizing the hydrological niche using the observed water depth in the Spring growing season and defining vegetation zones as the relative number of species adapted to terrestrial, amphibious, and submerged conditions (hydrological group richness, HGR). A regression model was first developed using the build data to predict the hydrological conditions (ie depth) marking transition points between these zones. The model explained 0.72 of variation in HGR, predicting submerged-terrestrial (median ± [95% credible intervals] = 13 [7, 22] cm) and submerged-amphibious transitions (62 [50, 72] cm) directly and estimating amphibious-terrestrial transitions (predicted to occur below ground; -8 [-18, -2] cm) from extrapolation. These thresholds were tested using the validation data and predictive performance decreased in the order submerged-terrestrial > amphibious-terrestrial > submerged-amphibious transitions. Only the latter proved unreliable (Cohen’s kappa = 0.73, 0.59 and 0.10 respectively). Results suggest it could be possible to identify critical wetland tipping points at the dry end of the hydrological niche using only simple summary statistics and show improved precision in predicting terrestrial-amphibious transitions (a critical tipping point in climatic drying risk) will require a focus shallow sub-surface (e.g., 0 to -20 cm) saturation dynamics.

Steps to reproduce

Data: Two independent datasets were used, one to build models, the second for model validation. The model building data were used in Deane et al. (2017, Ecol Appl V27) and consist of 817 1 x 1 m quadrats, in which presence of all species was recorded along with the water depth in centimetres. Quadrats were positioned haphazardly along multiple elevation gradients within 17 wetland basins from 13 wetland complexes with contrasting geomorphology and hydrology. Validation data were compiled from a range of studies collated within the South Australian Wetland Inventory Database from a range of published sources, and from unpublished regional monitoring programs (see manuscript for details). The R script imports the raw data (model build = "Goyder_all_dat.csv", validate = "sprdat.csv"), pre-processes it into a form used in modelling and validation and re-runs the analyses. Model building: the response variable was the relative number of species within each hydrological grouping (i.e., terrestrial, amphibious, or submerged species, denoted T, A and S respectively) from the Brock & Casanova water plant functional group classification. The number of species within each is referred to as hydrological group richness (HGR). The response was modelled using a Bayesian multilevel regression model fit in Stan, and implemented using R package brms. The model used a zero-inflated binomial (ZIB) response distribution to model HGR. In this multilevel regression the Level 1 predictors (i.e., fixed effects) were linear and quadratic observed depth of water in the Spring growing season while Level 2 predictors (i.e., random effects) were included for hydrological group, nested within wetland. The ZIB regression model was estimated in Stan, using R package brms. The R script will re-run the final model using the original data. Validation: Having fit the model, it was used to predict the depth where pairwise comparison of the number of species in each hydrological group indicated a change in zonation. To validate predicted transitions between T, A, and S dominated zones inferred from the model, a confusion matrix was calculated for each hydrological group using the validation data. If the 'wetter' adapted group in the comparison contained more species the observed value was assigned to that group, otherwise the drier adapted group was declared the 'winner'. This outcome was compared with the predicted outcome based on the observed depth for the quadrat. The confusion matrix compared the dominant hydrological group in the quadrat with the predicted and was analysed using common statistical metrics for pairwise classification (e.g., Cohen's kappa, specificity, etc). All steps in analysis are outlined in the R script. If using a Windows-based system, simply save all files to a common folder and the script should run.

Institutions

Categories

Funding

Licence