| The spatial distribution prediction of vegetation is of great importance not only to the conservation, utilization and restoration of the vegetation but also to the predictive distribution of vegetation communities in ecological restoration practices. In recent years with the development of the applied ecology and computer technology, lots of species distribution models are being emerging and also being used in nearly all branches of life and environmental sciences, such as biodiversity and conservation, forestry and so on. On one hand, much progress has been made in predictive models of species; on the other hand, it makes more difficult to use the species distribution models because the number of available techniques or models is large and is increasing steadily. What's more, there has many and big differences of the development background, theoretical basis, parameters demand and other aspects among different models, making it confused for users to select the most appropriate methodology or modeling approach for their needs, not only for predicting the spatial distribution of species, but also for projecting the potential distribution of species and its response to the climatic change. There is no doubt that it has been a practical problem in the ecological environment protection or management. A solution for this problem is to make direct comparisons of the predictions across models established on the same data, which can also provide an important basis for selecting suitable model. In this paper, according to the practice demand of the ecological restoration decision in Loess Plateau, Yanhe Basin, located in the loess hilly region of Loess Plateau, was selected as the study area. BIOMOD (BIOdiversity MODelling), a new computation framework based on R language, which aims to maximize the predictive accuracy of current species distributions and the reliability of future potential distributions using different types of statistical modelling methods was used and artificial neural networks (ANN) etc nine widely used models were choosed in this study. Then the potential distribution of the representative species in Yanhe River catchment were simulated and predicted combining the statistical method and GIS. On this basis, evaluated and compared the prediction results and predictive accuracy of the nine models, hoping to provide evidence for selecting best species distribution model. The main results are as follows:1. Based on vegetation investigation and data processing, using R and BIOMOD as the technology platform, the ANN etc nine models of the species distribution and the environmental factors were developed through constangly debugging, comparing the models and setting the models'parameters. Also calibrated and evaluated all the models, at the same time, the potential distribution of the fifteen main species were predicted by using the nine species-environment relation models which have been established.2. In BIOMOD, the relative importance of each variable in different models was calculated and extracted by using the function VarImportance. The results showed that: the importance of different environmental factors in different models is quite different; also there has a great difference in selecting environmenta factors across different models; furthermore, the models which selected more environmental factors had higher prediction accuracy comparing with the models which selected fewer factors.3. Three available techniques were used to assess each model's performance, Roc, Kappa and TSS. The evaluation indicated that whether according to the AUC, Kappa or TSS, RF model (the three value were 0.965,0.888,0.919 respectively), GBM model (the three value were 0.916,0.887,0.731 respectively) and ANN modle (the three value were 0.872,0.602,0.634 respectively) had higher prediction accuracy than the other six models. However, SRE model (the value of Kappa and TSS were 0.356,0.365 respectively) and GLM model (the three value were 0.764,0.461,0.549 respectively) had very low simulation precision. The evaluation results of Kappa were basically identical to TSS.4. Across different species, the relative model performance and simulation accuracy of nine models were quite different. The results show that the nine models have the highest predictive accuracy for the Thymus mongolicus Ronn.and Quercus liaotungensis distribution, the value of the Kappa,TSS and Roc were all very high, the simulation effects were excellent. Take the value of AUC for example, the AUC value for predicting the Thymus mongolicus Ronn.and Quercus liaotungensis were respectively 0.930 and 0.981. Yet the predictive accuracy of the nine models for Artemisia gmelinii web.ex stec distribution was lowest, the three values were all very low. Take the value of Kappa for example, the value of SRE,MARS,MDA and GLM were respectively 0.043,0.184,0.202 and 0.204. Obviously the predictions were failed. Furthermore, except for RF model and GBM model, the simulation effects of the other three models were also very poor. So we consider the seven models except for RF and GBM cannot be used for predicting the distribution of Artemisia gmelinii web.ex stec in pratical research.5. Comprehensive the comparisons of the theoretical assumptions, data characteristics, the algorithms, the limiting conditions of the models themselves and the prediction accuracy, the prediction results, the computation time and so on, and used the program for selecting best model in BIOMOD, it is considered that RF model which has the highest predictive accuracy using the three methods is the best one among the nine models. But in order to avoid the deviation by using a single model, four models such as RF, GBM, ANN and GAM which had higher prediction accuracy were finally selected. Then ensemble forecasting was implemented for prediciting the potential distribution of the six representative species such as Lespedeza davurica and so on by overlaying the prediction results of the four models.
|
PDF Full Text Request