Comparison of three modelling approaches for predicting plant species distribution in mountainous scrub vegetation (Semnan Rangelands, Iran)

The predictive modeling of plant species distribution has wide applications in vegetation studies. This study attempts to assess three modeling approaches to predict the plant distribution in the dry (precipitation 128–275 mm) mountainous (altitude 1129–2260 m a.s.l.) scrub vegetation on the example of the rangelands of northeastern Semnan, Iran. The vegetation of the study area belongs to the communities of Artemisia, Astralagus, Eurotia and other scrub species. The main objective of this study is to compare the predictive ability of three habitat models, and to find the most effective environmental factors for predicting the plant species occurrence. The Canonical Correspondence Analysis (CCA), Logistic Regression (LR), and Artificial Neural Network (ANN) models were chosen to model the spatial distribution pattern of vegetation communities. Plant density and cover, soil texture, available moisture, pH, electrical conductivity (EC), organic matter, lime, gravel and gypsum contents and topography (elevation, slope and aspect) are those variables that have been sampled using the randomized systematic method. Within each vegetation type, the samples were collected using 15 quadrates placed at an interval of 50 m along three 750 m transects. As a necessary step, the maps of all factors affecting the predictive capability of the models were generated. The results showed that the predictive models using the LR and ANN methods are more suitable to predict the distribution of individual species. In opposite, the CCA method is more suitable to predict the distribution of the all studied species together. Using the finalized models, maps of individual species (for different species) or for all the species were generated in the GIS environment. To evaluate the predictive ability of the models, the accuracy of the predicted maps was compared against real-world vegetation maps using the Kappa statistic. The Kappa (κ) statistic was also used to evaluate the adequacy of vegetation mapping. The comparison between the vegetation cover of a map generated using the CCA application and its corresponding actual map showed a good agreement (i.e. κ= 0.58). The results also revealed that maps generated using the LR and ANN models for Astragalus spp., Halocnemum strobilaceum, Zygophyllum eurypterum and Seidlitzia rosmarinus species have a high accordance with their corresponding actual maps of the study area. Due to the high level of adaptability of Artemisia sieberi, allowing this specie to grow in most parts of the study area with relatively different habitat conditions, a predictive model for this species could not be fixed. In such cases, a set of predictive models may be used to formulate the environment-vegetation relationship. Finally, the predictive ability of the LR and ANN models for mapping Astragalus spp. was determined as κ = 0.86 and κ = 0.91 respectively, implying a very good agreement between predictions and observations. It is concluded that the combination of mod- elling of the local species distribution constitutes a promising future research area, which has the potentiality to enhance assessments and conservation planning of vegetation (like rangelands) based on predictive species models.