Ensemble Modeling Approach to Predict the Potential Distribution of Artemisia sieberi in Desert Rangelands of Yazd Province, Central Iran
Authors
Abstract
The object of this study was to compare predictive accuracy of some individual modeling methods versus ensemble modeling approach in estimating the spatial potential distribution and identifying ecological requirements of Artemisia sieberi in desert rangelands of Yazd province, Central Iran. For this purpose, the species presence data were collected using the random systematic sampling method in 2019. Individual modeling of the species distribution was performed using Random Forest, Classification and Regression Tree and Generalized Additive Model after preparing environmental variable maps using GIS and geostatistics. Predictive performance of individual models was evaluated using Area Under Curve and Root Mean Square Error statistics. Furthermore, the Ensemble model was used based on the weighted average AUC. The appropriate threshold limit value was calculated based on True Skill Statistic for conversion of continuous maps to binary ones of habitat suitability. Comparison of the performance of individual models showed that the RF model had a more accurate prediction compared to the other models (AUC= 0.971 and RMSE= 0.256). Evaluation of the models implemented using threshold-dependent metrics such as Sensitivity, Specificity, and Kappa index also confirmed this finding. The overall comparison of the results from the three models versus the Ensemble model also indicates the high performance of this model compared to the individual models. Based on Ensemble model results, 45.38% of the study area had a high suitability for the establishment of A. sieberi. Based on the analysis of the importance of variables in the RF model, elevation (42%), Clay (40.02%) and pH (38.97%) in 0-30 cm soil depth had the highest effect on the presence of species. In general, Ensemble modeling can reduce the uncertainty and provide more reliable results by combining the results of the different algorithms of individual modeling.
