| Spatial fingerprints of climate change on species are usually associated with changes in their latitudinal or altitudinal distributions. Species distribution modeling (SDM) technique, also known as niche model (NM) or habitat suitability model, that attempts to provide detailed predictions of distributions by statistically relating present-day species distribution to environmental predictors, has been widely applied to detect and predict species geographic distribution under climate change. Spatial prediction of species'geographic distributions has become a fundamental component of conservation planning, resource management, and environmental decision-making. China has been implementing a series of large-scale afforestation and reforatation programs. There is a pressing need to define better habitat suitable areas and guide these programs based on ecological and biogeographical rationality. Ecologically sound effects to manage or guide large scale vegetation projects require detailed knowledge of species habitat requirements. SDM techniques can assist this practice since they attempt to identify the probable causes of species whereabouts.There is now a plethora of methods for modelling species distribution with the rise of new powerful statistical techniques and GIS tools. These SDMs do not perform equally in predicting species' distribuion and sometimes differ substantially in both magnitude and direction of species range change in response to climate change. Selecting the most suitable model for simulating and projecting a given species' distribution under future climate needs a comprehensive model performance comparision. However, little is know about the relative performance of different models. Furthermore, SDM faces special challenges because of uncertainties in prediction. Recent studies have noted significant variability in model projections, which reuslts in a pressing need to reduce uncertainties. One step toward improving the reliability of SDM is building a range of models across more than one set of dataset initial conditions (IC), model classes (MC), model parameters (MP) and boundary conditions (BC) combinations (herein termed ensemble forecasting), and then find consensus in model projections (herein termed consensus forecasting). However, ensemble forcasting is in its infancy in assessing the effects of climate change on species' distirbution, and little is known about the source of variations and its contribution of uncertainty in ensemble forecasting, and the performance of different consensus approaches in the assessing climate change effects on species distribution.In general, species distribution modeling techniques, which based on statistical relationships between species occurrence data and the underlying environmental conditions, are classified into the two groups. One is profile technique (PT), which uses presence-only records to stimulate species' distribution. The other is group discrimination technique (GDT), which requires both presence and absence records to stimulate species' distribution. GDTs have been increasingly used in modeling species distribution because of higher reliability compared to PTs. However, GDTs also faces challenges because species occurrence records are usually found from museum, herbarium and atlases while true absences of species distribution are not available. To overcome this problem, GDTs generally use randomly generated pseudo-absences as absences. However, pseudo-absences can be located withinenvironmentally suitable areas where there is no species presence records duo to a variety of disturbance and ecology reasons. These pseudo-absence data will affect model prediction, because they feed false information to a model to relate species distribution to environments. Hence, the main objectives of this study are to (1) examine the effects of pseudo-absence sample size on model performance, (2) compare model performance, (3) partition and map uncertainties in ensembles of forecasts of species range change under climate change, (4) investigate whether different consensus approaches perform equally in extrapolating species distribution into a novel climate scenario, and (5) develop simple vegetation-climate classification model to project the potential effects of climate change on vegetation distribution in China. The study was conducted at a spatial resolution of 8 km. The study area comprises the whole China, an area characterized by dramatic climate and topography variation.The main research methods and results are as follows:1) In this study, I used a combination of two models, DOMAIN and NeuralEnsembles, to forecast the effects of climate change on spatial ditribution of Moso Bamboo (Phyllostachys edulis). Firstly, I used the profile technique- DOMAIN, to map potential habitat suitability for the species, and then select pseudo-absences from the areas predicted to have low habitat suitability. Secondly, I input pseudo-absences generated by DOMAIN alongside true presences into GDT- NeuralEnsembles to predict the potential distribution of Moso Bamboo. Sensitivity, Cohen' k-test and the area under the curve (AUC) values of receiver operator characteristic (ROC) curve were employed to assess model predictive accuracy. Meanwhile, I investigated the size effects of pseudo-absences generated by DOMAIN on model performance. Results indictaed that the combination of the two models could achieve a higher accuracy in simulating distribution of Moso Bamboo. Sensitivity and AUC are relatively independence of pseudo-absence size but Cohen's K-test declines with the increasing pseudo-absence size. Climate change is likely to have dramatic effects on the potential distribution of Moso Bamboo, with the northward migration ranging from 33 to 266 km, and the area expansion by 7.4% to 13.9%, while the potential habitat tends to be fragmented under the future climate change.2) Using eight niche-based models (Random forest, RF; Generalized boosted method, GBT; Generalized linear models, GLM; Generalized additive models, GAM; Classisication tree analysis, CTA; Artifical neuranl network, ANN; Mixture discriminant analysis, MDA; multivariate adaptive regression splines, MARS), nine set of model training dataset (initial data were randomly divided into two parts: mode training data and testing data), three global circulation models (GCM) (MIROC32_medres, JP; CCCMA_CGCM3; CA; BCCR-BCM2.0, NW) and one pessimistic SRES emissions scenarios (A2), I simulated current potential distribution and projected the future potential distribution of 12 prevailing planting tree species in China for the 3 time slices (2010-2039, 2020s; 2040-2069, 2050s; 2070-2099, 2080s). Totally, I obtained 72 predictions for current distribution and 201 projections for future distribution. Environmential envelop model- SRE, a submodel of BIOMOD platform, is used to select absences of species. AUC, true skill statistic (TSS) and Cohen's k-test were employed to objectively assess model