Uncertainty Analysis Of Spatial Distribution Prediction Based On MaxEnt

To date, along with the widely application of species distribution models, accuracy and uncertainty about the outcomes of species prediction have been considered and assessed. The best performing models are not always the same for different species, due to the difference of uncertainty factors such as parameters setting, the ecological distribution prediction of species from the same model is also different, while quantify the contribution from different sources of uncertainty will help toreduce the variance of ecological predictions and decide where to take into account in the future research to reduce variability and improve the reliability in projections. However, very few researches quantify the uncertainties from single model, especially for MaxEnt applied widely.This study is as Angelica sinensis a case, analyzing the distribution uncertainty of A. sinensis from three aspects (pseudo-absences points selection, variable selection and parameter setting) based on MaxEnt. Pseudo-absences points selection use the method of two degrees distance and altitude< 1,500 m to get two sets of pseudo-absences points of A. sinensis and build model; Variable selection use principal component analysis(PCA), variance inflation factor(VIF), single factor selection(SFS) and ecological habits variables selection methods(EHVS) to select environmental factors, determining seven sets of variables set (principal component, two degrees VIF, altitude VIF, single factor altitude, single factor two degrees, single factor 25%, ecological variable group) to build model; Varameter setting use 25% of occurrence data and pseudo-absences points as test data to verify this model, at the same time we set the test percentage of occurrence data to 5%,5%,15%,20%,25%,30%,35%,40%,45% and 50% to investigate the effect of the set scale on the model; By analyzing the AUC value of training and test, spatial distribution and suitable area to get the optimal pseudo-absences points selection method, the optimal variable selection method and the optimal model parameters settings based on MaxEnt. At the same time, through the analysis of the spatial distribution and two factor analysis of variance, we get the contribution rate of pseudo-absences points selection, variable selection and parameter setting. Finally, according to the optimal pseudo-absences points selection method, the optimal variable selection method and the optimal model parameters settings, we determine the optimal species distribution model, evaluating and analyzing the result of the model, getting the potential spatial distribution, the optimum ecological factors and the optimum factor value range of A. sinensis, giving advices on the protection and cultivation of wild angelica germplasm resources. This information is of value to provide powerful reference for the wild resources protection and ecological planning of Chinese herbal medicine. The main conclusions of this study are as follows:1. Seven sets of variables and two sets of pseudo-absences point datas constructed 42 sets of model. The uncertainty analysis of space prediction distribution for A. sinensis based on pseudo-absences points selection shows that pseudo-absences datas obtained by two degrees distance method have better effect on the test precision of the model from the prediction accuracy of model and suitability area. The size of suitability area and the training precision of model have more close relationship with the quality and quantity of occurrence data.2. The uncertainty analysis of space prediction distribution for A sinensis based on parameter settings founds that by analyzing the suitability area and the model precision under seven sets of variables set, the model fitting accuracy is higher when 25% of the occurence point datas participate in the test. This is the reason that most studies commonly used 25% of the occurence point datas for test. However, based on the test percentage analysis of occurence points, it is concluded that when 15% of occurence point datas were used for test, the test precision of the model is the highest and the prediction accuracy of the model is the best.3. The uncertainty analysis of space prediction distribution for A. sinensis based on variable selection shows that by analyzing the AUC value of training and test, the appropriate area, space distribution as well as dominant environmental factor under seven sets of variables set, principal component analysis can eliminate redundant information between factors, the model based on variable selected by principal component analysis has higher accuracy, the suitable area is more in line with the actuality, the environment factors are more accord with the growth habit of A. sinensis. So it is more important for model prediction accuracy to eliminate redundant information between factors.4. The contribution analysis of the prediction results based on three uncertainty factors (pseudo-absences points selection, variable selection and parameter setting) shows that the sorting of importance on potential suitability distribution for A. sinensis based on MaxEnt is variable selection> parameter setting> pseudo-absences points selection.5. This study seleted the optimal situation that 120 pseudo-absences points selected by two degrees distance method, principal component variables set and 15% of occurrence points as test set. The optimal predicted results show that the moderately and highly suitable habitats of A. sinensis were mainly located in the southeast of Gansu and Tibet, the north of Sichuan and Yunnan. ASL(Elevation above sea level), BIO5(Max temperature of the warmest month), ATG(Average temperature of growth), BIO 1 (Annual mean temperature), ASH(Annual sunshine hours), BIO15(Coefficient variation of precipitation seasonality) and BIO12(Annual average precipitation) are the dominant environmental factors for the growth and distribution of A. sinensis. The ecological threshold of most of the environmental factors showed a decreasing trend from unsuitable habitats to highly suitable habitats, the mean and standard deviation of each ecological factor in the moderately and highly suitable habitats were similar, the ecological factor parameters show a consistent trend.This study discusses the influence of three uncertainty factors on the species distribution prediction based on MaxEnt, which can provide the reference for species distribution prediction in the future and improve the accuracy of the prediction. It can provide real and effective guidance for species protection or development and utilization.