predictive accuracy. ClimateChina software developed by our research group was used to downscale current and future (GCM) climate data and calculate seasonal and annual climate variables for specific locations based on latitude, longitude and elevation. Repeated measure ANOVA is used to assess the effect of model classes and species on model performance.Variation in model performance among niche models and tree species is significant. RF, GLM, GAM and GBM show higher model performance in simulating species' current distribution and be less sensitive to random data-splitting process compared to other models. Model performance varies with modelling methods, tree species and random data-splitting process. Future climate change is predicted to have substantial effects on 12 forestation tree species distribution and shows a general pattern of shifting northward and contracting their current distributions. In the case of Chinese pine, all projections suggested different species range change in response to climate change. Difference among current predictions mainly located surrounding area of current distribution of Chinese pine; difference among projections for future distribution increases with increasing time horizon along and the area that has great variation will expand and the area has little variation will reduce.3) I performed a three-way analysis of variance (ANOVA) to partition the source of uncertainties for each grid cell, with model training data, model techniques and GCMs as factor and species occurrence probability as response variable. I then obtained the sum of squares which can be atrributed to training data, model techniques and GCMs and their interaction (training data×model techniques, GCMs×training data, model techniques×GCMs, training data×model techniques×GCMs). Model technique contributed to the largest variation in projections, while GCM and model training data had little influence on the variability in projections. This practice will reinforce our understanding of the source of uncertainties in modelling species distribution unde climate change.4) To reduce uncertainty in modelling species distribution, I combined ensembles of forecasting and subsequently get final consensual prediction maps for target species by using three different consensus approaches (Average, Frequency and Median (PCA)). Pearson's correlation coefficient and Cohen's K-test were used to quantify the similarity of three consensual prediction maps. Results indicate the magnitude and direction of species' relative range change (area change percent, range shifting direction and distance and changes in elevation optimum) in response to climate change among three consensual projections is not significantly different. But, differences in spatial similarity of three consensual predictions were detected for each species. For most (but not all) species, the congruent area that species will present predicted by three consensus approaches together mainly located in the central area of species range, while incongruent area mainly located at the borders of species range or at discrete distributions area. Correlation among consensual maps was highest between Average and Frequency, while Median (PCA) deviated much from them two.5) Based on equilibrium relationship between climate and vegetation distribution and primary vegetation in every part of China, I related vegetation distribution to climate variables by using Random forest technique. Future climate change is projected to have substantial effects on vegetation distribution in China. Forests in East China show a general pattern of shifting northward under future climate, especially for the boreal coniferous forest (BCF) which will have a large ruduction in distribution range. Temperate meadow steppe (TMS) and temperate desert steppe (TDS) will emerge in Tibet-Qinghai plateau. Forest and temperate desert vegetation (Temperate shrub desert and dwarf semi-shrub desert + Temperate dwarf semi-arboreous desert) will expand under future climate scenarios, but for temperate grassland (temperate meadow steppe + temperate desert steppe + temperate typical steppe), Tibet-Qinghai plateau vegetation (Alpine steppe, AS; Alpine meadow, AM; Alpine desert, AD) and subtropical mountainous cool coniferous forest (SMCF) will shrink. This study provide insight into the vegetation-climate relationship in China and facilitate vegetation dynamic modeling.In this study, species distributions were modeled across more than one set of initial conditions (model building data), model classes and boundary conditions (GCM) combinations. Although the study only restricted to 12 tree species, it provides conceptual insights into the uncertainty in modelling species response to climate change. Niche models do not perform equally in simulating current distribution with highest variance in AUC, TSS and Cohen's k-test happed to MDA. GAM, GBM, GLM and RF have the potential to provide accurate models more often than the other methods employed in this study and be insensitive to random split-sample. My results support previous findings that model robustness is related to model complexity. Consensus approaches do not perform equally in projecting species distribution into a new climate scenarios. Species' relative range changes in response to climate change does not reflect the performance of consensus approaches. None the less, I concluded that climate change adaptation policies based on the consensus forecasting of species distribution may be more robust than based on single niche model since I demonstrated differences in predictive performance among different modeling methods, despite substantial variation at both regional and species levels. Additionally, it is better to base resource management and policy-making on species' probabilistic distribution map rather than binary distribution map.These results further our understanding of uncertainties in SDM and draw attention to the importance of selecting niche models and consensus approaches for projecting the effects of climate change on species potential distribution. The results of this study also emphasize that developing new model with high transferability will be a promising approach for assessing climate change impacts and for the effectiveness of ensemble forecasting. Moreover, biotic interactions, dispersal and niche evolution process are also at play and matter ecologically in determining species' distribution. Further steps would also need to incorporate scale-dependent effects and more determinants of species distribution into SDM and to explore how to combine ensembles of forecasting. Future climate change is likely to have a substantial effect on forestation tree species' distribution. This study could also help to define better habitat suitable areas for current large-scale reforestation and afforestation programs in China and provide insights into adaptive forestry management in the context of climate change.
|
PDF Full Text Request